first commit

LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2022 Machine Vision and Learning Group, LMU Munich
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

app.py

@@ -0,0 +1,219 @@
+import math
+import fire
+import gradio as gr
+import numpy as np
+import rich
+import torch
+from contextlib import nullcontext
+from einops import rearrange
+from functools import partial
+from ldm.models.diffusion.ddim import DDIMSampler
+from ldm.util import load_and_preprocess, instantiate_from_config
+from omegaconf import OmegaConf
+from PIL import Image
+from rich import print
+from torch import autocast
+from torchvision import transforms
+
+
+_SHOW_INTERMEDIATE = True
+_GPU_INDEX = 0
+# _GPU_INDEX = 2
+
+
+def load_model_from_config(config, ckpt, device, verbose=False):
+    print(f'Loading model from {ckpt}')
+    pl_sd = torch.load(ckpt, map_location=device)
+    if 'global_step' in pl_sd:
+        print(f'Global Step: {pl_sd["global_step"]}')
+    sd = pl_sd['state_dict']
+    model = instantiate_from_config(config.model)
+    m, u = model.load_state_dict(sd, strict=False)
+    if len(m) > 0 and verbose:
+        print('missing keys:')
+        print(m)
+    if len(u) > 0 and verbose:
+        print('unexpected keys:')
+        print(u)
+
+    model.to(device)
+    model.eval()
+    return model
+
+
+@torch.no_grad()
+def sample_model(input_im, model, sampler, precision, h, w, ddim_steps, n_samples, scale,
+                 ddim_eta, x, y, z):
+    precision_scope = autocast if precision == 'autocast' else nullcontext
+    with precision_scope('cuda'):
+        with model.ema_scope():
+            c = model.get_learned_conditioning(input_im).tile(n_samples, 1, 1)
+            T = torch.tensor([math.radians(x), math.sin(
+                math.radians(y)), math.cos(math.radians(y)), z])
+            T = T[None, None, :].repeat(n_samples, 1, 1).to(c.device)
+            c = torch.cat([c, T], dim=-1)
+            c = model.cc_projection(c)
+            cond = {}
+            cond['c_crossattn'] = [c]
+            c_concat = model.encode_first_stage((input_im.to(c.device))).mode().detach()
+            cond['c_concat'] = [model.encode_first_stage((input_im.to(c.device))).mode().detach()
+                                .repeat(n_samples, 1, 1, 1)]
+            if scale != 1.0:
+                uc = {}
+                uc['c_concat'] = [torch.zeros(n_samples, 4, h // 8, w // 8).to(c.device)]
+                uc['c_crossattn'] = [torch.zeros_like(c).to(c.device)]
+            else:
+                uc = None
+
+            shape = [4, h // 8, w // 8]
+            samples_ddim, _ = sampler.sample(S=ddim_steps,
+                                             conditioning=cond,
+                                             batch_size=n_samples,
+                                             shape=shape,
+                                             verbose=False,
+                                             unconditional_guidance_scale=scale,
+                                             unconditional_conditioning=uc,
+                                             eta=ddim_eta,
+                                             x_T=None)
+            print(samples_ddim.shape)
+            # samples_ddim = torch.nn.functional.interpolate(samples_ddim, 64, mode='nearest', antialias=False)
+            x_samples_ddim = model.decode_first_stage(samples_ddim)
+            return torch.clamp((x_samples_ddim + 1.0) / 2.0, min=0.0, max=1.0).cpu()
+
+
+def main(
+    model,
+    device,
+    input_im,
+    preprocess=True,
+    x=0.,
+    y=0.,
+    z=0.,
+    scale=3.0,
+    n_samples=4,
+    ddim_steps=50,
+    ddim_eta=1.0,
+    precision='fp32',
+    h=256,
+    w=256,
+):
+    # input_im[input_im == [0., 0., 0.]] = [1., 1., 1., 1.]
+    print('old input_im:', input_im.size)
+
+    if preprocess:
+        input_im = load_and_preprocess(input_im)
+        input_im = (input_im / 255.0).astype(np.float32)
+        # (H, W, 3) array in [0, 1].
+
+    else:
+        input_im = input_im.resize([256, 256], Image.Resampling.LANCZOS)
+        input_im = np.asarray(input_im, dtype=np.float32) / 255.0
+        # (H, W, 4) array in [0, 1].
+        
+        # old method: very important, thresholding background
+        # input_im[input_im[:, :, -1] <= 0.9] = [1., 1., 1., 1.]
+
+        # new method: apply correct method of compositing to avoid sudden transitions / thresholding
+        # (smoothly transition foreground to white background based on alpha values)
+        alpha = input_im[:, :, 3:4]
+        white_im = np.ones_like(input_im)
+        input_im = alpha * input_im + (1.0 - alpha) * white_im
+        
+        input_im = input_im[:, :, 0:3]
+        # (H, W, 3) array in [0, 1].
+
+    print('new input_im:', input_im.shape, input_im.dtype, input_im.min(), input_im.max())
+    show_in_im = Image.fromarray((input_im * 255).astype(np.uint8))
+
+    input_im = transforms.ToTensor()(input_im).unsqueeze(0).to(device)
+    input_im = input_im * 2 - 1
+    input_im = transforms.functional.resize(input_im, [h, w])
+
+    sampler = DDIMSampler(model)
+    x_samples_ddim = sample_model(input_im, model, sampler, precision, h, w,
+                                  ddim_steps, n_samples, scale, ddim_eta, x, y, z)
+
+    output_ims = []
+    for x_sample in x_samples_ddim:
+        x_sample = 255. * rearrange(x_sample.cpu().numpy(), 'c h w -> h w c')
+        output_ims.append(Image.fromarray(x_sample.astype(np.uint8)))
+    
+    if _SHOW_INTERMEDIATE:
+        return (output_ims, show_in_im)
+    else:
+        return output_ims
+
+
+description = '''
+Generate novel viewpoints of an object depicted in one input image using a fine-tuned version of Stable Diffusion.
+'''
+
+article = '''
+## How to use this?
+TBD
+## How does this work?
+TBD
+'''
+
+
+def run_demo(
+    device_idx=_GPU_INDEX,
+    ckpt='last.ckpt',
+    config='configs/sd-objaverse-finetune-c_concat-256.yaml',
+):
+
+    device = f'cuda:{device_idx}'
+    config = OmegaConf.load(config)
+    model = load_model_from_config(config, ckpt, device=device)
+
+    inputs = [
+        gr.Image(type='pil', image_mode='RGBA'),  # shape=[512, 512]
+        gr.Checkbox(True, label='Preprocess image (remove background and center)',
+        info='If enabled, the uploaded image will be preprocessed to remove the background and center the object by cropping and/or padding as necessary. '
+        'If disabled, the image will be used as-is, *BUT* a fully transparent or white background is required.'),
+        # gr.Number(label='polar (between axis z+)'),
+        # gr.Number(label='azimuth (between axis x+)'),
+        # gr.Number(label='z (distance from center)'),
+        gr.Slider(-90, 90, value=0, step=5, label='Polar angle (vertical rotation in degrees)',
+        info='Positive values move the camera down, while negative values move the camera up.'),
+        gr.Slider(-90, 90, value=0, step=5, label='Azimuth angle (horizontal rotation in degrees)',
+        info='Positive values move the camera right, while negative values move the camera left.'),
+        gr.Slider(-2, 2, value=0, step=0.5, label='Radius (distance from center)',
+        info='Positive values move the camera further away, while negative values move the camera closer.'),
+        gr.Slider(0, 30, value=3, step=1, label='cfg scale'),
+        gr.Slider(1, 8, value=4, step=1, label='Number of samples to generate'),
+        gr.Slider(5, 200, value=100, step=5, label='Number of steps'),
+    ]
+    output = [gr.Gallery(label='Generated images from specified new viewpoint')]
+    output[0].style(grid=2)
+    
+    if _SHOW_INTERMEDIATE:
+        output += [gr.Image(type='pil', image_mode='RGB', label='Preprocessed input image')]
+
+    fn_with_model = partial(main, model, device)
+    fn_with_model.__name__ = 'fn_with_model'
+
+    examples = [
+        # ['assets/zero-shot/bear.png', 0, 0, 0, 3, 4, 100],
+        # ['assets/zero-shot/car.png', 0, 0, 0, 3, 4, 100],
+        # ['assets/zero-shot/elephant.png', 0, 0, 0, 3, 4, 100],
+        # ['assets/zero-shot/pikachu.png', 0, 0, 0, 3, 4, 100],
+        # ['assets/zero-shot/spyro.png', 0, 0, 0, 3, 4, 100],
+        # ['assets/zero-shot/taxi.png', 0, 0, 0, 3, 4, 100],
+    ]
+
+    demo = gr.Interface(
+        fn=fn_with_model,
+        title='Demo for Zero-Shot Control of Camera Viewpoints within a Single Image',
+        description=description,
+        article=article,
+        inputs=inputs,
+        outputs=output,
+        examples=examples,
+        allow_flagging='never',
+    )
+    demo.launch(enable_queue=True, share=True)
+
+
+if __name__ == '__main__':
+    fire.Fire(run_demo)

configs/sd-objaverse-finetune-c_concat-256.yaml

@@ -0,0 +1,117 @@
+model:
+  base_learning_rate: 1.0e-04
+  target: ldm.models.diffusion.ddpm.LatentDiffusion
+  params:
+    linear_start: 0.00085
+    linear_end: 0.0120
+    num_timesteps_cond: 1
+    log_every_t: 200
+    timesteps: 1000
+    first_stage_key: "image_target"
+    cond_stage_key: "image_cond"
+    image_size: 32
+    channels: 4
+    cond_stage_trainable: false   # Note: different from the one we trained before
+    conditioning_key: hybrid
+    monitor: val/loss_simple_ema
+    scale_factor: 0.18215
+
+    scheduler_config: # 10000 warmup steps
+      target: ldm.lr_scheduler.LambdaLinearScheduler
+      params:
+        warm_up_steps: [ 100 ]
+        cycle_lengths: [ 10000000000000 ] # incredibly large number to prevent corner cases
+        f_start: [ 1.e-6 ]
+        f_max: [ 1. ]
+        f_min: [ 1. ]
+
+    unet_config:
+      target: ldm.modules.diffusionmodules.openaimodel.UNetModel
+      params:
+        image_size: 32 # unused
+        in_channels: 8
+        out_channels: 4
+        model_channels: 320
+        attention_resolutions: [ 4, 2, 1 ]
+        num_res_blocks: 2
+        channel_mult: [ 1, 2, 4, 4 ]
+        num_heads: 8
+        use_spatial_transformer: True
+        transformer_depth: 1
+        context_dim: 768
+        use_checkpoint: True
+        legacy: False
+
+    first_stage_config:
+      target: ldm.models.autoencoder.AutoencoderKL
+      params:
+        embed_dim: 4
+        monitor: val/rec_loss
+        ddconfig:
+          double_z: true
+          z_channels: 4
+          resolution: 256
+          in_channels: 3
+          out_ch: 3
+          ch: 128
+          ch_mult:
+          - 1
+          - 2
+          - 4
+          - 4
+          num_res_blocks: 2
+          attn_resolutions: []
+          dropout: 0.0
+        lossconfig:
+          target: torch.nn.Identity
+
+    cond_stage_config:
+      target: ldm.modules.encoders.modules.FrozenCLIPImageEmbedder
+
+
+data:
+  target: ldm.data.simple.ObjaverseDataModuleFromConfig
+  params:
+    root_dir: 'views_whole_sphere'
+    batch_size: 192
+    num_workers: 16
+    total_view: 4
+    train:
+      validation: False
+      image_transforms:
+        size: 256
+
+    validation:
+      validation: True
+      image_transforms:
+        size: 256
+
+
+lightning:
+  find_unused_parameters: false
+  metrics_over_trainsteps_checkpoint: True
+  modelcheckpoint:
+    params:
+      every_n_train_steps: 5000
+  callbacks:
+    image_logger:
+      target: main.ImageLogger
+      params:
+        batch_frequency: 500
+        max_images: 32
+        increase_log_steps: False
+        log_first_step: True
+        log_images_kwargs:
+          use_ema_scope: False
+          inpaint: False
+          plot_progressive_rows: False
+          plot_diffusion_rows: False
+          N: 32
+          unconditional_guidance_scale: 3.0
+          unconditional_guidance_label: [""]
+
+  trainer:
+    benchmark: True
+    val_check_interval: 5000000 # really sorry
+    num_sanity_val_steps: 0
+    accumulate_grad_batches: 1

gradio_new.py

@@ -0,0 +1,663 @@
+'''
+conda activate zero123
+cd stable-diffusion
+python gradio_new.py 0
+'''
+
+import diffusers  # 0.12.1
+import math
+import fire
+import gradio as gr
+import lovely_numpy
+import lovely_tensors
+import numpy as np
+import plotly.express as px
+import plotly.graph_objects as go
+import rich
+import sys
+import time
+import torch
+from contextlib import nullcontext
+from diffusers.pipelines.stable_diffusion import StableDiffusionSafetyChecker
+from einops import rearrange
+from functools import partial
+from ldm.models.diffusion.ddim import DDIMSampler
+from ldm.util import create_carvekit_interface, load_and_preprocess, instantiate_from_config
+from lovely_numpy import lo
+from omegaconf import OmegaConf
+from PIL import Image
+from rich import print
+from transformers import AutoFeatureExtractor #, CLIPImageProcessor
+from torch import autocast
+from torchvision import transforms
+
+
+_SHOW_DESC = True
+_SHOW_INTERMEDIATE = False
+# _SHOW_INTERMEDIATE = True
+_GPU_INDEX = 0
+# _GPU_INDEX = 2
+
+# _TITLE = 'Zero-Shot Control of Camera Viewpoints within a Single Image'
+_TITLE = 'Zero-1-to-3: Zero-shot One Image to 3D Object'
+
+# This demo allows you to generate novel viewpoints of an object depicted in an input image using a fine-tuned version of Stable Diffusion.
+_DESCRIPTION = '''
+This demo allows you to control camera rotation and thereby generate novel viewpoints of an object within a single image.
+It is based on Stable Diffusion. Check out our [project webpage](https://zero123.cs.columbia.edu/) and [paper](https://arxiv.org/) if you want to learn more about the method!
+Note that this model is not intended for images of humans or faces, and is unlikely to work well for them.
+'''
+
+_ARTICLE = 'See uses.md'
+
+
+def load_model_from_config(config, ckpt, device, verbose=False):
+    print(f'Loading model from {ckpt}')
+    pl_sd = torch.load(ckpt, map_location=device)
+    if 'global_step' in pl_sd:
+        print(f'Global Step: {pl_sd["global_step"]}')
+    sd = pl_sd['state_dict']
+    model = instantiate_from_config(config.model)
+    m, u = model.load_state_dict(sd, strict=False)
+    if len(m) > 0 and verbose:
+        print('missing keys:')
+        print(m)
+    if len(u) > 0 and verbose:
+        print('unexpected keys:')
+        print(u)
+
+    model.to(device)
+    model.eval()
+    return model
+
+
+@torch.no_grad()
+def sample_model(input_im, model, sampler, precision, h, w, ddim_steps, n_samples, scale,
+                 ddim_eta, x, y, z):
+    precision_scope = autocast if precision == 'autocast' else nullcontext
+    with precision_scope('cuda'):
+        with model.ema_scope():
+            c = model.get_learned_conditioning(input_im).tile(n_samples, 1, 1)
+            T = torch.tensor([math.radians(x), math.sin(
+                math.radians(y)), math.cos(math.radians(y)), z])
+            T = T[None, None, :].repeat(n_samples, 1, 1).to(c.device)
+            c = torch.cat([c, T], dim=-1)
+            c = model.cc_projection(c)
+            cond = {}
+            cond['c_crossattn'] = [c]
+            c_concat = model.encode_first_stage((input_im.to(c.device))).mode().detach()
+            cond['c_concat'] = [model.encode_first_stage((input_im.to(c.device))).mode().detach()
+                                .repeat(n_samples, 1, 1, 1)]
+            if scale != 1.0:
+                uc = {}
+                uc['c_concat'] = [torch.zeros(n_samples, 4, h // 8, w // 8).to(c.device)]
+                uc['c_crossattn'] = [torch.zeros_like(c).to(c.device)]
+            else:
+                uc = None
+
+            shape = [4, h // 8, w // 8]
+            samples_ddim, _ = sampler.sample(S=ddim_steps,
+                                             conditioning=cond,
+                                             batch_size=n_samples,
+                                             shape=shape,
+                                             verbose=False,
+                                             unconditional_guidance_scale=scale,
+                                             unconditional_conditioning=uc,
+                                             eta=ddim_eta,
+                                             x_T=None)
+            print(samples_ddim.shape)
+            # samples_ddim = torch.nn.functional.interpolate(samples_ddim, 64, mode='nearest', antialias=False)
+            x_samples_ddim = model.decode_first_stage(samples_ddim)
+            return torch.clamp((x_samples_ddim + 1.0) / 2.0, min=0.0, max=1.0).cpu()
+
+
+class CameraVisualizer:
+    def __init__(self, gradio_plot):
+        self._gradio_plot = gradio_plot
+        self._fig = None
+        self._polar = 0.0
+        self._azimuth = 0.0
+        self._radius = 0.0
+        self._raw_image = None
+        self._8bit_image = None
+        self._image_colorscale = None
+
+    def polar_change(self, value):
+        self._polar = value
+        # return self.update_figure()
+
+    def azimuth_change(self, value):
+        self._azimuth = value
+        # return self.update_figure()
+
+    def radius_change(self, value):
+        self._radius = value
+        # return self.update_figure()
+
+    def encode_image(self, raw_image):
+        '''
+        :param raw_image (H, W, 3) array of uint8 in [0, 255].
+        '''
+        # https://stackoverflow.com/questions/60685749/python-plotly-how-to-add-an-image-to-a-3d-scatter-plot
+
+        dum_img = Image.fromarray(np.ones((3, 3, 3), dtype='uint8')).convert('P', palette='WEB')
+        idx_to_color = np.array(dum_img.getpalette()).reshape((-1, 3))
+
+        self._raw_image = raw_image
+        self._8bit_image = Image.fromarray(raw_image).convert('P', palette='WEB', dither=None)
+        # self._8bit_image = Image.fromarray(raw_image.clip(0, 254)).convert(
+        #     'P', palette='WEB', dither=None)
+        self._image_colorscale = [
+            [i / 255.0, 'rgb({}, {}, {})'.format(*rgb)] for i, rgb in enumerate(idx_to_color)]
+
+        # return self.update_figure()
+
+    def update_figure(self):
+        fig = go.Figure()
+
+        if self._raw_image is not None:
+            (H, W, C) = self._raw_image.shape
+
+            x = np.zeros((H, W))
+            (y, z) = np.meshgrid(np.linspace(-1.0, 1.0, W), np.linspace(1.0, -1.0, H) * H / W)
+            print('x:', lo(x))
+            print('y:', lo(y))
+            print('z:', lo(z))
+
+            fig.add_trace(go.Surface(
+                x=x, y=y, z=z,
+                surfacecolor=self._8bit_image,
+                cmin=0,
+                cmax=255,
+                colorscale=self._image_colorscale,
+                showscale=False,
+                lighting_diffuse=1.0,
+                lighting_ambient=1.0,
+                lighting_fresnel=1.0,
+                lighting_roughness=1.0,
+                lighting_specular=0.3))
+
+            scene_bounds = 3.5
+            base_radius = 2.5
+            zoom_scale = 1.5  # Note that input radius offset is in [-0.5, 0.5].
+            fov_deg = 50.0
+            edges = [(0, 1), (0, 2), (0, 3), (0, 4), (1, 2), (2, 3), (3, 4), (4, 1)]
+
+            input_cone = calc_cam_cone_pts_3d(
+                0.0, 0.0, base_radius, fov_deg)  # (5, 3).
+            output_cone = calc_cam_cone_pts_3d(
+                self._polar, self._azimuth, base_radius + self._radius * zoom_scale, fov_deg)  # (5, 3).
+            # print('input_cone:', lo(input_cone).v)
+            # print('output_cone:', lo(output_cone).v)
+
+            for (cone, clr, legend) in [(input_cone, 'green', 'Input view'),
+                                        (output_cone, 'blue', 'Target view')]:
+
+                for (i, edge) in enumerate(edges):
+                    (x1, x2) = (cone[edge[0], 0], cone[edge[1], 0])
+                    (y1, y2) = (cone[edge[0], 1], cone[edge[1], 1])
+                    (z1, z2) = (cone[edge[0], 2], cone[edge[1], 2])
+                    fig.add_trace(go.Scatter3d(
+                        x=[x1, x2], y=[y1, y2], z=[z1, z2], mode='lines',
+                        line=dict(color=clr, width=3),
+                        name=legend, showlegend=(i == 0)))
+                    # text=(legend if i == 0 else None),
+                    # textposition='bottom center'))
+                    # hoverinfo='text',
+                    # hovertext='hovertext'))
+
+                # Add label.
+                if cone[0, 2] <= base_radius / 2.0:
+                    fig.add_trace(go.Scatter3d(
+                        x=[cone[0, 0]], y=[cone[0, 1]], z=[cone[0, 2] - 0.05], showlegend=False,
+                        mode='text', text=legend, textposition='bottom center'))
+                else:
+                    fig.add_trace(go.Scatter3d(
+                        x=[cone[0, 0]], y=[cone[0, 1]], z=[cone[0, 2] + 0.05], showlegend=False,
+                        mode='text', text=legend, textposition='top center'))
+
+            # look at center of scene
+            fig.update_layout(
+                # width=640,
+                # height=480,
+                # height=400,
+                height=360,
+                autosize=True,
+                hovermode=False,
+                margin=go.layout.Margin(l=0, r=0, b=0, t=0),
+                showlegend=True,
+                legend=dict(
+                    yanchor='bottom',
+                    y=0.01,
+                    xanchor='right',
+                    x=0.99,
+                ),
+                scene=dict(
+                    aspectmode='manual',
+                    aspectratio=dict(x=1, y=1, z=1.0),
+                    camera=dict(
+                        eye=dict(x=base_radius - 1.6, y=0.0, z=0.6),
+                        center=dict(x=0.0, y=0.0, z=0.0),
+                        up=dict(x=0.0, y=0.0, z=1.0)),
+                    xaxis_title='',
+                    yaxis_title='',
+                    zaxis_title='',
+                    xaxis=dict(
+                        range=[-scene_bounds, scene_bounds],
+                        showticklabels=False,
+                        showgrid=True,
+                        zeroline=False,
+                        showbackground=True,
+                        showspikes=False,
+                        showline=False,
+                        ticks=''),
+                    yaxis=dict(
+                        range=[-scene_bounds, scene_bounds],
+                        showticklabels=False,
+                        showgrid=True,
+                        zeroline=False,
+                        showbackground=True,
+                        showspikes=False,
+                        showline=False,
+                        ticks=''),
+                    zaxis=dict(
+                        range=[-scene_bounds, scene_bounds],
+                        showticklabels=False,
+                        showgrid=True,
+                        zeroline=False,
+                        showbackground=True,
+                        showspikes=False,
+                        showline=False,
+                        ticks='')))
+
+        self._fig = fig
+        return fig
+
+
+def preprocess_image(models, input_im, preprocess):
+    '''
+    :param input_im (PIL Image).
+    :return input_im (H, W, 3) array in [0, 1].
+    '''
+
+    print('old input_im:', input_im.size)
+    start_time = time.time()
+
+    if preprocess:
+        input_im = load_and_preprocess(models['carvekit'], input_im)
+        input_im = (input_im / 255.0).astype(np.float32)
+        # (H, W, 3) array in [0, 1].
+
+    else:
+        input_im = input_im.resize([256, 256], Image.Resampling.LANCZOS)
+        input_im = np.asarray(input_im, dtype=np.float32) / 255.0
+        # (H, W, 4) array in [0, 1].
+
+        # old method: thresholding background, very important
+        # input_im[input_im[:, :, -1] <= 0.9] = [1., 1., 1., 1.]
+
+        # new method: apply correct method of compositing to avoid sudden transitions / thresholding
+        # (smoothly transition foreground to white background based on alpha values)
+        alpha = input_im[:, :, 3:4]
+        white_im = np.ones_like(input_im)
+        input_im = alpha * input_im + (1.0 - alpha) * white_im
+
+        input_im = input_im[:, :, 0:3]
+        # (H, W, 3) array in [0, 1].
+
+    print(f'Infer foreground mask (preprocess_image) took {time.time() - start_time:.3f}s.')
+    print('new input_im:', lo(input_im))
+
+    return input_im
+
+
+def main_run(models, device, cam_vis, return_what,
+             x=0.0, y=0.0, z=0.0,
+             raw_im=None, preprocess=True,
+             scale=3.0, n_samples=4, ddim_steps=50, ddim_eta=1.0,
+             precision='fp32', h=256, w=256):
+    '''
+    :param raw_im (PIL Image).
+    '''
+    
+    safety_checker_input = models['clip_fe'](raw_im, return_tensors='pt').to(device)
+    (image, has_nsfw_concept) = models['nsfw'](
+        images=np.ones((1, 3)), clip_input=safety_checker_input.pixel_values)
+    print('has_nsfw_concept:', has_nsfw_concept)
+    if np.any(has_nsfw_concept):
+        print('NSFW content detected.')
+        to_return = [None] * 10
+        description = ('###  <span style="color:red"> Unfortunately, '
+                       'potential NSFW content was detected, '
+                       'which is not supported by our model. '
+                       'Please try again with a different image. </span>')
+        if 'angles' in return_what:
+            to_return[0] = 0.0
+            to_return[1] = 0.0
+            to_return[2] = 0.0
+            to_return[3] = description
+        else:
+            to_return[0] = description
+        return to_return
+
+    else:
+        print('Safety check passed.')
+
+    input_im = preprocess_image(models, raw_im, preprocess)
+
+    # if np.random.rand() < 0.3:
+    #     description = ('Unfortunately, a human, a face, or potential NSFW content was detected, '
+    #                    'which is not supported by our model.')
+    #     if vis_only:
+    #         return (None, None, description)
+    #     else:
+    #         return (None, None, None, description)
+
+    show_in_im1 = (input_im * 255.0).astype(np.uint8)
+    show_in_im2 = Image.fromarray(show_in_im1)
+
+    if 'rand' in return_what:
+        x = int(np.round(np.arcsin(np.random.uniform(-1.0, 1.0)) * 160.0 / np.pi))  # [-80, 80].
+        y = int(np.round(np.random.uniform(-150.0, 150.0)))
+        z = 0.0
+
+    cam_vis.polar_change(x)
+    cam_vis.azimuth_change(y)
+    cam_vis.radius_change(z)
+    cam_vis.encode_image(show_in_im1)
+    new_fig = cam_vis.update_figure()
+
+    if 'vis' in return_what:
+        description = ('The viewpoints are visualized on the top right. '
+                       'Click Run Generation to update the results on the bottom right.')
+
+        if 'angles' in return_what:
+            return (x, y, z, description, new_fig, show_in_im2)
+        else:
+            return (description, new_fig, show_in_im2)
+
+    elif 'gen' in return_what:
+        input_im = transforms.ToTensor()(input_im).unsqueeze(0).to(device)
+        input_im = input_im * 2 - 1
+        input_im = transforms.functional.resize(input_im, [h, w])
+
+        sampler = DDIMSampler(models['turncam'])
+        # used_x = -x  # NOTE: Polar makes more sense in Basile's opinion this way!
+        used_x = x  # NOTE: Set this way for consistency.
+        x_samples_ddim = sample_model(input_im, models['turncam'], sampler, precision, h, w,
+                                      ddim_steps, n_samples, scale, ddim_eta, used_x, y, z)
+
+        output_ims = []
+        for x_sample in x_samples_ddim:
+            x_sample = 255.0 * rearrange(x_sample.cpu().numpy(), 'c h w -> h w c')
+            output_ims.append(Image.fromarray(x_sample.astype(np.uint8)))
+
+        description = None
+
+        if 'angles' in return_what:
+            return (x, y, z, description, new_fig, show_in_im2, output_ims)
+        else:
+            return (description, new_fig, show_in_im2, output_ims)
+
+
+def calc_cam_cone_pts_3d(polar_deg, azimuth_deg, radius_m, fov_deg):
+    '''
+    :param polar_deg (float).
+    :param azimuth_deg (float).
+    :param radius_m (float).
+    :param fov_deg (float).
+    :return (5, 3) array of float with (x, y, z).
+    '''
+    polar_rad = np.deg2rad(polar_deg)
+    azimuth_rad = np.deg2rad(azimuth_deg)
+    fov_rad = np.deg2rad(fov_deg)
+    polar_rad = -polar_rad  # NOTE: Inverse of how used_x relates to x.
+
+    # Camera pose center:
+    cam_x = radius_m * np.cos(azimuth_rad) * np.cos(polar_rad)
+    cam_y = radius_m * np.sin(azimuth_rad) * np.cos(polar_rad)
+    cam_z = radius_m * np.sin(polar_rad)
+
+    # Obtain four corners of camera frustum, assuming it is looking at origin.
+    # First, obtain camera extrinsics (rotation matrix only):
+    camera_R = np.array([[np.cos(azimuth_rad) * np.cos(polar_rad),
+                          -np.sin(azimuth_rad),
+                          -np.cos(azimuth_rad) * np.sin(polar_rad)],
+                         [np.sin(azimuth_rad) * np.cos(polar_rad),
+                          np.cos(azimuth_rad),
+                          -np.sin(azimuth_rad) * np.sin(polar_rad)],
+                         [np.sin(polar_rad),
+                          0.0,
+                          np.cos(polar_rad)]])
+    # print('camera_R:', lo(camera_R).v)
+
+    # Multiply by corners in camera space to obtain go to space:
+    corn1 = [-1.0, np.tan(fov_rad / 2.0), np.tan(fov_rad / 2.0)]
+    corn2 = [-1.0, -np.tan(fov_rad / 2.0), np.tan(fov_rad / 2.0)]
+    corn3 = [-1.0, -np.tan(fov_rad / 2.0), -np.tan(fov_rad / 2.0)]
+    corn4 = [-1.0, np.tan(fov_rad / 2.0), -np.tan(fov_rad / 2.0)]
+    corn1 = np.dot(camera_R, corn1)
+    corn2 = np.dot(camera_R, corn2)
+    corn3 = np.dot(camera_R, corn3)
+    corn4 = np.dot(camera_R, corn4)
+
+    # Now attach as offset to actual 3D camera position:
+    corn1 = np.array(corn1) / np.linalg.norm(corn1, ord=2)
+    corn_x1 = cam_x + corn1[0]
+    corn_y1 = cam_y + corn1[1]
+    corn_z1 = cam_z + corn1[2]
+    corn2 = np.array(corn2) / np.linalg.norm(corn2, ord=2)
+    corn_x2 = cam_x + corn2[0]
+    corn_y2 = cam_y + corn2[1]
+    corn_z2 = cam_z + corn2[2]
+    corn3 = np.array(corn3) / np.linalg.norm(corn3, ord=2)
+    corn_x3 = cam_x + corn3[0]
+    corn_y3 = cam_y + corn3[1]
+    corn_z3 = cam_z + corn3[2]
+    corn4 = np.array(corn4) / np.linalg.norm(corn4, ord=2)
+    corn_x4 = cam_x + corn4[0]
+    corn_y4 = cam_y + corn4[1]
+    corn_z4 = cam_z + corn4[2]
+
+    xs = [cam_x, corn_x1, corn_x2, corn_x3, corn_x4]
+    ys = [cam_y, corn_y1, corn_y2, corn_y3, corn_y4]
+    zs = [cam_z, corn_z1, corn_z2, corn_z3, corn_z4]
+
+    return np.array([xs, ys, zs]).T
+
+
+def run_demo(
+        device_idx=_GPU_INDEX,
+        ckpt='105000.ckpt',
+        config='configs/sd-objaverse-finetune-c_concat-256.yaml'):
+
+    print('sys.argv:', sys.argv)
+    if len(sys.argv) > 1:
+        print('old device_idx:', device_idx)
+        device_idx = int(sys.argv[1])
+        print('new device_idx:', device_idx)
+
+    device = f'cuda:{device_idx}'
+    config = OmegaConf.load(config)
+
+    # Instantiate all models beforehand for efficiency.
+    models = dict()
+    print('Instantiating LatentDiffusion...')
+    models['turncam'] = load_model_from_config(config, ckpt, device=device)
+    print('Instantiating Carvekit HiInterface...')
+    models['carvekit'] = create_carvekit_interface()
+    print('Instantiating StableDiffusionSafetyChecker...')
+    models['nsfw'] = StableDiffusionSafetyChecker.from_pretrained(
+        'CompVis/stable-diffusion-safety-checker').to(device)
+    print('Instantiating AutoFeatureExtractor...')
+    models['clip_fe'] = AutoFeatureExtractor.from_pretrained(
+        'CompVis/stable-diffusion-safety-checker')
+
+    # Reduce NSFW false positives.
+    # NOTE: At the time of writing, and for diffusers 0.12.1, the default parameters are:
+    # models['nsfw'].concept_embeds_weights:
+    # [0.1800, 0.1900, 0.2060, 0.2100, 0.1950, 0.1900, 0.1940, 0.1900, 0.1900, 0.2200, 0.1900,
+    #  0.1900, 0.1950, 0.1984, 0.2100, 0.2140, 0.2000].
+    # models['nsfw'].special_care_embeds_weights:
+    # [0.1950, 0.2000, 0.2200].
+    # We multiply all by some factor > 1 to make them less likely to be triggered.
+    models['nsfw'].concept_embeds_weights *= 1.07
+    models['nsfw'].special_care_embeds_weights *= 1.07
+
+    with open('instructions.md', 'r') as f:
+        article = f.read()
+
+    # Compose demo layout & data flow.
+    demo = gr.Blocks(title=_TITLE)
+
+    with demo:
+        gr.Markdown('# ' + _TITLE)
+        gr.Markdown(_DESCRIPTION)
+
+        with gr.Row():
+            with gr.Column(scale=0.9, variant='panel'):
+
+                image_block = gr.Image(type='pil', image_mode='RGBA',
+                                       label='Input image of single object')
+                preprocess_chk = gr.Checkbox(
+                    True, label='Preprocess image automatically (remove background and recenter object)')
+                # info='If enabled, the uploaded image will be preprocessed to remove the background and recenter the object by cropping and/or padding as necessary. '
+                # 'If disabled, the image will be used as-is, *BUT* a fully transparent or white background is required.'),
+
+                gr.Markdown('*Try camera position presets:*')
+                with gr.Row():
+                    left_btn = gr.Button('View from the Left', variant='primary')
+                    above_btn = gr.Button('View from Above', variant='primary')
+                    right_btn = gr.Button('View from the Right', variant='primary')
+                with gr.Row():
+                    random_btn = gr.Button('Random Rotation', variant='primary')
+                    below_btn = gr.Button('View from Below', variant='primary')
+                    behind_btn = gr.Button('View from Behind', variant='primary')
+
+                gr.Markdown('*Control camera position manually:*')
+                polar_slider = gr.Slider(
+                    -90, 90, value=0, step=5, label='Polar angle (vertical rotation in degrees)')
+                # info='Positive values move the camera down, while negative values move the camera up.')
+                azimuth_slider = gr.Slider(
+                    -180, 180, value=0, step=5, label='Azimuth angle (horizontal rotation in degrees)')
+                # info='Positive values move the camera right, while negative values move the camera left.')
+                radius_slider = gr.Slider(
+                    -0.5, 0.5, value=0.0, step=0.1, label='Zoom (relative distance from center)')
+                # info='Positive values move the camera further away, while negative values move the camera closer.')
+
+                samples_slider = gr.Slider(1, 8, value=4, step=1,
+                                           label='Number of samples to generate')
+
+                with gr.Accordion('Advanced options', open=False):
+                    scale_slider = gr.Slider(0, 30, value=3, step=1,
+                                             label='Diffusion guidance scale')
+                    steps_slider = gr.Slider(5, 200, value=75, step=5,
+                                             label='Number of diffusion inference steps')
+
+                with gr.Row():
+                    vis_btn = gr.Button('Visualize Angles', variant='secondary')
+                    run_btn = gr.Button('Run Generation', variant='primary')
+
+                desc_output = gr.Markdown('The results will appear on the right.', visible=_SHOW_DESC)
+
+            with gr.Column(scale=1.1, variant='panel'):
+
+                vis_output = gr.Plot(
+                    label='Relationship between input (green) and output (blue) camera poses')
+
+                gen_output = gr.Gallery(label='Generated images from specified new viewpoint')
+                gen_output.style(grid=2)
+
+                preproc_output = gr.Image(type='pil', image_mode='RGB',
+                                          label='Preprocessed input image', visible=_SHOW_INTERMEDIATE)
+
+        gr.Markdown(article)
+
+        # NOTE: I am forced to update vis_output for these preset buttons,
+        # because otherwise the gradio plot always resets the plotly 3D viewpoint for some reason,
+        # which might confuse the user into thinking that the plot has been updated too.
+
+        # OLD 1:
+        # left_btn.click(fn=lambda: [0.0, -90.0], #, 0.0],
+        #                inputs=[], outputs=[polar_slider, azimuth_slider]), #], radius_slider])
+        # above_btn.click(fn=lambda: [90.0, 0.0], #, 0.0],
+        #                 inputs=[], outputs=[polar_slider, azimuth_slider]), #, radius_slider])
+        # right_btn.click(fn=lambda: [0.0, 90.0], #, 0.0],
+        #                 inputs=[], outputs=[polar_slider, azimuth_slider]), #, radius_slider])
+        # random_btn.click(fn=lambda: [int(np.round(np.random.uniform(-60.0, 60.0))),
+        #                              int(np.round(np.random.uniform(-150.0, 150.0)))], #, 0.0],
+        #                 inputs=[], outputs=[polar_slider, azimuth_slider]), #, radius_slider])
+        # below_btn.click(fn=lambda: [-90.0, 0.0], #, 0.0],
+        #                 inputs=[], outputs=[polar_slider, azimuth_slider]), #, radius_slider])
+        # behind_btn.click(fn=lambda: [0.0, 180.0], #, 0.0],
+        #                  inputs=[], outputs=[polar_slider, azimuth_slider]), #, radius_slider])
+
+        # OLD 2:
+        # preset_text = ('You have selected a preset target camera view. '
+        #                'Now click Run Generation to update the results!')
+
+        # left_btn.click(fn=lambda: [0.0, -90.0, None, preset_text],
+        #                inputs=[], outputs=[polar_slider, azimuth_slider, vis_output, desc_output])
+        # above_btn.click(fn=lambda: [90.0, 0.0, None, preset_text],
+        #                 inputs=[], outputs=[polar_slider, azimuth_slider, vis_output, desc_output])
+        # right_btn.click(fn=lambda: [0.0, 90.0, None, preset_text],
+        #                 inputs=[], outputs=[polar_slider, azimuth_slider, vis_output, desc_output])
+        # random_btn.click(fn=lambda: [int(np.round(np.random.uniform(-60.0, 60.0))),
+        #                              int(np.round(np.random.uniform(-150.0, 150.0))),
+        #                              None, preset_text],
+        #                 inputs=[], outputs=[polar_slider, azimuth_slider, vis_output, desc_output])
+        # below_btn.click(fn=lambda: [-90.0, 0.0, None, preset_text],
+        #                 inputs=[], outputs=[polar_slider, azimuth_slider, vis_output, desc_output])
+        # behind_btn.click(fn=lambda: [0.0, 180.0, None, preset_text],
+        #                  inputs=[], outputs=[polar_slider, azimuth_slider, vis_output, desc_output])
+
+        # OLD 3 (does not work at all):
+        # def a():
+        #     polar_slider.value = 77.7
+        #     polar_slider.postprocess(77.7)
+        #     print('testa')
+        # left_btn.click(fn=a)
+
+        cam_vis = CameraVisualizer(vis_output)
+
+        vis_btn.click(fn=partial(main_run, models, device, cam_vis, 'vis'),
+                      inputs=[polar_slider, azimuth_slider, radius_slider,
+                              image_block, preprocess_chk],
+                      outputs=[desc_output, vis_output, preproc_output])
+
+        run_btn.click(fn=partial(main_run, models, device, cam_vis, 'gen'),
+                      inputs=[polar_slider, azimuth_slider, radius_slider,
+                              image_block, preprocess_chk,
+                              scale_slider, samples_slider, steps_slider],
+                      outputs=[desc_output, vis_output, preproc_output, gen_output])
+
+        # NEW:
+        preset_inputs = [image_block, preprocess_chk,
+                         scale_slider, samples_slider, steps_slider]
+        preset_outputs = [polar_slider, azimuth_slider, radius_slider,
+                          desc_output, vis_output, preproc_output, gen_output]
+        left_btn.click(fn=partial(main_run, models, device, cam_vis, 'angles_gen',
+                                  0.0, -90.0, 0.0),
+                       inputs=preset_inputs, outputs=preset_outputs)
+        above_btn.click(fn=partial(main_run, models, device, cam_vis, 'angles_gen',
+                                   -90.0, 0.0, 0.0),
+                       inputs=preset_inputs, outputs=preset_outputs)
+        right_btn.click(fn=partial(main_run, models, device, cam_vis, 'angles_gen',
+                                   0.0, 90.0, 0.0),
+                       inputs=preset_inputs, outputs=preset_outputs)
+        random_btn.click(fn=partial(main_run, models, device, cam_vis, 'rand_angles_gen',
+                                    -1.0, -1.0, -1.0),
+                       inputs=preset_inputs, outputs=preset_outputs)
+        below_btn.click(fn=partial(main_run, models, device, cam_vis, 'angles_gen',
+                                   90.0, 0.0, 0.0),
+                       inputs=preset_inputs, outputs=preset_outputs)
+        behind_btn.click(fn=partial(main_run, models, device, cam_vis, 'angles_gen',
+                                    0.0, 180.0, 0.0),
+                       inputs=preset_inputs, outputs=preset_outputs)
+
+    demo.launch(enable_queue=True, share=True)
+
+
+if __name__ == '__main__':
+
+    fire.Fire(run_demo)

gradio_objaverse.py

@@ -0,0 +1,184 @@
+from contextlib import nullcontext
+from functools import partial
+
+import math
+import fire
+import gradio as gr
+import numpy as np
+import torch
+from einops import rearrange
+from ldm.models.diffusion.ddim import DDIMSampler
+from omegaconf import OmegaConf
+from PIL import Image
+from torch import autocast
+from torchvision import transforms
+from ldm.util import load_and_preprocess, instantiate_from_config
+
+def load_model_from_config(config, ckpt, device, verbose=False):
+    print(f"Loading model from {ckpt}")
+    pl_sd = torch.load(ckpt, map_location=device)
+    if "global_step" in pl_sd:
+        print(f"Global Step: {pl_sd['global_step']}")
+    sd = pl_sd["state_dict"]
+    model = instantiate_from_config(config.model)
+    m, u = model.load_state_dict(sd, strict=False)
+    if len(m) > 0 and verbose:
+        print("missing keys:")
+        print(m)
+    if len(u) > 0 and verbose:
+        print("unexpected keys:")
+        print(u)
+
+    model.to(device)
+    model.eval()
+    return model
+
+@torch.no_grad()
+def sample_model(input_im, model, sampler, precision, h, w, ddim_steps, n_samples, scale, \
+                 ddim_eta, x, y, z):
+    precision_scope = autocast if precision=="autocast" else nullcontext
+    with precision_scope("cuda"):
+        with model.ema_scope():
+            c = model.get_learned_conditioning(input_im).tile(n_samples,1,1)
+            T = torch.tensor([math.radians(x), math.sin(math.radians(y)), math.cos(math.radians(y)), z])
+            T = T[None, None, :].repeat(n_samples, 1, 1).to(c.device)
+            c = torch.cat([c, T], dim=-1)
+            c = model.cc_projection(c)
+            cond = {}
+            cond['c_crossattn'] = [c]
+            c_concat = model.encode_first_stage((input_im.to(c.device))).mode().detach()
+            cond['c_concat'] = [model.encode_first_stage((input_im.to(c.device))).mode().detach()\
+                               .repeat(n_samples, 1, 1, 1)]
+            if scale != 1.0:
+                uc = {}
+                uc['c_concat'] = [torch.zeros(n_samples, 4, h // 8, w // 8).to(c.device)]
+                uc['c_crossattn'] = [torch.zeros_like(c).to(c.device)]
+            else:
+                uc = None
+
+            shape = [4, h // 8, w // 8]
+            samples_ddim, _ = sampler.sample(S=ddim_steps,
+                                             conditioning=cond,
+                                             batch_size=n_samples,
+                                             shape=shape,
+                                             verbose=False,
+                                             unconditional_guidance_scale=scale,
+                                             unconditional_conditioning=uc,
+                                             eta=ddim_eta,
+                                             x_T=None)
+            print(samples_ddim.shape)
+            # samples_ddim = torch.nn.functional.interpolate(samples_ddim, 64, mode='nearest', antialias=False)
+            x_samples_ddim = model.decode_first_stage(samples_ddim)
+            return torch.clamp((x_samples_ddim + 1.0) / 2.0, min=0.0, max=1.0).cpu()
+
+
+def main(
+    model,
+    device,
+    input_im,
+    x=0.,
+    y=0.,
+    z=0.,
+    scale=3.0,
+    n_samples=4,
+    ddim_steps=50,
+    preprocess=True,
+    ddim_eta=1.0,
+    precision="fp32",
+    h=256,
+    w=256,
+    ):
+    # input_im[input_im == [0., 0., 0.]] = [1., 1., 1., 1.]
+    print(input_im.size)
+    if preprocess:
+        input_im = load_and_preprocess(input_im)
+    else:
+        input_im = input_im.resize([256, 256], Image.Resampling.LANCZOS)
+        input_im = np.asarray(input_im, dtype=np.float32) / 255.
+        input_im[input_im[:, :, -1] <= 0.9] = [1., 1., 1., 1.] # very important, thresholding background
+        input_im = input_im[:, :, :3]
+    print(input_im.shape)
+    input_im = transforms.ToTensor()(input_im).unsqueeze(0).to(device)
+    input_im = input_im * 2 - 1
+    input_im = transforms.functional.resize(input_im, [h, w])
+
+    sampler = DDIMSampler(model)
+
+    x_samples_ddim = sample_model(input_im, model, sampler, precision, h, w,\
+                                  ddim_steps, n_samples, scale, ddim_eta, x, y, z)
+    output_ims = []
+    for x_sample in x_samples_ddim:
+        x_sample = 255. * rearrange(x_sample.cpu().numpy(), 'c h w -> h w c')
+        output_ims.append(Image.fromarray(x_sample.astype(np.uint8)))
+    return output_ims
+
+
+description = \
+"""Generate variations on an input image using a fine-tuned version of Stable Diffision.
+Trained by [Justin Pinkney](https://www.justinpinkney.com) ([@Buntworthy](https://twitter.com/Buntworthy)) at [Lambda](https://lambdalabs.com/)
+__Get the [code](https://github.com/justinpinkney/stable-diffusion) and [model](https://huggingface.co/lambdalabs/stable-diffusion-image-conditioned).__
+![](https://raw.githubusercontent.com/justinpinkney/stable-diffusion/main/assets/im-vars-thin.jpg)
+"""
+
+article = \
+"""
+## How does this work?
+The normal Stable Diffusion model is trained to be conditioned on text input. This version has had the original text encoder (from CLIP) removed, and replaced with
+the CLIP _image_ encoder instead. So instead of generating images based a text input, images are generated to match CLIP's embedding of the image.
+This creates images which have the same rough style and content, but different details, in particular the composition is generally quite different.
+This is a totally different approach to the img2img script of the original Stable Diffusion and gives very different results.
+The model was fine tuned on the [LAION aethetics v2 6+ dataset](https://laion.ai/blog/laion-aesthetics/) to accept the new conditioning.
+Training was done on 4xA6000 GPUs on [Lambda GPU Cloud](https://lambdalabs.com/service/gpu-cloud).
+More details on the method and training will come in a future blog post.
+"""
+
+
+def run_demo(
+    device_idx=0,
+    ckpt="last.ckpt",
+    config="configs/sd-objaverse-finetune-c_concat-256.yaml",
+    ):
+
+    device = f"cuda:{device_idx}"
+    config = OmegaConf.load(config)
+    model = load_model_from_config(config, ckpt, device=device)
+
+    inputs = [
+        gr.Image(type="pil", image_mode="RGBA"), # shape=[512, 512]
+        gr.Number(label="polar (between axis z+)"),
+        gr.Number(label="azimuth (between axis x+)"),
+        gr.Number(label="z (distance from center)"),
+        gr.Slider(0, 100, value=3, step=1, label="cfg scale"),
+        gr.Slider(1, 8, value=4, step=1, label="Number images"),
+        gr.Slider(5, 200, value=100, step=5, label="steps"),
+        gr.Checkbox(True, label="image preprocess (background removal and recenter)"),
+    ]
+    output = gr.Gallery(label="Generated variations")
+    output.style(grid=2)
+
+    fn_with_model = partial(main, model, device)
+    fn_with_model.__name__ = "fn_with_model"
+
+    examples = [
+        # ["assets/zero-shot/bear.png", 0, 0, 0, 3, 4, 100],
+        # ["assets/zero-shot/car.png", 0, 0, 0, 3, 4, 100],
+        # ["assets/zero-shot/elephant.png", 0, 0, 0, 3, 4, 100],
+        # ["assets/zero-shot/pikachu.png", 0, 0, 0, 3, 4, 100],
+        # ["assets/zero-shot/spyro.png", 0, 0, 0, 3, 4, 100],
+        # ["assets/zero-shot/taxi.png", 0, 0, 0, 3, 4, 100],
+    ]
+
+    demo = gr.Interface(
+        fn=fn_with_model,
+        title="Stable Diffusion Novel View Synthesis (Image)",
+        # description=description,
+        # article=article,
+        inputs=inputs,
+        outputs=output,
+        examples=examples,
+        allow_flagging="never",
+        )
+    demo.launch(enable_queue=True, share=True)
+
+if __name__ == "__main__":
+    fire.Fire(run_demo)

instructions.md

@@ -0,0 +1,11 @@
+## Tips if not getting expected results :detective:
+1. If you are trying out images of humans, especially faces, note that it's unfortunately not the intended use cases. We would encourage to try out images of everyday objects or even artworks.
+2. If some part of the object is missing, check the interactive angle visualization pane (top right) where you can find a panel of the actual input image to the model after preprocessing steps and see if the segmented image contains the entire object you are trying to visualize.
+3. The model is probabilistic, therefore, if the number of samples is selected to be bigger than 1 and results look different, that's expected as the model tries to predict a diverse set of possibilities given the input image and the specified camera viewpoint.
+4. Under "advanced options", you can tune two parameters as you can typically find in other stable diffusion demos as well:
+	 - Diffusion Guidance Scale defines how much you want the model to respect the input information (image + angles). Higher scale typically leads to less diversity and higher image distortion.
+	 - Number of diffusion inference steps controls the number of diffusion steps is applied to generate each image. Usually the higher the better with a diminishing return.
+
+Have fun!
+
+A model card can be found here: uses.md

ldm/data/base.py

@@ -0,0 +1,40 @@
+import os
+import numpy as np
+from abc import abstractmethod
+from torch.utils.data import Dataset, ConcatDataset, ChainDataset, IterableDataset
+
+
+class Txt2ImgIterableBaseDataset(IterableDataset):
+    '''
+    Define an interface to make the IterableDatasets for text2img data chainable
+    '''
+    def __init__(self, num_records=0, valid_ids=None, size=256):
+        super().__init__()
+        self.num_records = num_records
+        self.valid_ids = valid_ids
+        self.sample_ids = valid_ids
+        self.size = size
+
+        print(f'{self.__class__.__name__} dataset contains {self.__len__()} examples.')
+
+    def __len__(self):
+        return self.num_records
+
+    @abstractmethod
+    def __iter__(self):
+        pass
+
+
+class PRNGMixin(object):
+    """
+    Adds a prng property which is a numpy RandomState which gets
+    reinitialized whenever the pid changes to avoid synchronized sampling
+    behavior when used in conjunction with multiprocessing.
+    """
+    @property
+    def prng(self):
+        currentpid = os.getpid()
+        if getattr(self, "_initpid", None) != currentpid:
+            self._initpid = currentpid
+            self._prng = np.random.RandomState()
+        return self._prng

ldm/data/coco.py

@@ -0,0 +1,253 @@
+import os
+import json
+import albumentations
+import numpy as np
+from PIL import Image
+from tqdm import tqdm
+from torch.utils.data import Dataset
+from abc import abstractmethod
+
+
+class CocoBase(Dataset):
+    """needed for (image, caption, segmentation) pairs"""
+    def __init__(self, size=None, dataroot="", datajson="", onehot_segmentation=False, use_stuffthing=False,
+                 crop_size=None, force_no_crop=False, given_files=None, use_segmentation=True,crop_type=None):
+        self.split = self.get_split()
+        self.size = size
+        if crop_size is None:
+            self.crop_size = size
+        else:
+            self.crop_size = crop_size
+
+        assert crop_type in [None, 'random', 'center']
+        self.crop_type = crop_type
+        self.use_segmenation = use_segmentation
+        self.onehot = onehot_segmentation       # return segmentation as rgb or one hot
+        self.stuffthing = use_stuffthing        # include thing in segmentation
+        if self.onehot and not self.stuffthing:
+            raise NotImplemented("One hot mode is only supported for the "
+                                 "stuffthings version because labels are stored "
+                                 "a bit different.")
+
+        data_json = datajson
+        with open(data_json) as json_file:
+            self.json_data = json.load(json_file)
+            self.img_id_to_captions = dict()
+            self.img_id_to_filepath = dict()
+            self.img_id_to_segmentation_filepath = dict()
+
+        assert data_json.split("/")[-1] in [f"captions_train{self.year()}.json",
+                                            f"captions_val{self.year()}.json"]
+        # TODO currently hardcoded paths, would be better to follow logic in
+        # cocstuff pixelmaps
+        if self.use_segmenation:
+            if self.stuffthing:
+                self.segmentation_prefix = (
+                    f"data/cocostuffthings/val{self.year()}" if
+                    data_json.endswith(f"captions_val{self.year()}.json") else
+                    f"data/cocostuffthings/train{self.year()}")
+            else:
+                self.segmentation_prefix = (
+                    f"data/coco/annotations/stuff_val{self.year()}_pixelmaps" if
+                    data_json.endswith(f"captions_val{self.year()}.json") else
+                    f"data/coco/annotations/stuff_train{self.year()}_pixelmaps")
+
+        imagedirs = self.json_data["images"]
+        self.labels = {"image_ids": list()}
+        for imgdir in tqdm(imagedirs, desc="ImgToPath"):
+            self.img_id_to_filepath[imgdir["id"]] = os.path.join(dataroot, imgdir["file_name"])
+            self.img_id_to_captions[imgdir["id"]] = list()
+            pngfilename = imgdir["file_name"].replace("jpg", "png")
+            if self.use_segmenation:
+                self.img_id_to_segmentation_filepath[imgdir["id"]] = os.path.join(
+                    self.segmentation_prefix, pngfilename)
+            if given_files is not None:
+                if pngfilename in given_files:
+                    self.labels["image_ids"].append(imgdir["id"])
+            else:
+                self.labels["image_ids"].append(imgdir["id"])
+
+        capdirs = self.json_data["annotations"]
+        for capdir in tqdm(capdirs, desc="ImgToCaptions"):
+            # there are in average 5 captions per image
+            #self.img_id_to_captions[capdir["image_id"]].append(np.array([capdir["caption"]]))
+            self.img_id_to_captions[capdir["image_id"]].append(capdir["caption"])
+
+        self.rescaler = albumentations.SmallestMaxSize(max_size=self.size)
+        if self.split=="validation":
+            self.cropper = albumentations.CenterCrop(height=self.crop_size, width=self.crop_size)
+        else:
+            # default option for train is random crop
+            if self.crop_type in [None, 'random']:
+                self.cropper = albumentations.RandomCrop(height=self.crop_size, width=self.crop_size)
+            else:
+                self.cropper = albumentations.CenterCrop(height=self.crop_size, width=self.crop_size)
+        self.preprocessor = albumentations.Compose(
+            [self.rescaler, self.cropper],
+            additional_targets={"segmentation": "image"})
+        if force_no_crop:
+            self.rescaler = albumentations.Resize(height=self.size, width=self.size)
+            self.preprocessor = albumentations.Compose(
+                [self.rescaler],
+                additional_targets={"segmentation": "image"})
+
+    @abstractmethod
+    def year(self):
+        raise NotImplementedError()
+
+    def __len__(self):
+        return len(self.labels["image_ids"])
+
+    def preprocess_image(self, image_path, segmentation_path=None):
+        image = Image.open(image_path)
+        if not image.mode == "RGB":
+            image = image.convert("RGB")
+        image = np.array(image).astype(np.uint8)
+        if segmentation_path:
+            segmentation = Image.open(segmentation_path)
+            if not self.onehot and not segmentation.mode == "RGB":
+                segmentation = segmentation.convert("RGB")
+            segmentation = np.array(segmentation).astype(np.uint8)
+            if self.onehot:
+                assert self.stuffthing
+                # stored in caffe format: unlabeled==255. stuff and thing from
+                # 0-181. to be compatible with the labels in
+                # https://github.com/nightrome/cocostuff/blob/master/labels.txt
+                # we shift stuffthing one to the right and put unlabeled in zero
+                # as long as segmentation is uint8 shifting to right handles the
+                # latter too
+                assert segmentation.dtype == np.uint8
+                segmentation = segmentation + 1
+
+            processed = self.preprocessor(image=image, segmentation=segmentation)
+
+            image, segmentation = processed["image"], processed["segmentation"]
+        else:
+            image = self.preprocessor(image=image,)['image']
+
+        image = (image / 127.5 - 1.0).astype(np.float32)
+        if segmentation_path:
+            if self.onehot:
+                assert segmentation.dtype == np.uint8
+                # make it one hot
+                n_labels = 183
+                flatseg = np.ravel(segmentation)
+                onehot = np.zeros((flatseg.size, n_labels), dtype=np.bool)
+                onehot[np.arange(flatseg.size), flatseg] = True
+                onehot = onehot.reshape(segmentation.shape + (n_labels,)).astype(int)
+                segmentation = onehot
+            else:
+                segmentation = (segmentation / 127.5 - 1.0).astype(np.float32)
+            return image, segmentation
+        else:
+            return image
+
+    def __getitem__(self, i):
+        img_path = self.img_id_to_filepath[self.labels["image_ids"][i]]
+        if self.use_segmenation:
+            seg_path = self.img_id_to_segmentation_filepath[self.labels["image_ids"][i]]
+            image, segmentation = self.preprocess_image(img_path, seg_path)
+        else:
+            image = self.preprocess_image(img_path)
+        captions = self.img_id_to_captions[self.labels["image_ids"][i]]
+        # randomly draw one of all available captions per image
+        caption = captions[np.random.randint(0, len(captions))]
+        example = {"image": image,
+                   #"caption": [str(caption[0])],
+                   "caption": caption,
+                   "img_path": img_path,
+                   "filename_": img_path.split(os.sep)[-1]
+                    }
+        if self.use_segmenation:
+            example.update({"seg_path": seg_path, 'segmentation': segmentation})
+        return example
+
+
+class CocoImagesAndCaptionsTrain2017(CocoBase):
+    """returns a pair of (image, caption)"""
+    def __init__(self, size, onehot_segmentation=False, use_stuffthing=False, crop_size=None, force_no_crop=False,):
+        super().__init__(size=size,
+                         dataroot="data/coco/train2017",
+                         datajson="data/coco/annotations/captions_train2017.json",
+                         onehot_segmentation=onehot_segmentation,
+                         use_stuffthing=use_stuffthing, crop_size=crop_size, force_no_crop=force_no_crop)
+
+    def get_split(self):
+        return "train"
+
+    def year(self):
+        return '2017'
+
+
+class CocoImagesAndCaptionsValidation2017(CocoBase):
+    """returns a pair of (image, caption)"""
+    def __init__(self, size, onehot_segmentation=False, use_stuffthing=False, crop_size=None, force_no_crop=False,
+                 given_files=None):
+        super().__init__(size=size,
+                         dataroot="data/coco/val2017",
+                         datajson="data/coco/annotations/captions_val2017.json",
+                         onehot_segmentation=onehot_segmentation,
+                         use_stuffthing=use_stuffthing, crop_size=crop_size, force_no_crop=force_no_crop,
+                         given_files=given_files)
+
+    def get_split(self):
+        return "validation"
+
+    def year(self):
+        return '2017'
+
+
+
+class CocoImagesAndCaptionsTrain2014(CocoBase):
+    """returns a pair of (image, caption)"""
+    def __init__(self, size, onehot_segmentation=False, use_stuffthing=False, crop_size=None, force_no_crop=False,crop_type='random'):
+        super().__init__(size=size,
+                         dataroot="data/coco/train2014",
+                         datajson="data/coco/annotations2014/annotations/captions_train2014.json",
+                         onehot_segmentation=onehot_segmentation,
+                         use_stuffthing=use_stuffthing, crop_size=crop_size, force_no_crop=force_no_crop,
+                         use_segmentation=False,
+                         crop_type=crop_type)
+
+    def get_split(self):
+        return "train"
+
+    def year(self):
+        return '2014'
+
+class CocoImagesAndCaptionsValidation2014(CocoBase):
+    """returns a pair of (image, caption)"""
+    def __init__(self, size, onehot_segmentation=False, use_stuffthing=False, crop_size=None, force_no_crop=False,
+                 given_files=None,crop_type='center',**kwargs):
+        super().__init__(size=size,
+                         dataroot="data/coco/val2014",
+                         datajson="data/coco/annotations2014/annotations/captions_val2014.json",
+                         onehot_segmentation=onehot_segmentation,
+                         use_stuffthing=use_stuffthing, crop_size=crop_size, force_no_crop=force_no_crop,
+                         given_files=given_files,
+                         use_segmentation=False,
+                         crop_type=crop_type)
+
+    def get_split(self):
+        return "validation"
+
+    def year(self):
+        return '2014'
+
+if __name__ == '__main__':
+    with open("data/coco/annotations2014/annotations/captions_val2014.json", "r") as json_file:
+        json_data = json.load(json_file)
+        capdirs = json_data["annotations"]
+        import pudb; pudb.set_trace()
+    #d2 = CocoImagesAndCaptionsTrain2014(size=256)
+    d2 = CocoImagesAndCaptionsValidation2014(size=256)
+    print("constructed dataset.")
+    print(f"length of {d2.__class__.__name__}: {len(d2)}")
+
+    ex2 = d2[0]
+    # ex3 = d3[0]
+    # print(ex1["image"].shape)
+    print(ex2["image"].shape)
+    # print(ex3["image"].shape)
+    # print(ex1["segmentation"].shape)
+    print(ex2["caption"].__class__.__name__)

ldm/data/dummy.py

@@ -0,0 +1,34 @@
+import numpy as np
+import random
+import string
+from torch.utils.data import Dataset, Subset
+
+class DummyData(Dataset):
+    def __init__(self, length, size):
+        self.length = length
+        self.size = size
+
+    def __len__(self):
+        return self.length
+
+    def __getitem__(self, i):
+        x = np.random.randn(*self.size)
+        letters = string.ascii_lowercase
+        y = ''.join(random.choice(string.ascii_lowercase) for i in range(10))
+        return {"jpg": x, "txt": y}
+
+
+class DummyDataWithEmbeddings(Dataset):
+    def __init__(self, length, size, emb_size):
+        self.length = length
+        self.size = size
+        self.emb_size = emb_size
+
+    def __len__(self):
+        return self.length
+
+    def __getitem__(self, i):
+        x = np.random.randn(*self.size)
+        y = np.random.randn(*self.emb_size).astype(np.float32)
+        return {"jpg": x, "txt": y}
+

ldm/data/imagenet.py

@@ -0,0 +1,394 @@
+import os, yaml, pickle, shutil, tarfile, glob
+import cv2
+import albumentations
+import PIL
+import numpy as np
+import torchvision.transforms.functional as TF
+from omegaconf import OmegaConf
+from functools import partial
+from PIL import Image
+from tqdm import tqdm
+from torch.utils.data import Dataset, Subset
+
+import taming.data.utils as tdu
+from taming.data.imagenet import str_to_indices, give_synsets_from_indices, download, retrieve
+from taming.data.imagenet import ImagePaths
+
+from ldm.modules.image_degradation import degradation_fn_bsr, degradation_fn_bsr_light
+
+
+def synset2idx(path_to_yaml="data/index_synset.yaml"):
+    with open(path_to_yaml) as f:
+        di2s = yaml.load(f)
+    return dict((v,k) for k,v in di2s.items())
+
+
+class ImageNetBase(Dataset):
+    def __init__(self, config=None):
+        self.config = config or OmegaConf.create()
+        if not type(self.config)==dict:
+            self.config = OmegaConf.to_container(self.config)
+        self.keep_orig_class_label = self.config.get("keep_orig_class_label", False)
+        self.process_images = True  # if False we skip loading & processing images and self.data contains filepaths
+        self._prepare()
+        self._prepare_synset_to_human()
+        self._prepare_idx_to_synset()
+        self._prepare_human_to_integer_label()
+        self._load()
+
+    def __len__(self):
+        return len(self.data)
+
+    def __getitem__(self, i):
+        return self.data[i]
+
+    def _prepare(self):
+        raise NotImplementedError()
+
+    def _filter_relpaths(self, relpaths):
+        ignore = set([
+            "n06596364_9591.JPEG",
+        ])
+        relpaths = [rpath for rpath in relpaths if not rpath.split("/")[-1] in ignore]
+        if "sub_indices" in self.config:
+            indices = str_to_indices(self.config["sub_indices"])
+            synsets = give_synsets_from_indices(indices, path_to_yaml=self.idx2syn)  # returns a list of strings
+            self.synset2idx = synset2idx(path_to_yaml=self.idx2syn)
+            files = []
+            for rpath in relpaths:
+                syn = rpath.split("/")[0]
+                if syn in synsets:
+                    files.append(rpath)
+            return files
+        else:
+            return relpaths
+
+    def _prepare_synset_to_human(self):
+        SIZE = 2655750
+        URL = "https://heibox.uni-heidelberg.de/f/9f28e956cd304264bb82/?dl=1"
+        self.human_dict = os.path.join(self.root, "synset_human.txt")
+        if (not os.path.exists(self.human_dict) or
+                not os.path.getsize(self.human_dict)==SIZE):
+            download(URL, self.human_dict)
+
+    def _prepare_idx_to_synset(self):
+        URL = "https://heibox.uni-heidelberg.de/f/d835d5b6ceda4d3aa910/?dl=1"
+        self.idx2syn = os.path.join(self.root, "index_synset.yaml")
+        if (not os.path.exists(self.idx2syn)):
+            download(URL, self.idx2syn)
+
+    def _prepare_human_to_integer_label(self):
+        URL = "https://heibox.uni-heidelberg.de/f/2362b797d5be43b883f6/?dl=1"
+        self.human2integer = os.path.join(self.root, "imagenet1000_clsidx_to_labels.txt")
+        if (not os.path.exists(self.human2integer)):
+            download(URL, self.human2integer)
+        with open(self.human2integer, "r") as f:
+            lines = f.read().splitlines()
+            assert len(lines) == 1000
+            self.human2integer_dict = dict()
+            for line in lines:
+                value, key = line.split(":")
+                self.human2integer_dict[key] = int(value)
+
+    def _load(self):
+        with open(self.txt_filelist, "r") as f:
+            self.relpaths = f.read().splitlines()
+            l1 = len(self.relpaths)
+            self.relpaths = self._filter_relpaths(self.relpaths)
+            print("Removed {} files from filelist during filtering.".format(l1 - len(self.relpaths)))
+
+        self.synsets = [p.split("/")[0] for p in self.relpaths]
+        self.abspaths = [os.path.join(self.datadir, p) for p in self.relpaths]
+
+        unique_synsets = np.unique(self.synsets)
+        class_dict = dict((synset, i) for i, synset in enumerate(unique_synsets))
+        if not self.keep_orig_class_label:
+            self.class_labels = [class_dict[s] for s in self.synsets]
+        else:
+            self.class_labels = [self.synset2idx[s] for s in self.synsets]
+
+        with open(self.human_dict, "r") as f:
+            human_dict = f.read().splitlines()
+            human_dict = dict(line.split(maxsplit=1) for line in human_dict)
+
+        self.human_labels = [human_dict[s] for s in self.synsets]
+
+        labels = {
+            "relpath": np.array(self.relpaths),
+            "synsets": np.array(self.synsets),
+            "class_label": np.array(self.class_labels),
+            "human_label": np.array(self.human_labels),
+        }
+
+        if self.process_images:
+            self.size = retrieve(self.config, "size", default=256)
+            self.data = ImagePaths(self.abspaths,
+                                   labels=labels,
+                                   size=self.size,
+                                   random_crop=self.random_crop,
+                                   )
+        else:
+            self.data = self.abspaths
+
+
+class ImageNetTrain(ImageNetBase):
+    NAME = "ILSVRC2012_train"
+    URL = "http://www.image-net.org/challenges/LSVRC/2012/"
+    AT_HASH = "a306397ccf9c2ead27155983c254227c0fd938e2"
+    FILES = [
+        "ILSVRC2012_img_train.tar",
+    ]
+    SIZES = [
+        147897477120,
+    ]
+
+    def __init__(self, process_images=True, data_root=None, **kwargs):
+        self.process_images = process_images
+        self.data_root = data_root
+        super().__init__(**kwargs)
+
+    def _prepare(self):
+        if self.data_root:
+            self.root = os.path.join(self.data_root, self.NAME)
+        else:
+            cachedir = os.environ.get("XDG_CACHE_HOME", os.path.expanduser("~/.cache"))
+            self.root = os.path.join(cachedir, "autoencoders/data", self.NAME)
+
+        self.datadir = os.path.join(self.root, "data")
+        self.txt_filelist = os.path.join(self.root, "filelist.txt")
+        self.expected_length = 1281167
+        self.random_crop = retrieve(self.config, "ImageNetTrain/random_crop",
+                                    default=True)
+        if not tdu.is_prepared(self.root):
+            # prep
+            print("Preparing dataset {} in {}".format(self.NAME, self.root))
+
+            datadir = self.datadir
+            if not os.path.exists(datadir):
+                path = os.path.join(self.root, self.FILES[0])
+                if not os.path.exists(path) or not os.path.getsize(path)==self.SIZES[0]:
+                    import academictorrents as at
+                    atpath = at.get(self.AT_HASH, datastore=self.root)
+                    assert atpath == path
+
+                print("Extracting {} to {}".format(path, datadir))
+                os.makedirs(datadir, exist_ok=True)
+                with tarfile.open(path, "r:") as tar:
+                    tar.extractall(path=datadir)
+
+                print("Extracting sub-tars.")
+                subpaths = sorted(glob.glob(os.path.join(datadir, "*.tar")))
+                for subpath in tqdm(subpaths):
+                    subdir = subpath[:-len(".tar")]
+                    os.makedirs(subdir, exist_ok=True)
+                    with tarfile.open(subpath, "r:") as tar:
+                        tar.extractall(path=subdir)
+
+            filelist = glob.glob(os.path.join(datadir, "**", "*.JPEG"))
+            filelist = [os.path.relpath(p, start=datadir) for p in filelist]
+            filelist = sorted(filelist)
+            filelist = "\n".join(filelist)+"\n"
+            with open(self.txt_filelist, "w") as f:
+                f.write(filelist)
+
+            tdu.mark_prepared(self.root)
+
+
+class ImageNetValidation(ImageNetBase):
+    NAME = "ILSVRC2012_validation"
+    URL = "http://www.image-net.org/challenges/LSVRC/2012/"
+    AT_HASH = "5d6d0df7ed81efd49ca99ea4737e0ae5e3a5f2e5"
+    VS_URL = "https://heibox.uni-heidelberg.de/f/3e0f6e9c624e45f2bd73/?dl=1"
+    FILES = [
+        "ILSVRC2012_img_val.tar",
+        "validation_synset.txt",
+    ]
+    SIZES = [
+        6744924160,
+        1950000,
+    ]
+
+    def __init__(self, process_images=True, data_root=None, **kwargs):
+        self.data_root = data_root
+        self.process_images = process_images
+        super().__init__(**kwargs)
+
+    def _prepare(self):
+        if self.data_root:
+            self.root = os.path.join(self.data_root, self.NAME)
+        else:
+            cachedir = os.environ.get("XDG_CACHE_HOME", os.path.expanduser("~/.cache"))
+            self.root = os.path.join(cachedir, "autoencoders/data", self.NAME)
+        self.datadir = os.path.join(self.root, "data")
+        self.txt_filelist = os.path.join(self.root, "filelist.txt")
+        self.expected_length = 50000
+        self.random_crop = retrieve(self.config, "ImageNetValidation/random_crop",
+                                    default=False)
+        if not tdu.is_prepared(self.root):
+            # prep
+            print("Preparing dataset {} in {}".format(self.NAME, self.root))
+
+            datadir = self.datadir
+            if not os.path.exists(datadir):
+                path = os.path.join(self.root, self.FILES[0])
+                if not os.path.exists(path) or not os.path.getsize(path)==self.SIZES[0]:
+                    import academictorrents as at
+                    atpath = at.get(self.AT_HASH, datastore=self.root)
+                    assert atpath == path
+
+                print("Extracting {} to {}".format(path, datadir))
+                os.makedirs(datadir, exist_ok=True)
+                with tarfile.open(path, "r:") as tar:
+                    tar.extractall(path=datadir)
+
+                vspath = os.path.join(self.root, self.FILES[1])
+                if not os.path.exists(vspath) or not os.path.getsize(vspath)==self.SIZES[1]:
+                    download(self.VS_URL, vspath)
+
+                with open(vspath, "r") as f:
+                    synset_dict = f.read().splitlines()
+                    synset_dict = dict(line.split() for line in synset_dict)
+
+                print("Reorganizing into synset folders")
+                synsets = np.unique(list(synset_dict.values()))
+                for s in synsets:
+                    os.makedirs(os.path.join(datadir, s), exist_ok=True)
+                for k, v in synset_dict.items():
+                    src = os.path.join(datadir, k)
+                    dst = os.path.join(datadir, v)
+                    shutil.move(src, dst)
+
+            filelist = glob.glob(os.path.join(datadir, "**", "*.JPEG"))
+            filelist = [os.path.relpath(p, start=datadir) for p in filelist]
+            filelist = sorted(filelist)
+            filelist = "\n".join(filelist)+"\n"
+            with open(self.txt_filelist, "w") as f:
+                f.write(filelist)
+
+            tdu.mark_prepared(self.root)
+
+
+
+class ImageNetSR(Dataset):
+    def __init__(self, size=None,
+                 degradation=None, downscale_f=4, min_crop_f=0.5, max_crop_f=1.,
+                 random_crop=True):
+        """
+        Imagenet Superresolution Dataloader
+        Performs following ops in order:
+        1.  crops a crop of size s from image either as random or center crop
+        2.  resizes crop to size with cv2.area_interpolation
+        3.  degrades resized crop with degradation_fn
+
+        :param size: resizing to size after cropping
+        :param degradation: degradation_fn, e.g. cv_bicubic or bsrgan_light
+        :param downscale_f: Low Resolution Downsample factor
+        :param min_crop_f: determines crop size s,
+          where s = c * min_img_side_len with c sampled from interval (min_crop_f, max_crop_f)
+        :param max_crop_f: ""
+        :param data_root:
+        :param random_crop:
+        """
+        self.base = self.get_base()
+        assert size
+        assert (size / downscale_f).is_integer()
+        self.size = size
+        self.LR_size = int(size / downscale_f)
+        self.min_crop_f = min_crop_f
+        self.max_crop_f = max_crop_f
+        assert(max_crop_f <= 1.)
+        self.center_crop = not random_crop
+
+        self.image_rescaler = albumentations.SmallestMaxSize(max_size=size, interpolation=cv2.INTER_AREA)
+
+        self.pil_interpolation = False # gets reset later if incase interp_op is from pillow
+
+        if degradation == "bsrgan":
+            self.degradation_process = partial(degradation_fn_bsr, sf=downscale_f)
+
+        elif degradation == "bsrgan_light":
+            self.degradation_process = partial(degradation_fn_bsr_light, sf=downscale_f)
+
+        else:
+            interpolation_fn = {
+            "cv_nearest": cv2.INTER_NEAREST,
+            "cv_bilinear": cv2.INTER_LINEAR,
+            "cv_bicubic": cv2.INTER_CUBIC,
+            "cv_area": cv2.INTER_AREA,
+            "cv_lanczos": cv2.INTER_LANCZOS4,
+            "pil_nearest": PIL.Image.NEAREST,
+            "pil_bilinear": PIL.Image.BILINEAR,
+            "pil_bicubic": PIL.Image.BICUBIC,
+            "pil_box": PIL.Image.BOX,
+            "pil_hamming": PIL.Image.HAMMING,
+            "pil_lanczos": PIL.Image.LANCZOS,
+            }[degradation]
+
+            self.pil_interpolation = degradation.startswith("pil_")
+
+            if self.pil_interpolation:
+                self.degradation_process = partial(TF.resize, size=self.LR_size, interpolation=interpolation_fn)
+
+            else:
+                self.degradation_process = albumentations.SmallestMaxSize(max_size=self.LR_size,
+                                                                          interpolation=interpolation_fn)
+
+    def __len__(self):
+        return len(self.base)
+
+    def __getitem__(self, i):
+        example = self.base[i]
+        image = Image.open(example["file_path_"])
+
+        if not image.mode == "RGB":
+            image = image.convert("RGB")
+
+        image = np.array(image).astype(np.uint8)
+
+        min_side_len = min(image.shape[:2])
+        crop_side_len = min_side_len * np.random.uniform(self.min_crop_f, self.max_crop_f, size=None)
+        crop_side_len = int(crop_side_len)
+
+        if self.center_crop:
+            self.cropper = albumentations.CenterCrop(height=crop_side_len, width=crop_side_len)
+
+        else:
+            self.cropper = albumentations.RandomCrop(height=crop_side_len, width=crop_side_len)
+
+        image = self.cropper(image=image)["image"]
+        image = self.image_rescaler(image=image)["image"]
+
+        if self.pil_interpolation:
+            image_pil = PIL.Image.fromarray(image)
+            LR_image = self.degradation_process(image_pil)
+            LR_image = np.array(LR_image).astype(np.uint8)
+
+        else:
+            LR_image = self.degradation_process(image=image)["image"]
+
+        example["image"] = (image/127.5 - 1.0).astype(np.float32)
+        example["LR_image"] = (LR_image/127.5 - 1.0).astype(np.float32)
+        example["caption"] = example["human_label"]  # dummy caption
+        return example
+
+
+class ImageNetSRTrain(ImageNetSR):
+    def __init__(self, **kwargs):
+        super().__init__(**kwargs)
+
+    def get_base(self):
+        with open("data/imagenet_train_hr_indices.p", "rb") as f:
+            indices = pickle.load(f)
+        dset = ImageNetTrain(process_images=False,)
+        return Subset(dset, indices)
+
+
+class ImageNetSRValidation(ImageNetSR):
+    def __init__(self, **kwargs):
+        super().__init__(**kwargs)
+
+    def get_base(self):
+        with open("data/imagenet_val_hr_indices.p", "rb") as f:
+            indices = pickle.load(f)
+        dset = ImageNetValidation(process_images=False,)
+        return Subset(dset, indices)

ldm/data/inpainting/synthetic_mask.py

@@ -0,0 +1,166 @@
+from PIL import Image, ImageDraw
+import numpy as np
+
+settings = {
+    "256narrow": {
+        "p_irr": 1,
+        "min_n_irr": 4,
+        "max_n_irr": 50,
+        "max_l_irr": 40,
+        "max_w_irr": 10,
+        "min_n_box": None,
+        "max_n_box": None,
+        "min_s_box": None,
+        "max_s_box": None,
+        "marg": None,
+    },
+    "256train": {
+        "p_irr": 0.5,
+        "min_n_irr": 1,
+        "max_n_irr": 5,
+        "max_l_irr": 200,
+        "max_w_irr": 100,
+        "min_n_box": 1,
+        "max_n_box": 4,
+        "min_s_box": 30,
+        "max_s_box": 150,
+        "marg": 10,
+    },
+    "512train": {    # TODO: experimental
+            "p_irr": 0.5,
+            "min_n_irr": 1,
+            "max_n_irr": 5,
+            "max_l_irr": 450,
+            "max_w_irr": 250,
+            "min_n_box": 1,
+            "max_n_box": 4,
+            "min_s_box": 30,
+            "max_s_box": 300,
+            "marg": 10,
+        },
+    "512train-large": {    # TODO: experimental
+            "p_irr": 0.5,
+            "min_n_irr": 1,
+            "max_n_irr": 5,
+            "max_l_irr": 450,
+            "max_w_irr": 400,
+            "min_n_box": 1,
+            "max_n_box": 4,
+            "min_s_box": 75,
+            "max_s_box": 450,
+            "marg": 10,
+        },
+}
+
+
+def gen_segment_mask(mask, start, end, brush_width):
+    mask = mask > 0
+    mask = (255 * mask).astype(np.uint8)
+    mask = Image.fromarray(mask)
+    draw = ImageDraw.Draw(mask)
+    draw.line([start, end], fill=255, width=brush_width, joint="curve")
+    mask = np.array(mask) / 255
+    return mask
+
+
+def gen_box_mask(mask, masked):
+    x_0, y_0, w, h = masked
+    mask[y_0:y_0 + h, x_0:x_0 + w] = 1
+    return mask
+
+
+def gen_round_mask(mask, masked, radius):
+    x_0, y_0, w, h = masked
+    xy = [(x_0, y_0), (x_0 + w, y_0 + w)]
+
+    mask = mask > 0
+    mask = (255 * mask).astype(np.uint8)
+    mask = Image.fromarray(mask)
+    draw = ImageDraw.Draw(mask)
+    draw.rounded_rectangle(xy, radius=radius, fill=255)
+    mask = np.array(mask) / 255
+    return mask
+
+
+def gen_large_mask(prng, img_h, img_w,
+                   marg, p_irr, min_n_irr, max_n_irr, max_l_irr, max_w_irr,
+                   min_n_box, max_n_box, min_s_box, max_s_box):
+    """
+    img_h: int, an image height
+    img_w: int, an image width
+    marg: int, a margin for a box starting coordinate
+    p_irr: float, 0 <= p_irr <= 1, a probability of a polygonal chain mask
+
+    min_n_irr: int, min number of segments
+    max_n_irr: int, max number of segments
+    max_l_irr: max length of a segment in polygonal chain
+    max_w_irr: max width of a segment in polygonal chain
+
+    min_n_box: int, min bound for the number of box primitives
+    max_n_box: int, max bound for the number of box primitives
+    min_s_box: int, min length of a box side
+    max_s_box: int, max length of a box side
+    """
+
+    mask = np.zeros((img_h, img_w))
+    uniform = prng.randint
+
+    if np.random.uniform(0, 1) < p_irr:  # generate polygonal chain
+        n = uniform(min_n_irr, max_n_irr)  # sample number of segments
+
+        for _ in range(n):
+            y = uniform(0, img_h)  # sample a starting point
+            x = uniform(0, img_w)
+
+            a = uniform(0, 360)  # sample angle
+            l = uniform(10, max_l_irr)  # sample segment length
+            w = uniform(5, max_w_irr)  # sample a segment width
+
+            # draw segment starting from (x,y) to (x_,y_) using brush of width w
+            x_ = x + l * np.sin(a)
+            y_ = y + l * np.cos(a)
+
+            mask = gen_segment_mask(mask, start=(x, y), end=(x_, y_), brush_width=w)
+            x, y = x_, y_
+    else:  # generate Box masks
+        n = uniform(min_n_box, max_n_box)  # sample number of rectangles
+
+        for _ in range(n):
+            h = uniform(min_s_box, max_s_box)  # sample box shape
+            w = uniform(min_s_box, max_s_box)
+
+            x_0 = uniform(marg, img_w - marg - w)  # sample upper-left coordinates of box
+            y_0 = uniform(marg, img_h - marg - h)
+
+            if np.random.uniform(0, 1) < 0.5:
+                mask = gen_box_mask(mask, masked=(x_0, y_0, w, h))
+            else:
+                r = uniform(0, 60)  # sample radius
+                mask = gen_round_mask(mask, masked=(x_0, y_0, w, h), radius=r)
+    return mask
+
+
+make_lama_mask = lambda prng, h, w: gen_large_mask(prng, h, w, **settings["256train"])
+make_narrow_lama_mask = lambda prng, h, w: gen_large_mask(prng, h, w, **settings["256narrow"])
+make_512_lama_mask = lambda prng, h, w: gen_large_mask(prng, h, w, **settings["512train"])
+make_512_lama_mask_large = lambda prng, h, w: gen_large_mask(prng, h, w, **settings["512train-large"])
+
+
+MASK_MODES = {
+    "256train": make_lama_mask,
+    "256narrow": make_narrow_lama_mask,
+    "512train": make_512_lama_mask,
+    "512train-large": make_512_lama_mask_large
+}
+
+if __name__ == "__main__":
+    import sys
+
+    out = sys.argv[1]
+
+    prng = np.random.RandomState(1)
+    kwargs = settings["256train"]
+    mask = gen_large_mask(prng, 256, 256, **kwargs)
+    mask = (255 * mask).astype(np.uint8)
+    mask = Image.fromarray(mask)
+    mask.save(out)

ldm/data/laion.py

@@ -0,0 +1,537 @@
+import webdataset as wds
+import kornia
+from PIL import Image
+import io
+import os
+import torchvision
+from PIL import Image
+import glob
+import random
+import numpy as np
+import pytorch_lightning as pl
+from tqdm import tqdm
+from omegaconf import OmegaConf
+from einops import rearrange
+import torch
+from webdataset.handlers import warn_and_continue
+
+
+from ldm.util import instantiate_from_config
+from ldm.data.inpainting.synthetic_mask import gen_large_mask, MASK_MODES
+from ldm.data.base import PRNGMixin
+
+
+class DataWithWings(torch.utils.data.IterableDataset):
+    def __init__(self, min_size, transform=None, target_transform=None):
+        self.min_size = min_size
+        self.transform = transform if transform is not None else nn.Identity()
+        self.target_transform = target_transform if target_transform is not None else nn.Identity()
+        self.kv = OnDiskKV(file='/home/ubuntu/laion5B-watermark-safety-ordered', key_format='q', value_format='ee')
+        self.kv_aesthetic = OnDiskKV(file='/home/ubuntu/laion5B-aesthetic-tags-kv', key_format='q', value_format='e')
+        self.pwatermark_threshold = 0.8
+        self.punsafe_threshold = 0.5
+        self.aesthetic_threshold = 5.
+        self.total_samples = 0
+        self.samples = 0
+        location = 'pipe:aws s3 cp --quiet s3://s-datasets/laion5b/laion2B-data/{000000..231349}.tar -'
+
+        self.inner_dataset = wds.DataPipeline(
+            wds.ResampledShards(location),
+            wds.tarfile_to_samples(handler=wds.warn_and_continue),
+            wds.shuffle(1000, handler=wds.warn_and_continue),
+            wds.decode('pilrgb', handler=wds.warn_and_continue),
+            wds.map(self._add_tags, handler=wds.ignore_and_continue),
+            wds.select(self._filter_predicate),
+            wds.map_dict(jpg=self.transform, txt=self.target_transform, punsafe=self._punsafe_to_class, handler=wds.warn_and_continue),
+            wds.to_tuple('jpg', 'txt', 'punsafe', handler=wds.warn_and_continue),
+        )
+
+    @staticmethod
+    def _compute_hash(url, text):
+        if url is None:
+            url = ''
+        if text is None:
+            text = ''
+        total = (url + text).encode('utf-8')
+        return mmh3.hash64(total)[0]
+
+    def _add_tags(self, x):
+        hsh = self._compute_hash(x['json']['url'], x['txt'])
+        pwatermark, punsafe = self.kv[hsh]
+        aesthetic = self.kv_aesthetic[hsh][0]
+        return {**x, 'pwatermark': pwatermark, 'punsafe': punsafe, 'aesthetic': aesthetic}
+
+    def _punsafe_to_class(self, punsafe):
+        return torch.tensor(punsafe >= self.punsafe_threshold).long()
+
+    def _filter_predicate(self, x):
+        try:
+            return x['pwatermark'] < self.pwatermark_threshold and x['aesthetic'] >= self.aesthetic_threshold and x['json']['original_width'] >= self.min_size and x['json']['original_height'] >= self.min_size
+        except:
+            return False
+
+    def __iter__(self):
+        return iter(self.inner_dataset)
+
+
+def dict_collation_fn(samples, combine_tensors=True, combine_scalars=True):
+    """Take a list  of samples (as dictionary) and create a batch, preserving the keys.
+    If `tensors` is True, `ndarray` objects are combined into
+    tensor batches.
+    :param dict samples: list of samples
+    :param bool tensors: whether to turn lists of ndarrays into a single ndarray
+    :returns: single sample consisting of a batch
+    :rtype: dict
+    """
+    keys = set.intersection(*[set(sample.keys()) for sample in samples])
+    batched = {key: [] for key in keys}
+
+    for s in samples:
+        [batched[key].append(s[key]) for key in batched]
+
+    result = {}
+    for key in batched:
+        if isinstance(batched[key][0], (int, float)):
+            if combine_scalars:
+                result[key] = np.array(list(batched[key]))
+        elif isinstance(batched[key][0], torch.Tensor):
+            if combine_tensors:
+                result[key] = torch.stack(list(batched[key]))
+        elif isinstance(batched[key][0], np.ndarray):
+            if combine_tensors:
+                result[key] = np.array(list(batched[key]))
+        else:
+            result[key] = list(batched[key])
+    return result
+
+
+class WebDataModuleFromConfig(pl.LightningDataModule):
+    def __init__(self, tar_base, batch_size, train=None, validation=None,
+                 test=None, num_workers=4, multinode=True, min_size=None,
+                 max_pwatermark=1.0,
+                 **kwargs):
+        super().__init__(self)
+        print(f'Setting tar base to {tar_base}')
+        self.tar_base = tar_base
+        self.batch_size = batch_size
+        self.num_workers = num_workers
+        self.train = train
+        self.validation = validation
+        self.test = test
+        self.multinode = multinode
+        self.min_size = min_size  # filter out very small images
+        self.max_pwatermark = max_pwatermark # filter out watermarked images
+
+    def make_loader(self, dataset_config, train=True):
+        if 'image_transforms' in dataset_config:
+            image_transforms = [instantiate_from_config(tt) for tt in dataset_config.image_transforms]
+        else:
+            image_transforms = []
+
+        image_transforms.extend([torchvision.transforms.ToTensor(),
+                                 torchvision.transforms.Lambda(lambda x: rearrange(x * 2. - 1., 'c h w -> h w c'))])
+        image_transforms = torchvision.transforms.Compose(image_transforms)
+
+        if 'transforms' in dataset_config:
+            transforms_config = OmegaConf.to_container(dataset_config.transforms)
+        else:
+            transforms_config = dict()
+
+        transform_dict = {dkey: load_partial_from_config(transforms_config[dkey])
+                if transforms_config[dkey] != 'identity' else identity
+                for dkey in transforms_config}
+        img_key = dataset_config.get('image_key', 'jpeg')
+        transform_dict.update({img_key: image_transforms})
+
+        if 'postprocess' in dataset_config:
+            postprocess = instantiate_from_config(dataset_config['postprocess'])
+        else:
+            postprocess = None
+
+        shuffle = dataset_config.get('shuffle', 0)
+        shardshuffle = shuffle > 0
+
+        nodesplitter = wds.shardlists.split_by_node if self.multinode else wds.shardlists.single_node_only
+
+        if self.tar_base == "__improvedaesthetic__":
+            print("## Warning, loading the same improved aesthetic dataset "
+                    "for all splits and ignoring shards parameter.")
+            tars = "pipe:aws s3 cp s3://s-laion/improved-aesthetics-laion-2B-en-subsets/aesthetics_tars/{000000..060207}.tar -"
+        else:
+            tars = os.path.join(self.tar_base, dataset_config.shards)
+
+        dset = wds.WebDataset(
+                tars,
+                nodesplitter=nodesplitter,
+                shardshuffle=shardshuffle,
+                handler=wds.warn_and_continue).repeat().shuffle(shuffle)
+        print(f'Loading webdataset with {len(dset.pipeline[0].urls)} shards.')
+
+        dset = (dset
+                .select(self.filter_keys)
+                .decode('pil', handler=wds.warn_and_continue)
+                .select(self.filter_size)
+                .map_dict(**transform_dict, handler=wds.warn_and_continue)
+                )
+        if postprocess is not None:
+            dset = dset.map(postprocess)
+        dset = (dset
+                .batched(self.batch_size, partial=False,
+                    collation_fn=dict_collation_fn)
+                )
+
+        loader = wds.WebLoader(dset, batch_size=None, shuffle=False,
+                               num_workers=self.num_workers)
+
+        return loader
+
+    def filter_size(self, x):
+        try:
+            valid = True
+            if self.min_size is not None and self.min_size > 1:
+                try:
+                    valid = valid and x['json']['original_width'] >= self.min_size and x['json']['original_height'] >= self.min_size
+                except Exception:
+                    valid = False
+            if self.max_pwatermark is not None and self.max_pwatermark < 1.0:
+                try:
+                    valid = valid and  x['json']['pwatermark'] <= self.max_pwatermark
+                except Exception:
+                    valid = False
+            return valid
+        except Exception:
+            return False
+
+    def filter_keys(self, x):
+        try:
+            return ("jpg" in x) and ("txt" in x)
+        except Exception:
+            return False
+
+    def train_dataloader(self):
+        return self.make_loader(self.train)
+
+    def val_dataloader(self):
+        return self.make_loader(self.validation, train=False)
+
+    def test_dataloader(self):
+        return self.make_loader(self.test, train=False)
+
+
+from ldm.modules.image_degradation import degradation_fn_bsr_light
+import cv2
+
+class AddLR(object):
+    def __init__(self, factor, output_size, initial_size=None, image_key="jpg"):
+        self.factor = factor
+        self.output_size = output_size
+        self.image_key = image_key
+        self.initial_size = initial_size
+
+    def pt2np(self, x):
+        x = ((x+1.0)*127.5).clamp(0, 255).to(dtype=torch.uint8).detach().cpu().numpy()
+        return x
+
+    def np2pt(self, x):
+        x = torch.from_numpy(x)/127.5-1.0
+        return x
+
+    def __call__(self, sample):
+        # sample['jpg'] is tensor hwc in [-1, 1] at this point
+        x = self.pt2np(sample[self.image_key])
+        if self.initial_size is not None:
+            x = cv2.resize(x, (self.initial_size, self.initial_size), interpolation=2)
+        x = degradation_fn_bsr_light(x, sf=self.factor)['image']
+        x = cv2.resize(x, (self.output_size, self.output_size), interpolation=2)
+        x = self.np2pt(x)
+        sample['lr'] = x
+        return sample
+
+class AddBW(object):
+    def __init__(self, image_key="jpg"):
+        self.image_key = image_key
+
+    def pt2np(self, x):
+        x = ((x+1.0)*127.5).clamp(0, 255).to(dtype=torch.uint8).detach().cpu().numpy()
+        return x
+
+    def np2pt(self, x):
+        x = torch.from_numpy(x)/127.5-1.0
+        return x
+
+    def __call__(self, sample):
+        # sample['jpg'] is tensor hwc in [-1, 1] at this point
+        x = sample[self.image_key]
+        w = torch.rand(3, device=x.device)
+        w /= w.sum()
+        out = torch.einsum('hwc,c->hw', x, w)
+
+        # Keep as 3ch so we can pass to encoder, also we might want to add hints
+        sample['lr'] = out.unsqueeze(-1).tile(1,1,3)
+        return sample
+
+class AddMask(PRNGMixin):
+    def __init__(self, mode="512train", p_drop=0.):
+        super().__init__()
+        assert mode in list(MASK_MODES.keys()), f'unknown mask generation mode "{mode}"'
+        self.make_mask = MASK_MODES[mode]
+        self.p_drop = p_drop
+
+    def __call__(self, sample):
+        # sample['jpg'] is tensor hwc in [-1, 1] at this point
+        x = sample['jpg']
+        mask = self.make_mask(self.prng, x.shape[0], x.shape[1])
+        if self.prng.choice(2, p=[1 - self.p_drop, self.p_drop]):
+            mask = np.ones_like(mask)
+        mask[mask < 0.5] = 0
+        mask[mask > 0.5] = 1
+        mask = torch.from_numpy(mask[..., None])
+        sample['mask'] = mask
+        sample['masked_image'] = x * (mask < 0.5)
+        return sample
+
+
+class AddEdge(PRNGMixin):
+    def __init__(self, mode="512train", mask_edges=True):
+        super().__init__()
+        assert mode in list(MASK_MODES.keys()), f'unknown mask generation mode "{mode}"'
+        self.make_mask = MASK_MODES[mode]
+        self.n_down_choices = [0]
+        self.sigma_choices = [1, 2]
+        self.mask_edges = mask_edges
+
+    @torch.no_grad()
+    def __call__(self, sample):
+        # sample['jpg'] is tensor hwc in [-1, 1] at this point
+        x = sample['jpg']
+
+        mask = self.make_mask(self.prng, x.shape[0], x.shape[1])
+        mask[mask < 0.5] = 0
+        mask[mask > 0.5] = 1
+        mask = torch.from_numpy(mask[..., None])
+        sample['mask'] = mask
+
+        n_down_idx = self.prng.choice(len(self.n_down_choices))
+        sigma_idx = self.prng.choice(len(self.sigma_choices))
+
+        n_choices = len(self.n_down_choices)*len(self.sigma_choices)
+        raveled_idx = np.ravel_multi_index((n_down_idx, sigma_idx),
+                                           (len(self.n_down_choices), len(self.sigma_choices)))
+        normalized_idx = raveled_idx/max(1, n_choices-1)
+
+        n_down = self.n_down_choices[n_down_idx]
+        sigma = self.sigma_choices[sigma_idx]
+
+        kernel_size = 4*sigma+1
+        kernel_size = (kernel_size, kernel_size)
+        sigma = (sigma, sigma)
+        canny = kornia.filters.Canny(
+                low_threshold=0.1,
+                high_threshold=0.2,
+                kernel_size=kernel_size,
+                sigma=sigma,
+                hysteresis=True,
+                )
+        y = (x+1.0)/2.0 # in 01
+        y = y.unsqueeze(0).permute(0, 3, 1, 2).contiguous()
+
+        # down
+        for i_down in range(n_down):
+            size = min(y.shape[-2], y.shape[-1])//2
+            y = kornia.geometry.transform.resize(y, size, antialias=True)
+
+        # edge
+        _, y = canny(y)
+
+        if n_down > 0:
+            size = x.shape[0], x.shape[1]
+            y = kornia.geometry.transform.resize(y, size, interpolation="nearest")
+
+        y = y.permute(0, 2, 3, 1)[0].expand(-1, -1, 3).contiguous()
+        y = y*2.0-1.0
+
+        if self.mask_edges:
+            sample['masked_image'] = y * (mask < 0.5)
+        else:
+            sample['masked_image'] = y
+            sample['mask'] = torch.zeros_like(sample['mask'])
+
+        # concat normalized idx
+        sample['smoothing_strength'] = torch.ones_like(sample['mask'])*normalized_idx
+
+        return sample
+
+
+def example00():
+    url = "pipe:aws s3 cp s3://s-datasets/laion5b/laion2B-data/000000.tar -"
+    dataset = wds.WebDataset(url)
+    example = next(iter(dataset))
+    for k in example:
+        print(k, type(example[k]))
+
+    print(example["__key__"])
+    for k in ["json", "txt"]:
+        print(example[k].decode())
+
+    image = Image.open(io.BytesIO(example["jpg"]))
+    outdir = "tmp"
+    os.makedirs(outdir, exist_ok=True)
+    image.save(os.path.join(outdir, example["__key__"] + ".png"))
+
+
+    def load_example(example):
+        return {
+            "key": example["__key__"],
+            "image": Image.open(io.BytesIO(example["jpg"])),
+            "text": example["txt"].decode(),
+        }
+
+
+    for i, example in tqdm(enumerate(dataset)):
+        ex = load_example(example)
+        print(ex["image"].size, ex["text"])
+        if i >= 100:
+            break
+
+
+def example01():
+    # the first laion shards contain ~10k examples each
+    url = "pipe:aws s3 cp s3://s-datasets/laion5b/laion2B-data/{000000..000002}.tar -"
+
+    batch_size = 3
+    shuffle_buffer = 10000
+    dset = wds.WebDataset(
+            url,
+            nodesplitter=wds.shardlists.split_by_node,
+            shardshuffle=True,
+            )
+    dset = (dset
+            .shuffle(shuffle_buffer, initial=shuffle_buffer)
+            .decode('pil', handler=warn_and_continue)
+            .batched(batch_size, partial=False,
+                collation_fn=dict_collation_fn)
+            )
+
+    num_workers = 2
+    loader = wds.WebLoader(dset, batch_size=None, shuffle=False, num_workers=num_workers)
+
+    batch_sizes = list()
+    keys_per_epoch = list()
+    for epoch in range(5):
+        keys = list()
+        for batch in tqdm(loader):
+            batch_sizes.append(len(batch["__key__"]))
+            keys.append(batch["__key__"])
+
+        for bs in batch_sizes:
+            assert bs==batch_size
+        print(f"{len(batch_sizes)} batches of size {batch_size}.")
+        batch_sizes = list()
+
+        keys_per_epoch.append(keys)
+        for i_batch in [0, 1, -1]:
+            print(f"Batch {i_batch} of epoch {epoch}:")
+            print(keys[i_batch])
+        print("next epoch.")
+
+
+def example02():
+    from omegaconf import OmegaConf
+    from torch.utils.data.distributed import DistributedSampler
+    from torch.utils.data import IterableDataset
+    from torch.utils.data import DataLoader, RandomSampler, Sampler, SequentialSampler
+    from pytorch_lightning.trainer.supporters import CombinedLoader, CycleIterator
+
+    #config = OmegaConf.load("configs/stable-diffusion/txt2img-1p4B-multinode-clip-encoder-high-res-512.yaml")
+    #config = OmegaConf.load("configs/stable-diffusion/txt2img-upscale-clip-encoder-f16-1024.yaml")
+    config = OmegaConf.load("configs/stable-diffusion/txt2img-v2-clip-encoder-improved_aesthetics-256.yaml")
+    datamod = WebDataModuleFromConfig(**config["data"]["params"])
+    dataloader = datamod.train_dataloader()
+
+    for batch in dataloader:
+        print(batch.keys())
+        print(batch["jpg"].shape)
+        break
+
+
+def example03():
+    # improved aesthetics
+    tars = "pipe:aws s3 cp s3://s-laion/improved-aesthetics-laion-2B-en-subsets/aesthetics_tars/{000000..060207}.tar -"
+    dataset = wds.WebDataset(tars)
+
+    def filter_keys(x):
+        try:
+            return ("jpg" in x) and ("txt" in x)
+        except Exception:
+            return False
+
+    def filter_size(x):
+        try:
+            return x['json']['original_width'] >= 512 and x['json']['original_height'] >= 512
+        except Exception:
+            return False
+
+    def filter_watermark(x):
+        try:
+            return x['json']['pwatermark'] < 0.5
+        except Exception:
+            return False
+
+    dataset = (dataset
+                .select(filter_keys)
+                .decode('pil', handler=wds.warn_and_continue))
+    n_save = 20
+    n_total = 0
+    n_large = 0
+    n_large_nowm = 0
+    for i, example in enumerate(dataset):
+        n_total += 1
+        if filter_size(example):
+            n_large += 1
+            if filter_watermark(example):
+                n_large_nowm += 1
+                if n_large_nowm < n_save+1:
+                    image = example["jpg"]
+                    image.save(os.path.join("tmp", f"{n_large_nowm-1:06}.png"))
+
+        if i%500 == 0:
+            print(i)
+            print(f"Large: {n_large}/{n_total} | {n_large/n_total*100:.2f}%")
+            if n_large > 0:
+                print(f"No Watermark: {n_large_nowm}/{n_large} | {n_large_nowm/n_large*100:.2f}%")
+
+
+
+def example04():
+    # improved aesthetics
+    for i_shard in range(60208)[::-1]:
+        print(i_shard)
+        tars = "pipe:aws s3 cp s3://s-laion/improved-aesthetics-laion-2B-en-subsets/aesthetics_tars/{:06}.tar -".format(i_shard)
+        dataset = wds.WebDataset(tars)
+
+        def filter_keys(x):
+            try:
+                return ("jpg" in x) and ("txt" in x)
+            except Exception:
+                return False
+
+        def filter_size(x):
+            try:
+                return x['json']['original_width'] >= 512 and x['json']['original_height'] >= 512
+            except Exception:
+                return False
+
+        dataset = (dataset
+                    .select(filter_keys)
+                    .decode('pil', handler=wds.warn_and_continue))
+        try:
+            example = next(iter(dataset))
+        except Exception:
+            print(f"Error @ {i_shard}")
+
+
+if __name__ == "__main__":
+    #example01()
+    #example02()
+    example03()
+    #example04()

ldm/data/lsun.py

@@ -0,0 +1,92 @@
+import os
+import numpy as np
+import PIL
+from PIL import Image
+from torch.utils.data import Dataset
+from torchvision import transforms
+
+
+class LSUNBase(Dataset):
+    def __init__(self,
+                 txt_file,
+                 data_root,
+                 size=None,
+                 interpolation="bicubic",
+                 flip_p=0.5
+                 ):
+        self.data_paths = txt_file
+        self.data_root = data_root
+        with open(self.data_paths, "r") as f:
+            self.image_paths = f.read().splitlines()
+        self._length = len(self.image_paths)
+        self.labels = {
+            "relative_file_path_": [l for l in self.image_paths],
+            "file_path_": [os.path.join(self.data_root, l)
+                           for l in self.image_paths],
+        }
+
+        self.size = size
+        self.interpolation = {"linear": PIL.Image.LINEAR,
+                              "bilinear": PIL.Image.BILINEAR,
+                              "bicubic": PIL.Image.BICUBIC,
+                              "lanczos": PIL.Image.LANCZOS,
+                              }[interpolation]
+        self.flip = transforms.RandomHorizontalFlip(p=flip_p)
+
+    def __len__(self):
+        return self._length
+
+    def __getitem__(self, i):
+        example = dict((k, self.labels[k][i]) for k in self.labels)
+        image = Image.open(example["file_path_"])
+        if not image.mode == "RGB":
+            image = image.convert("RGB")
+
+        # default to score-sde preprocessing
+        img = np.array(image).astype(np.uint8)
+        crop = min(img.shape[0], img.shape[1])
+        h, w, = img.shape[0], img.shape[1]
+        img = img[(h - crop) // 2:(h + crop) // 2,
+              (w - crop) // 2:(w + crop) // 2]
+
+        image = Image.fromarray(img)
+        if self.size is not None:
+            image = image.resize((self.size, self.size), resample=self.interpolation)
+
+        image = self.flip(image)
+        image = np.array(image).astype(np.uint8)
+        example["image"] = (image / 127.5 - 1.0).astype(np.float32)
+        return example
+
+
+class LSUNChurchesTrain(LSUNBase):
+    def __init__(self, **kwargs):
+        super().__init__(txt_file="data/lsun/church_outdoor_train.txt", data_root="data/lsun/churches", **kwargs)
+
+
+class LSUNChurchesValidation(LSUNBase):
+    def __init__(self, flip_p=0., **kwargs):
+        super().__init__(txt_file="data/lsun/church_outdoor_val.txt", data_root="data/lsun/churches",
+                         flip_p=flip_p, **kwargs)
+
+
+class LSUNBedroomsTrain(LSUNBase):
+    def __init__(self, **kwargs):
+        super().__init__(txt_file="data/lsun/bedrooms_train.txt", data_root="data/lsun/bedrooms", **kwargs)
+
+
+class LSUNBedroomsValidation(LSUNBase):
+    def __init__(self, flip_p=0.0, **kwargs):
+        super().__init__(txt_file="data/lsun/bedrooms_val.txt", data_root="data/lsun/bedrooms",
+                         flip_p=flip_p, **kwargs)
+
+
+class LSUNCatsTrain(LSUNBase):
+    def __init__(self, **kwargs):
+        super().__init__(txt_file="data/lsun/cat_train.txt", data_root="data/lsun/cats", **kwargs)
+
+
+class LSUNCatsValidation(LSUNBase):
+    def __init__(self, flip_p=0., **kwargs):
+        super().__init__(txt_file="data/lsun/cat_val.txt", data_root="data/lsun/cats",
+                         flip_p=flip_p, **kwargs)

ldm/data/nerf_like.py

@@ -0,0 +1,165 @@
+from torch.utils.data import Dataset
+import os
+import json
+import numpy as np
+import torch
+import imageio
+import math
+import cv2
+from torchvision import transforms
+
+def cartesian_to_spherical(xyz):
+    ptsnew = np.hstack((xyz, np.zeros(xyz.shape)))
+    xy = xyz[:,0]**2 + xyz[:,1]**2
+    z = np.sqrt(xy + xyz[:,2]**2)
+    theta = np.arctan2(np.sqrt(xy), xyz[:,2]) # for elevation angle defined from Z-axis down
+    #ptsnew[:,4] = np.arctan2(xyz[:,2], np.sqrt(xy)) # for elevation angle defined from XY-plane up
+    azimuth = np.arctan2(xyz[:,1], xyz[:,0])
+    return np.array([theta, azimuth, z])
+
+
+def get_T(T_target, T_cond):
+    theta_cond, azimuth_cond, z_cond = cartesian_to_spherical(T_cond[None, :])
+    theta_target, azimuth_target, z_target = cartesian_to_spherical(T_target[None, :])
+    
+    d_theta = theta_target - theta_cond
+    d_azimuth = (azimuth_target - azimuth_cond) % (2 * math.pi)
+    d_z = z_target - z_cond
+    
+    d_T = torch.tensor([d_theta.item(), math.sin(d_azimuth.item()), math.cos(d_azimuth.item()), d_z.item()])
+    return d_T
+
+def get_spherical(T_target, T_cond):
+    theta_cond, azimuth_cond, z_cond = cartesian_to_spherical(T_cond[None, :])
+    theta_target, azimuth_target, z_target = cartesian_to_spherical(T_target[None, :])
+    
+    d_theta = theta_target - theta_cond
+    d_azimuth = (azimuth_target - azimuth_cond) % (2 * math.pi)
+    d_z = z_target - z_cond
+    
+    d_T = torch.tensor([math.degrees(d_theta.item()), math.degrees(d_azimuth.item()), d_z.item()])
+    return d_T
+
+class RTMV(Dataset):
+    def __init__(self, root_dir='datasets/RTMV/google_scanned',\
+                 first_K=64, resolution=256, load_target=False):
+        self.root_dir = root_dir
+        self.scene_list = sorted(next(os.walk(root_dir))[1])
+        self.resolution = resolution
+        self.first_K = first_K
+        self.load_target = load_target
+
+    def __len__(self):
+        return len(self.scene_list)
+
+    def __getitem__(self, idx):
+        scene_dir = os.path.join(self.root_dir, self.scene_list[idx])
+        with open(os.path.join(scene_dir, 'transforms.json'), "r") as f:
+            meta = json.load(f)
+        imgs = []
+        poses = []
+        for i_img in range(self.first_K):
+            meta_img = meta['frames'][i_img]
+
+            if i_img == 0 or self.load_target:
+                img_path = os.path.join(scene_dir, meta_img['file_path'])
+                img = imageio.imread(img_path)
+                img = cv2.resize(img, (self.resolution, self.resolution), interpolation = cv2.INTER_LINEAR)
+                imgs.append(img)
+            
+            c2w = meta_img['transform_matrix']
+            poses.append(c2w)
+            
+        imgs = (np.array(imgs) / 255.).astype(np.float32)  # (RGBA) imgs
+        imgs = torch.tensor(self.blend_rgba(imgs)).permute(0, 3, 1, 2)
+        imgs = imgs * 2 - 1. # convert to stable diffusion range
+        poses = torch.tensor(np.array(poses).astype(np.float32))
+        return imgs, poses
+                
+    def blend_rgba(self, img):
+        img = img[..., :3] * img[..., -1:] + (1. - img[..., -1:])  # blend A to RGB
+        return img
+            
+
+class GSO(Dataset):
+    def __init__(self, root_dir='datasets/GoogleScannedObjects',\
+                 split='val', first_K=5, resolution=256, load_target=False, name='render_mvs'):
+        self.root_dir = root_dir
+        with open(os.path.join(root_dir, '%s.json' % split), "r") as f:
+            self.scene_list = json.load(f)
+        self.resolution = resolution
+        self.first_K = first_K
+        self.load_target = load_target
+        self.name = name
+
+    def __len__(self):
+        return len(self.scene_list)
+
+    def __getitem__(self, idx):
+        scene_dir = os.path.join(self.root_dir, self.scene_list[idx])
+        with open(os.path.join(scene_dir, 'transforms_%s.json' % self.name), "r") as f:
+            meta = json.load(f)
+        imgs = []
+        poses = []
+        for i_img in range(self.first_K):
+            meta_img = meta['frames'][i_img]
+
+            if i_img == 0 or self.load_target:
+                img_path = os.path.join(scene_dir, meta_img['file_path'])
+                img = imageio.imread(img_path)
+                img = cv2.resize(img, (self.resolution, self.resolution), interpolation = cv2.INTER_LINEAR)
+                imgs.append(img)
+            
+            c2w = meta_img['transform_matrix']
+            poses.append(c2w)
+            
+        imgs = (np.array(imgs) / 255.).astype(np.float32)  # (RGBA) imgs
+        mask = imgs[:, :, :, -1]
+        imgs = torch.tensor(self.blend_rgba(imgs)).permute(0, 3, 1, 2)
+        imgs = imgs * 2 - 1. # convert to stable diffusion range
+        poses = torch.tensor(np.array(poses).astype(np.float32))
+        return imgs, poses
+                
+    def blend_rgba(self, img):
+        img = img[..., :3] * img[..., -1:] + (1. - img[..., -1:])  # blend A to RGB
+        return img
+             
+class WILD(Dataset):
+    def __init__(self, root_dir='data/nerf_wild',\
+                 first_K=33, resolution=256, load_target=False):
+        self.root_dir = root_dir
+        self.scene_list = sorted(next(os.walk(root_dir))[1])
+        self.resolution = resolution
+        self.first_K = first_K
+        self.load_target = load_target
+
+    def __len__(self):
+        return len(self.scene_list)
+
+    def __getitem__(self, idx):
+        scene_dir = os.path.join(self.root_dir, self.scene_list[idx])
+        with open(os.path.join(scene_dir, 'transforms_train.json'), "r") as f:
+            meta = json.load(f)
+        imgs = []
+        poses = []
+        for i_img in range(self.first_K):
+            meta_img = meta['frames'][i_img]
+
+            if i_img == 0 or self.load_target:
+                img_path = os.path.join(scene_dir, meta_img['file_path'])
+                img = imageio.imread(img_path + '.png')
+                img = cv2.resize(img, (self.resolution, self.resolution), interpolation = cv2.INTER_LINEAR)
+                imgs.append(img)
+            
+            c2w = meta_img['transform_matrix']
+            poses.append(c2w)
+            
+        imgs = (np.array(imgs) / 255.).astype(np.float32)  # (RGBA) imgs
+        imgs = torch.tensor(self.blend_rgba(imgs)).permute(0, 3, 1, 2)
+        imgs = imgs * 2 - 1. # convert to stable diffusion range
+        poses = torch.tensor(np.array(poses).astype(np.float32))
+        return imgs, poses
+                
+    def blend_rgba(self, img):
+        img = img[..., :3] * img[..., -1:] + (1. - img[..., -1:])  # blend A to RGB
+        return img

ldm/data/simple.py

@@ -0,0 +1,526 @@
+from typing import Dict
+import webdataset as wds
+import numpy as np
+from omegaconf import DictConfig, ListConfig
+import torch
+from torch.utils.data import Dataset
+from pathlib import Path
+import json
+from PIL import Image
+from torchvision import transforms
+import torchvision
+from einops import rearrange
+from ldm.util import instantiate_from_config
+from datasets import load_dataset
+import pytorch_lightning as pl
+import copy
+import csv
+import cv2
+import random
+import matplotlib.pyplot as plt
+from torch.utils.data import DataLoader
+import json
+import os
+import webdataset as wds
+import math
+from torch.utils.data.distributed import DistributedSampler
+
+# Some hacky things to make experimentation easier
+def make_transform_multi_folder_data(paths, caption_files=None, **kwargs):
+    ds = make_multi_folder_data(paths, caption_files, **kwargs)
+    return TransformDataset(ds)
+
+def make_nfp_data(base_path):
+    dirs = list(Path(base_path).glob("*/"))
+    print(f"Found {len(dirs)} folders")
+    print(dirs)
+    tforms = [transforms.Resize(512), transforms.CenterCrop(512)]
+    datasets = [NfpDataset(x, image_transforms=copy.copy(tforms), default_caption="A view from a train window") for x in dirs]
+    return torch.utils.data.ConcatDataset(datasets)
+
+
+class VideoDataset(Dataset):
+    def __init__(self, root_dir, image_transforms, caption_file, offset=8, n=2):
+        self.root_dir = Path(root_dir)
+        self.caption_file = caption_file
+        self.n = n
+        ext = "mp4"
+        self.paths = sorted(list(self.root_dir.rglob(f"*.{ext}")))
+        self.offset = offset
+
+        if isinstance(image_transforms, ListConfig):
+            image_transforms = [instantiate_from_config(tt) for tt in image_transforms]
+        image_transforms.extend([transforms.ToTensor(),
+                                 transforms.Lambda(lambda x: rearrange(x * 2. - 1., 'c h w -> h w c'))])
+        image_transforms = transforms.Compose(image_transforms)
+        self.tform = image_transforms
+        with open(self.caption_file) as f:
+            reader = csv.reader(f)
+            rows = [row for row in reader]
+        self.captions = dict(rows)
+
+    def __len__(self):
+        return len(self.paths)
+
+    def __getitem__(self, index):
+        for i in range(10):
+            try:
+                return self._load_sample(index)
+            except Exception:
+                # Not really good enough but...
+                print("uh oh")
+
+    def _load_sample(self, index):
+        n = self.n
+        filename = self.paths[index]
+        min_frame = 2*self.offset + 2
+        vid = cv2.VideoCapture(str(filename))
+        max_frames = int(vid.get(cv2.CAP_PROP_FRAME_COUNT))
+        curr_frame_n = random.randint(min_frame, max_frames)
+        vid.set(cv2.CAP_PROP_POS_FRAMES,curr_frame_n)
+        _, curr_frame = vid.read()
+
+        prev_frames = []
+        for i in range(n):
+            prev_frame_n = curr_frame_n - (i+1)*self.offset
+            vid.set(cv2.CAP_PROP_POS_FRAMES,prev_frame_n)
+            _, prev_frame = vid.read()
+            prev_frame = self.tform(Image.fromarray(prev_frame[...,::-1]))
+            prev_frames.append(prev_frame)
+
+        vid.release()
+        caption = self.captions[filename.name]
+        data = {
+            "image": self.tform(Image.fromarray(curr_frame[...,::-1])),
+            "prev": torch.cat(prev_frames, dim=-1),
+            "txt": caption
+        }
+        return data
+
+# end hacky things
+
+
+def make_tranforms(image_transforms):
+    # if isinstance(image_transforms, ListConfig):
+    #     image_transforms = [instantiate_from_config(tt) for tt in image_transforms]
+    image_transforms = []
+    image_transforms.extend([transforms.ToTensor(),
+                                transforms.Lambda(lambda x: rearrange(x * 2. - 1., 'c h w -> h w c'))])
+    image_transforms = transforms.Compose(image_transforms)
+    return image_transforms
+
+
+def make_multi_folder_data(paths, caption_files=None, **kwargs):
+    """Make a concat dataset from multiple folders
+    Don't suport captions yet
+
+    If paths is a list, that's ok, if it's a Dict interpret it as:
+    k=folder v=n_times to repeat that
+    """
+    list_of_paths = []
+    if isinstance(paths, (Dict, DictConfig)):
+        assert caption_files is None, \
+            "Caption files not yet supported for repeats"
+        for folder_path, repeats in paths.items():
+            list_of_paths.extend([folder_path]*repeats)
+        paths = list_of_paths
+
+    if caption_files is not None:
+        datasets = [FolderData(p, caption_file=c, **kwargs) for (p, c) in zip(paths, caption_files)]
+    else:
+        datasets = [FolderData(p, **kwargs) for p in paths]
+    return torch.utils.data.ConcatDataset(datasets)
+
+
+
+class NfpDataset(Dataset):
+    def __init__(self,
+        root_dir,
+        image_transforms=[],
+        ext="jpg",
+        default_caption="",
+        ) -> None:
+        """assume sequential frames and a deterministic transform"""
+
+        self.root_dir = Path(root_dir)
+        self.default_caption = default_caption
+
+        self.paths = sorted(list(self.root_dir.rglob(f"*.{ext}")))
+        self.tform = make_tranforms(image_transforms)
+
+    def __len__(self):
+        return len(self.paths) - 1
+
+
+    def __getitem__(self, index):
+        prev = self.paths[index]
+        curr = self.paths[index+1]
+        data = {}
+        data["image"] = self._load_im(curr)
+        data["prev"] = self._load_im(prev)
+        data["txt"] = self.default_caption
+        return data
+
+    def _load_im(self, filename):
+        im = Image.open(filename).convert("RGB")
+        return self.tform(im)
+
+class ObjaverseDataModuleFromConfig(pl.LightningDataModule):
+    def __init__(self, root_dir, batch_size, total_view, train=None, validation=None,
+                 test=None, num_workers=4, **kwargs):
+        super().__init__(self)
+        self.root_dir = root_dir
+        self.batch_size = batch_size
+        self.num_workers = num_workers
+        self.total_view = total_view
+
+        if train is not None:
+            dataset_config = train
+        if validation is not None:
+            dataset_config = validation
+
+        if 'image_transforms' in dataset_config:
+            image_transforms = [torchvision.transforms.Resize(dataset_config.image_transforms.size)]
+        else:
+            image_transforms = []
+        image_transforms.extend([transforms.ToTensor(),
+                                transforms.Lambda(lambda x: rearrange(x * 2. - 1., 'c h w -> h w c'))])
+        self.image_transforms = torchvision.transforms.Compose(image_transforms)
+
+
+    def train_dataloader(self):
+        dataset = ObjaverseData(root_dir=self.root_dir, total_view=self.total_view, validation=False, \
+                                image_transforms=self.image_transforms)
+        sampler = DistributedSampler(dataset)
+        return wds.WebLoader(dataset, batch_size=self.batch_size, num_workers=self.num_workers, shuffle=False, sampler=sampler)
+
+    def val_dataloader(self):
+        dataset = ObjaverseData(root_dir=self.root_dir, total_view=self.total_view, validation=True, \
+                                image_transforms=self.image_transforms)
+        sampler = DistributedSampler(dataset)
+        return wds.WebLoader(dataset, batch_size=self.batch_size, num_workers=self.num_workers, shuffle=False)
+    
+    def test_dataloader(self):
+        return wds.WebLoader(ObjaverseData(root_dir=self.root_dir, total_view=self.total_view, validation=self.validation),\
+                          batch_size=self.batch_size, num_workers=self.num_workers, shuffle=False)
+
+
+class ObjaverseData(Dataset):
+    def __init__(self,
+        root_dir='.objaverse/hf-objaverse-v1/views',
+        image_transforms=[],
+        ext="png",
+        default_trans=torch.zeros(3),
+        postprocess=None,
+        return_paths=False,
+        total_view=4,
+        validation=False
+        ) -> None:
+        """Create a dataset from a folder of images.
+        If you pass in a root directory it will be searched for images
+        ending in ext (ext can be a list)
+        """
+        self.root_dir = Path(root_dir)
+        self.default_trans = default_trans
+        self.return_paths = return_paths
+        if isinstance(postprocess, DictConfig):
+            postprocess = instantiate_from_config(postprocess)
+        self.postprocess = postprocess
+        self.total_view = total_view
+
+        if not isinstance(ext, (tuple, list, ListConfig)):
+            ext = [ext]
+
+        with open(os.path.join(root_dir, 'valid_paths.json')) as f:
+            self.paths = json.load(f)
+            
+        total_objects = len(self.paths)
+        if validation:
+            self.paths = self.paths[math.floor(total_objects / 100. * 99.):] # used last 1% as validation
+        else:
+            self.paths = self.paths[:math.floor(total_objects / 100. * 99.)] # used first 99% as training
+        print('============= length of dataset %d =============' % len(self.paths))
+        self.tform = image_transforms
+
+    def __len__(self):
+        return len(self.paths)
+        
+    def cartesian_to_spherical(self, xyz):
+        ptsnew = np.hstack((xyz, np.zeros(xyz.shape)))
+        xy = xyz[:,0]**2 + xyz[:,1]**2
+        z = np.sqrt(xy + xyz[:,2]**2)
+        theta = np.arctan2(np.sqrt(xy), xyz[:,2]) # for elevation angle defined from Z-axis down
+        #ptsnew[:,4] = np.arctan2(xyz[:,2], np.sqrt(xy)) # for elevation angle defined from XY-plane up
+        azimuth = np.arctan2(xyz[:,1], xyz[:,0])
+        return np.array([theta, azimuth, z])
+
+    def get_T(self, target_RT, cond_RT):
+        R, T = target_RT[:3, :3], target_RT[:, -1]
+        T_target = -R.T @ T
+
+        R, T = cond_RT[:3, :3], cond_RT[:, -1]
+        T_cond = -R.T @ T
+
+        theta_cond, azimuth_cond, z_cond = self.cartesian_to_spherical(T_cond[None, :])
+        theta_target, azimuth_target, z_target = self.cartesian_to_spherical(T_target[None, :])
+        
+        d_theta = theta_target - theta_cond
+        d_azimuth = (azimuth_target - azimuth_cond) % (2 * math.pi)
+        d_z = z_target - z_cond
+        
+        d_T = torch.tensor([d_theta.item(), math.sin(d_azimuth.item()), math.cos(d_azimuth.item()), d_z.item()])
+        return d_T
+
+    def load_im(self, path, color):
+        '''
+        replace background pixel with random color in rendering
+        '''
+        try:
+            img = plt.imread(path)
+        except:
+            print(path)
+            sys.exit()
+        img[img[:, :, -1] == 0.] = color
+        img = Image.fromarray(np.uint8(img[:, :, :3] * 255.))
+        return img
+
+    def __getitem__(self, index):
+
+        data = {}
+        if self.paths[index][-2:] == '_1': # dirty fix for rendering dataset twice
+            total_view = 8
+        else:
+            total_view = 4
+        index_target, index_cond = random.sample(range(total_view), 2) # without replacement
+        filename = os.path.join(self.root_dir, self.paths[index])
+
+        # print(self.paths[index])
+
+        if self.return_paths:
+            data["path"] = str(filename)
+        
+        color = [1., 1., 1., 1.]
+
+        try:
+            target_im = self.process_im(self.load_im(os.path.join(filename, '%03d.png' % index_target), color))
+            cond_im = self.process_im(self.load_im(os.path.join(filename, '%03d.png' % index_cond), color))
+            target_RT = np.load(os.path.join(filename, '%03d.npy' % index_target))
+            cond_RT = np.load(os.path.join(filename, '%03d.npy' % index_cond))
+        except:
+            # very hacky solution, sorry about this
+            filename = os.path.join(self.root_dir, '692db5f2d3a04bb286cb977a7dba903e_1') # this one we know is valid
+            target_im = self.process_im(self.load_im(os.path.join(filename, '%03d.png' % index_target), color))
+            cond_im = self.process_im(self.load_im(os.path.join(filename, '%03d.png' % index_cond), color))
+            target_RT = np.load(os.path.join(filename, '%03d.npy' % index_target))
+            cond_RT = np.load(os.path.join(filename, '%03d.npy' % index_cond))
+            target_im = torch.zeros_like(target_im)
+            cond_im = torch.zeros_like(cond_im)
+
+        data["image_target"] = target_im
+        data["image_cond"] = cond_im
+        data["T"] = self.get_T(target_RT, cond_RT)
+
+        if self.postprocess is not None:
+            data = self.postprocess(data)
+
+        return data
+
+    def process_im(self, im):
+        im = im.convert("RGB")
+        return self.tform(im)
+
+class FolderData(Dataset):
+    def __init__(self,
+        root_dir,
+        caption_file=None,
+        image_transforms=[],
+        ext="jpg",
+        default_caption="",
+        postprocess=None,
+        return_paths=False,
+        ) -> None:
+        """Create a dataset from a folder of images.
+        If you pass in a root directory it will be searched for images
+        ending in ext (ext can be a list)
+        """
+        self.root_dir = Path(root_dir)
+        self.default_caption = default_caption
+        self.return_paths = return_paths
+        if isinstance(postprocess, DictConfig):
+            postprocess = instantiate_from_config(postprocess)
+        self.postprocess = postprocess
+        if caption_file is not None:
+            with open(caption_file, "rt") as f:
+                ext = Path(caption_file).suffix.lower()
+                if ext == ".json":
+                    captions = json.load(f)
+                elif ext == ".jsonl":
+                    lines = f.readlines()
+                    lines = [json.loads(x) for x in lines]
+                    captions = {x["file_name"]: x["text"].strip("\n") for x in lines}
+                else:
+                    raise ValueError(f"Unrecognised format: {ext}")
+            self.captions = captions
+        else:
+            self.captions = None
+
+        if not isinstance(ext, (tuple, list, ListConfig)):
+            ext = [ext]
+
+        # Only used if there is no caption file
+        self.paths = []
+        for e in ext:
+            self.paths.extend(sorted(list(self.root_dir.rglob(f"*.{e}"))))
+        self.tform = make_tranforms(image_transforms)
+
+    def __len__(self):
+        if self.captions is not None:
+            return len(self.captions.keys())
+        else:
+            return len(self.paths)
+
+    def __getitem__(self, index):
+        data = {}
+        if self.captions is not None:
+            chosen = list(self.captions.keys())[index]
+            caption = self.captions.get(chosen, None)
+            if caption is None:
+                caption = self.default_caption
+            filename = self.root_dir/chosen
+        else:
+            filename = self.paths[index]
+
+        if self.return_paths:
+            data["path"] = str(filename)
+
+        im = Image.open(filename).convert("RGB")
+        im = self.process_im(im)
+        data["image"] = im
+
+        if self.captions is not None:
+            data["txt"] = caption
+        else:
+            data["txt"] = self.default_caption
+
+        if self.postprocess is not None:
+            data = self.postprocess(data)
+
+        return data
+
+    def process_im(self, im):
+        im = im.convert("RGB")
+        return self.tform(im)
+import random
+
+class TransformDataset():
+    def __init__(self, ds, extra_label="sksbspic"):
+        self.ds = ds
+        self.extra_label = extra_label
+        self.transforms = {
+            "align": transforms.Resize(768),
+            "centerzoom": transforms.CenterCrop(768),
+            "randzoom": transforms.RandomCrop(768),
+        }
+
+
+    def __getitem__(self, index):
+        data = self.ds[index]
+
+        im = data['image']
+        im = im.permute(2,0,1)
+        # In case data is smaller than expected
+        im = transforms.Resize(1024)(im)
+
+        tform_name = random.choice(list(self.transforms.keys()))
+        im = self.transforms[tform_name](im)
+
+        im = im.permute(1,2,0)
+
+        data['image'] = im
+        data['txt'] = data['txt'] + f" {self.extra_label} {tform_name}"
+
+        return data
+
+    def __len__(self):
+        return len(self.ds)
+
+def hf_dataset(
+    name,
+    image_transforms=[],
+    image_column="image",
+    text_column="text",
+    split='train',
+    image_key='image',
+    caption_key='txt',
+    ):
+    """Make huggingface dataset with appropriate list of transforms applied
+    """
+    ds = load_dataset(name, split=split)
+    tform = make_tranforms(image_transforms)
+
+    assert image_column in ds.column_names, f"Didn't find column {image_column} in {ds.column_names}"
+    assert text_column in ds.column_names, f"Didn't find column {text_column} in {ds.column_names}"
+
+    def pre_process(examples):
+        processed = {}
+        processed[image_key] = [tform(im) for im in examples[image_column]]
+        processed[caption_key] = examples[text_column]
+        return processed
+
+    ds.set_transform(pre_process)
+    return ds
+
+class TextOnly(Dataset):
+    def __init__(self, captions, output_size, image_key="image", caption_key="txt", n_gpus=1):
+        """Returns only captions with dummy images"""
+        self.output_size = output_size
+        self.image_key = image_key
+        self.caption_key = caption_key
+        if isinstance(captions, Path):
+            self.captions = self._load_caption_file(captions)
+        else:
+            self.captions = captions
+
+        if n_gpus > 1:
+            # hack to make sure that all the captions appear on each gpu
+            repeated = [n_gpus*[x] for x in self.captions]
+            self.captions = []
+            [self.captions.extend(x) for x in repeated]
+
+    def __len__(self):
+        return len(self.captions)
+
+    def __getitem__(self, index):
+        dummy_im = torch.zeros(3, self.output_size, self.output_size)
+        dummy_im = rearrange(dummy_im * 2. - 1., 'c h w -> h w c')
+        return {self.image_key: dummy_im, self.caption_key: self.captions[index]}
+
+    def _load_caption_file(self, filename):
+        with open(filename, 'rt') as f:
+            captions = f.readlines()
+        return [x.strip('\n') for x in captions]
+
+
+
+import random
+import json
+class IdRetreivalDataset(FolderData):
+    def __init__(self, ret_file, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        with open(ret_file, "rt") as f:
+            self.ret = json.load(f)
+
+    def __getitem__(self, index):
+        data = super().__getitem__(index)
+        key = self.paths[index].name
+        matches = self.ret[key]
+        if len(matches) > 0:
+            retreived = random.choice(matches)
+        else:
+            retreived = key
+        filename = self.root_dir/retreived
+        im = Image.open(filename).convert("RGB")
+        im = self.process_im(im)
+        # data["match"] = im
+        data["match"] = torch.cat((data["image"], im), dim=-1)
+        return data

ldm/extras.py

@@ -0,0 +1,77 @@
+from pathlib import Path
+from omegaconf import OmegaConf
+import torch
+from ldm.util import instantiate_from_config
+import logging
+from contextlib import contextmanager
+
+from contextlib import contextmanager
+import logging
+
+@contextmanager
+def all_logging_disabled(highest_level=logging.CRITICAL):
+    """
+    A context manager that will prevent any logging messages
+    triggered during the body from being processed.
+
+    :param highest_level: the maximum logging level in use.
+      This would only need to be changed if a custom level greater than CRITICAL
+      is defined.
+
+    https://gist.github.com/simon-weber/7853144
+    """
+    # two kind-of hacks here:
+    #    * can't get the highest logging level in effect => delegate to the user
+    #    * can't get the current module-level override => use an undocumented
+    #       (but non-private!) interface
+
+    previous_level = logging.root.manager.disable
+
+    logging.disable(highest_level)
+
+    try:
+        yield
+    finally:
+        logging.disable(previous_level)
+
+def load_training_dir(train_dir, device, epoch="last"):
+    """Load a checkpoint and config from training directory"""
+    train_dir = Path(train_dir)
+    ckpt = list(train_dir.rglob(f"*{epoch}.ckpt"))
+    assert len(ckpt) == 1, f"found {len(ckpt)} matching ckpt files"
+    config = list(train_dir.rglob(f"*-project.yaml"))
+    assert len(ckpt) > 0, f"didn't find any config in {train_dir}"
+    if len(config) > 1:
+        print(f"found {len(config)} matching config files")
+        config = sorted(config)[-1]
+        print(f"selecting {config}")
+    else:
+        config = config[0]
+
+
+    config = OmegaConf.load(config)
+    return load_model_from_config(config, ckpt[0], device)
+
+def load_model_from_config(config, ckpt, device="cpu", verbose=False):
+    """Loads a model from config and a ckpt
+    if config is a path will use omegaconf to load
+    """
+    if isinstance(config, (str, Path)):
+        config = OmegaConf.load(config)
+
+    with all_logging_disabled():
+        print(f"Loading model from {ckpt}")
+        pl_sd = torch.load(ckpt, map_location="cpu")
+        global_step = pl_sd["global_step"]
+        sd = pl_sd["state_dict"]
+        model = instantiate_from_config(config.model)
+        m, u = model.load_state_dict(sd, strict=False)
+        if len(m) > 0 and verbose:
+            print("missing keys:")
+            print(m)
+        if len(u) > 0 and verbose:
+            print("unexpected keys:")
+        model.to(device)
+        model.eval()
+        model.cond_stage_model.device = device
+        return model

ldm/guidance.py

@@ -0,0 +1,96 @@
+from typing import List, Tuple
+from scipy import interpolate
+import numpy as np
+import torch
+import matplotlib.pyplot as plt
+from IPython.display import clear_output
+import abc
+
+
+class GuideModel(torch.nn.Module, abc.ABC):
+    def __init__(self) -> None:
+        super().__init__()
+
+    @abc.abstractmethod
+    def preprocess(self, x_img):
+        pass
+
+    @abc.abstractmethod
+    def compute_loss(self, inp):
+        pass
+
+
+class Guider(torch.nn.Module):
+    def __init__(self, sampler, guide_model, scale=1.0, verbose=False):
+        """Apply classifier guidance
+
+        Specify a guidance scale as either a scalar
+        Or a schedule as a list of tuples t = 0->1 and scale, e.g.
+        [(0, 10), (0.5, 20), (1, 50)]
+        """
+        super().__init__()
+        self.sampler = sampler
+        self.index = 0
+        self.show = verbose
+        self.guide_model = guide_model
+        self.history = []
+
+        if isinstance(scale, (Tuple, List)):
+            times = np.array([x[0] for x in scale])
+            values = np.array([x[1] for x in scale])
+            self.scale_schedule = {"times": times, "values": values}
+        else:
+            self.scale_schedule = float(scale)
+
+        self.ddim_timesteps = sampler.ddim_timesteps
+        self.ddpm_num_timesteps = sampler.ddpm_num_timesteps
+
+
+    def get_scales(self):
+        if isinstance(self.scale_schedule, float):
+            return len(self.ddim_timesteps)*[self.scale_schedule]
+
+        interpolater = interpolate.interp1d(self.scale_schedule["times"], self.scale_schedule["values"])
+        fractional_steps = np.array(self.ddim_timesteps)/self.ddpm_num_timesteps
+        return interpolater(fractional_steps)
+
+    def modify_score(self, model, e_t, x, t, c):
+
+        # TODO look up index by t
+        scale = self.get_scales()[self.index]
+
+        if (scale == 0):
+            return e_t
+
+        sqrt_1ma = self.sampler.ddim_sqrt_one_minus_alphas[self.index].to(x.device)
+        with torch.enable_grad():
+            x_in = x.detach().requires_grad_(True)
+            pred_x0 = model.predict_start_from_noise(x_in, t=t, noise=e_t)
+            x_img = model.first_stage_model.decode((1/0.18215)*pred_x0)
+
+            inp = self.guide_model.preprocess(x_img)
+            loss = self.guide_model.compute_loss(inp)
+            grads = torch.autograd.grad(loss.sum(), x_in)[0]
+            correction = grads * scale
+
+            if self.show:
+                clear_output(wait=True)
+                print(loss.item(), scale, correction.abs().max().item(), e_t.abs().max().item())
+                self.history.append([loss.item(), scale, correction.min().item(), correction.max().item()])
+                plt.imshow((inp[0].detach().permute(1,2,0).clamp(-1,1).cpu()+1)/2)
+                plt.axis('off')
+                plt.show()
+                plt.imshow(correction[0][0].detach().cpu())
+                plt.axis('off')
+                plt.show()
+
+
+        e_t_mod = e_t - sqrt_1ma*correction
+        if self.show:
+            fig, axs = plt.subplots(1, 3)
+            axs[0].imshow(e_t[0][0].detach().cpu(), vmin=-2, vmax=+2)
+            axs[1].imshow(e_t_mod[0][0].detach().cpu(), vmin=-2, vmax=+2)
+            axs[2].imshow(correction[0][0].detach().cpu(), vmin=-2, vmax=+2)
+            plt.show()
+        self.index += 1
+        return e_t_mod

ldm/lr_scheduler.py

@@ -0,0 +1,98 @@
+import numpy as np
+
+
+class LambdaWarmUpCosineScheduler:
+    """
+    note: use with a base_lr of 1.0
+    """
+    def __init__(self, warm_up_steps, lr_min, lr_max, lr_start, max_decay_steps, verbosity_interval=0):
+        self.lr_warm_up_steps = warm_up_steps
+        self.lr_start = lr_start
+        self.lr_min = lr_min
+        self.lr_max = lr_max
+        self.lr_max_decay_steps = max_decay_steps
+        self.last_lr = 0.
+        self.verbosity_interval = verbosity_interval
+
+    def schedule(self, n, **kwargs):
+        if self.verbosity_interval > 0:
+            if n % self.verbosity_interval == 0: print(f"current step: {n}, recent lr-multiplier: {self.last_lr}")
+        if n < self.lr_warm_up_steps:
+            lr = (self.lr_max - self.lr_start) / self.lr_warm_up_steps * n + self.lr_start
+            self.last_lr = lr
+            return lr
+        else:
+            t = (n - self.lr_warm_up_steps) / (self.lr_max_decay_steps - self.lr_warm_up_steps)
+            t = min(t, 1.0)
+            lr = self.lr_min + 0.5 * (self.lr_max - self.lr_min) * (
+                    1 + np.cos(t * np.pi))
+            self.last_lr = lr
+            return lr
+
+    def __call__(self, n, **kwargs):
+        return self.schedule(n,**kwargs)
+
+
+class LambdaWarmUpCosineScheduler2:
+    """
+    supports repeated iterations, configurable via lists
+    note: use with a base_lr of 1.0.
+    """
+    def __init__(self, warm_up_steps, f_min, f_max, f_start, cycle_lengths, verbosity_interval=0):
+        assert len(warm_up_steps) == len(f_min) == len(f_max) == len(f_start) == len(cycle_lengths)
+        self.lr_warm_up_steps = warm_up_steps
+        self.f_start = f_start
+        self.f_min = f_min
+        self.f_max = f_max
+        self.cycle_lengths = cycle_lengths
+        self.cum_cycles = np.cumsum([0] + list(self.cycle_lengths))
+        self.last_f = 0.
+        self.verbosity_interval = verbosity_interval
+
+    def find_in_interval(self, n):
+        interval = 0
+        for cl in self.cum_cycles[1:]:
+            if n <= cl:
+                return interval
+            interval += 1
+
+    def schedule(self, n, **kwargs):
+        cycle = self.find_in_interval(n)
+        n = n - self.cum_cycles[cycle]
+        if self.verbosity_interval > 0:
+            if n % self.verbosity_interval == 0: print(f"current step: {n}, recent lr-multiplier: {self.last_f}, "
+                                                       f"current cycle {cycle}")
+        if n < self.lr_warm_up_steps[cycle]:
+            f = (self.f_max[cycle] - self.f_start[cycle]) / self.lr_warm_up_steps[cycle] * n + self.f_start[cycle]
+            self.last_f = f
+            return f
+        else:
+            t = (n - self.lr_warm_up_steps[cycle]) / (self.cycle_lengths[cycle] - self.lr_warm_up_steps[cycle])
+            t = min(t, 1.0)
+            f = self.f_min[cycle] + 0.5 * (self.f_max[cycle] - self.f_min[cycle]) * (
+                    1 + np.cos(t * np.pi))
+            self.last_f = f
+            return f
+
+    def __call__(self, n, **kwargs):
+        return self.schedule(n, **kwargs)
+
+
+class LambdaLinearScheduler(LambdaWarmUpCosineScheduler2):
+
+    def schedule(self, n, **kwargs):
+        cycle = self.find_in_interval(n)
+        n = n - self.cum_cycles[cycle]
+        if self.verbosity_interval > 0:
+            if n % self.verbosity_interval == 0: print(f"current step: {n}, recent lr-multiplier: {self.last_f}, "
+                                                       f"current cycle {cycle}")
+
+        if n < self.lr_warm_up_steps[cycle]:
+            f = (self.f_max[cycle] - self.f_start[cycle]) / self.lr_warm_up_steps[cycle] * n + self.f_start[cycle]
+            self.last_f = f
+            return f
+        else:
+            f = self.f_min[cycle] + (self.f_max[cycle] - self.f_min[cycle]) * (self.cycle_lengths[cycle] - n) / (self.cycle_lengths[cycle])
+            self.last_f = f
+            return f
+

ldm/models/autoencoder.py

@@ -0,0 +1,443 @@
+import torch
+import pytorch_lightning as pl
+import torch.nn.functional as F
+from contextlib import contextmanager
+
+from taming.modules.vqvae.quantize import VectorQuantizer2 as VectorQuantizer
+
+from ldm.modules.diffusionmodules.model import Encoder, Decoder
+from ldm.modules.distributions.distributions import DiagonalGaussianDistribution
+
+from ldm.util import instantiate_from_config
+
+
+class VQModel(pl.LightningModule):
+    def __init__(self,
+                 ddconfig,
+                 lossconfig,
+                 n_embed,
+                 embed_dim,
+                 ckpt_path=None,
+                 ignore_keys=[],
+                 image_key="image",
+                 colorize_nlabels=None,
+                 monitor=None,
+                 batch_resize_range=None,
+                 scheduler_config=None,
+                 lr_g_factor=1.0,
+                 remap=None,
+                 sane_index_shape=False, # tell vector quantizer to return indices as bhw
+                 use_ema=False
+                 ):
+        super().__init__()
+        self.embed_dim = embed_dim
+        self.n_embed = n_embed
+        self.image_key = image_key
+        self.encoder = Encoder(**ddconfig)
+        self.decoder = Decoder(**ddconfig)
+        self.loss = instantiate_from_config(lossconfig)
+        self.quantize = VectorQuantizer(n_embed, embed_dim, beta=0.25,
+                                        remap=remap,
+                                        sane_index_shape=sane_index_shape)
+        self.quant_conv = torch.nn.Conv2d(ddconfig["z_channels"], embed_dim, 1)
+        self.post_quant_conv = torch.nn.Conv2d(embed_dim, ddconfig["z_channels"], 1)
+        if colorize_nlabels is not None:
+            assert type(colorize_nlabels)==int
+            self.register_buffer("colorize", torch.randn(3, colorize_nlabels, 1, 1))
+        if monitor is not None:
+            self.monitor = monitor
+        self.batch_resize_range = batch_resize_range
+        if self.batch_resize_range is not None:
+            print(f"{self.__class__.__name__}: Using per-batch resizing in range {batch_resize_range}.")
+
+        self.use_ema = use_ema
+        if self.use_ema:
+            self.model_ema = LitEma(self)
+            print(f"Keeping EMAs of {len(list(self.model_ema.buffers()))}.")
+
+        if ckpt_path is not None:
+            self.init_from_ckpt(ckpt_path, ignore_keys=ignore_keys)
+        self.scheduler_config = scheduler_config
+        self.lr_g_factor = lr_g_factor
+
+    @contextmanager
+    def ema_scope(self, context=None):
+        if self.use_ema:
+            self.model_ema.store(self.parameters())
+            self.model_ema.copy_to(self)
+            if context is not None:
+                print(f"{context}: Switched to EMA weights")
+        try:
+            yield None
+        finally:
+            if self.use_ema:
+                self.model_ema.restore(self.parameters())
+                if context is not None:
+                    print(f"{context}: Restored training weights")
+
+    def init_from_ckpt(self, path, ignore_keys=list()):
+        sd = torch.load(path, map_location="cpu")["state_dict"]
+        keys = list(sd.keys())
+        for k in keys:
+            for ik in ignore_keys:
+                if k.startswith(ik):
+                    print("Deleting key {} from state_dict.".format(k))
+                    del sd[k]
+        missing, unexpected = self.load_state_dict(sd, strict=False)
+        print(f"Restored from {path} with {len(missing)} missing and {len(unexpected)} unexpected keys")
+        if len(missing) > 0:
+            print(f"Missing Keys: {missing}")
+            print(f"Unexpected Keys: {unexpected}")
+
+    def on_train_batch_end(self, *args, **kwargs):
+        if self.use_ema:
+            self.model_ema(self)
+
+    def encode(self, x):
+        h = self.encoder(x)
+        h = self.quant_conv(h)
+        quant, emb_loss, info = self.quantize(h)
+        return quant, emb_loss, info
+
+    def encode_to_prequant(self, x):
+        h = self.encoder(x)
+        h = self.quant_conv(h)
+        return h
+
+    def decode(self, quant):
+        quant = self.post_quant_conv(quant)
+        dec = self.decoder(quant)
+        return dec
+
+    def decode_code(self, code_b):
+        quant_b = self.quantize.embed_code(code_b)
+        dec = self.decode(quant_b)
+        return dec
+
+    def forward(self, input, return_pred_indices=False):
+        quant, diff, (_,_,ind) = self.encode(input)
+        dec = self.decode(quant)
+        if return_pred_indices:
+            return dec, diff, ind
+        return dec, diff
+
+    def get_input(self, batch, k):
+        x = batch[k]
+        if len(x.shape) == 3:
+            x = x[..., None]
+        x = x.permute(0, 3, 1, 2).to(memory_format=torch.contiguous_format).float()
+        if self.batch_resize_range is not None:
+            lower_size = self.batch_resize_range[0]
+            upper_size = self.batch_resize_range[1]
+            if self.global_step <= 4:
+                # do the first few batches with max size to avoid later oom
+                new_resize = upper_size
+            else:
+                new_resize = np.random.choice(np.arange(lower_size, upper_size+16, 16))
+            if new_resize != x.shape[2]:
+                x = F.interpolate(x, size=new_resize, mode="bicubic")
+            x = x.detach()
+        return x
+
+    def training_step(self, batch, batch_idx, optimizer_idx):
+        # https://github.com/pytorch/pytorch/issues/37142
+        # try not to fool the heuristics
+        x = self.get_input(batch, self.image_key)
+        xrec, qloss, ind = self(x, return_pred_indices=True)
+
+        if optimizer_idx == 0:
+            # autoencode
+            aeloss, log_dict_ae = self.loss(qloss, x, xrec, optimizer_idx, self.global_step,
+                                            last_layer=self.get_last_layer(), split="train",
+                                            predicted_indices=ind)
+
+            self.log_dict(log_dict_ae, prog_bar=False, logger=True, on_step=True, on_epoch=True)
+            return aeloss
+
+        if optimizer_idx == 1:
+            # discriminator
+            discloss, log_dict_disc = self.loss(qloss, x, xrec, optimizer_idx, self.global_step,
+                                            last_layer=self.get_last_layer(), split="train")
+            self.log_dict(log_dict_disc, prog_bar=False, logger=True, on_step=True, on_epoch=True)
+            return discloss
+
+    def validation_step(self, batch, batch_idx):
+        log_dict = self._validation_step(batch, batch_idx)
+        with self.ema_scope():
+            log_dict_ema = self._validation_step(batch, batch_idx, suffix="_ema")
+        return log_dict
+
+    def _validation_step(self, batch, batch_idx, suffix=""):
+        x = self.get_input(batch, self.image_key)
+        xrec, qloss, ind = self(x, return_pred_indices=True)
+        aeloss, log_dict_ae = self.loss(qloss, x, xrec, 0,
+                                        self.global_step,
+                                        last_layer=self.get_last_layer(),
+                                        split="val"+suffix,
+                                        predicted_indices=ind
+                                        )
+
+        discloss, log_dict_disc = self.loss(qloss, x, xrec, 1,
+                                            self.global_step,
+                                            last_layer=self.get_last_layer(),
+                                            split="val"+suffix,
+                                            predicted_indices=ind
+                                            )
+        rec_loss = log_dict_ae[f"val{suffix}/rec_loss"]
+        self.log(f"val{suffix}/rec_loss", rec_loss,
+                   prog_bar=True, logger=True, on_step=False, on_epoch=True, sync_dist=True)
+        self.log(f"val{suffix}/aeloss", aeloss,
+                   prog_bar=True, logger=True, on_step=False, on_epoch=True, sync_dist=True)
+        if version.parse(pl.__version__) >= version.parse('1.4.0'):
+            del log_dict_ae[f"val{suffix}/rec_loss"]
+        self.log_dict(log_dict_ae)
+        self.log_dict(log_dict_disc)
+        return self.log_dict
+
+    def configure_optimizers(self):
+        lr_d = self.learning_rate
+        lr_g = self.lr_g_factor*self.learning_rate
+        print("lr_d", lr_d)
+        print("lr_g", lr_g)
+        opt_ae = torch.optim.Adam(list(self.encoder.parameters())+
+                                  list(self.decoder.parameters())+
+                                  list(self.quantize.parameters())+
+                                  list(self.quant_conv.parameters())+
+                                  list(self.post_quant_conv.parameters()),
+                                  lr=lr_g, betas=(0.5, 0.9))
+        opt_disc = torch.optim.Adam(self.loss.discriminator.parameters(),
+                                    lr=lr_d, betas=(0.5, 0.9))
+
+        if self.scheduler_config is not None:
+            scheduler = instantiate_from_config(self.scheduler_config)
+
+            print("Setting up LambdaLR scheduler...")
+            scheduler = [
+                {
+                    'scheduler': LambdaLR(opt_ae, lr_lambda=scheduler.schedule),
+                    'interval': 'step',
+                    'frequency': 1
+                },
+                {
+                    'scheduler': LambdaLR(opt_disc, lr_lambda=scheduler.schedule),
+                    'interval': 'step',
+                    'frequency': 1
+                },
+            ]
+            return [opt_ae, opt_disc], scheduler
+        return [opt_ae, opt_disc], []
+
+    def get_last_layer(self):
+        return self.decoder.conv_out.weight
+
+    def log_images(self, batch, only_inputs=False, plot_ema=False, **kwargs):
+        log = dict()
+        x = self.get_input(batch, self.image_key)
+        x = x.to(self.device)
+        if only_inputs:
+            log["inputs"] = x
+            return log
+        xrec, _ = self(x)
+        if x.shape[1] > 3:
+            # colorize with random projection
+            assert xrec.shape[1] > 3
+            x = self.to_rgb(x)
+            xrec = self.to_rgb(xrec)
+        log["inputs"] = x
+        log["reconstructions"] = xrec
+        if plot_ema:
+            with self.ema_scope():
+                xrec_ema, _ = self(x)
+                if x.shape[1] > 3: xrec_ema = self.to_rgb(xrec_ema)
+                log["reconstructions_ema"] = xrec_ema
+        return log
+
+    def to_rgb(self, x):
+        assert self.image_key == "segmentation"
+        if not hasattr(self, "colorize"):
+            self.register_buffer("colorize", torch.randn(3, x.shape[1], 1, 1).to(x))
+        x = F.conv2d(x, weight=self.colorize)
+        x = 2.*(x-x.min())/(x.max()-x.min()) - 1.
+        return x
+
+
+class VQModelInterface(VQModel):
+    def __init__(self, embed_dim, *args, **kwargs):
+        super().__init__(embed_dim=embed_dim, *args, **kwargs)
+        self.embed_dim = embed_dim
+
+    def encode(self, x):
+        h = self.encoder(x)
+        h = self.quant_conv(h)
+        return h
+
+    def decode(self, h, force_not_quantize=False):
+        # also go through quantization layer
+        if not force_not_quantize:
+            quant, emb_loss, info = self.quantize(h)
+        else:
+            quant = h
+        quant = self.post_quant_conv(quant)
+        dec = self.decoder(quant)
+        return dec
+
+
+class AutoencoderKL(pl.LightningModule):
+    def __init__(self,
+                 ddconfig,
+                 lossconfig,
+                 embed_dim,
+                 ckpt_path=None,
+                 ignore_keys=[],
+                 image_key="image",
+                 colorize_nlabels=None,
+                 monitor=None,
+                 ):
+        super().__init__()
+        self.image_key = image_key
+        self.encoder = Encoder(**ddconfig)
+        self.decoder = Decoder(**ddconfig)
+        self.loss = instantiate_from_config(lossconfig)
+        assert ddconfig["double_z"]
+        self.quant_conv = torch.nn.Conv2d(2*ddconfig["z_channels"], 2*embed_dim, 1)
+        self.post_quant_conv = torch.nn.Conv2d(embed_dim, ddconfig["z_channels"], 1)
+        self.embed_dim = embed_dim
+        if colorize_nlabels is not None:
+            assert type(colorize_nlabels)==int
+            self.register_buffer("colorize", torch.randn(3, colorize_nlabels, 1, 1))
+        if monitor is not None:
+            self.monitor = monitor
+        if ckpt_path is not None:
+            self.init_from_ckpt(ckpt_path, ignore_keys=ignore_keys)
+
+    def init_from_ckpt(self, path, ignore_keys=list()):
+        sd = torch.load(path, map_location="cpu")["state_dict"]
+        keys = list(sd.keys())
+        for k in keys:
+            for ik in ignore_keys:
+                if k.startswith(ik):
+                    print("Deleting key {} from state_dict.".format(k))
+                    del sd[k]
+        self.load_state_dict(sd, strict=False)
+        print(f"Restored from {path}")
+
+    def encode(self, x):
+        h = self.encoder(x)
+        moments = self.quant_conv(h)
+        posterior = DiagonalGaussianDistribution(moments)
+        return posterior
+
+    def decode(self, z):
+        z = self.post_quant_conv(z)
+        dec = self.decoder(z)
+        return dec
+
+    def forward(self, input, sample_posterior=True):
+        posterior = self.encode(input)
+        if sample_posterior:
+            z = posterior.sample()
+        else:
+            z = posterior.mode()
+        dec = self.decode(z)
+        return dec, posterior
+
+    def get_input(self, batch, k):
+        x = batch[k]
+        if len(x.shape) == 3:
+            x = x[..., None]
+        x = x.permute(0, 3, 1, 2).to(memory_format=torch.contiguous_format).float()
+        return x
+
+    def training_step(self, batch, batch_idx, optimizer_idx):
+        inputs = self.get_input(batch, self.image_key)
+        reconstructions, posterior = self(inputs)
+
+        if optimizer_idx == 0:
+            # train encoder+decoder+logvar
+            aeloss, log_dict_ae = self.loss(inputs, reconstructions, posterior, optimizer_idx, self.global_step,
+                                            last_layer=self.get_last_layer(), split="train")
+            self.log("aeloss", aeloss, prog_bar=True, logger=True, on_step=True, on_epoch=True)
+            self.log_dict(log_dict_ae, prog_bar=False, logger=True, on_step=True, on_epoch=False)
+            return aeloss
+
+        if optimizer_idx == 1:
+            # train the discriminator
+            discloss, log_dict_disc = self.loss(inputs, reconstructions, posterior, optimizer_idx, self.global_step,
+                                                last_layer=self.get_last_layer(), split="train")
+
+            self.log("discloss", discloss, prog_bar=True, logger=True, on_step=True, on_epoch=True)
+            self.log_dict(log_dict_disc, prog_bar=False, logger=True, on_step=True, on_epoch=False)
+            return discloss
+
+    def validation_step(self, batch, batch_idx):
+        inputs = self.get_input(batch, self.image_key)
+        reconstructions, posterior = self(inputs)
+        aeloss, log_dict_ae = self.loss(inputs, reconstructions, posterior, 0, self.global_step,
+                                        last_layer=self.get_last_layer(), split="val")
+
+        discloss, log_dict_disc = self.loss(inputs, reconstructions, posterior, 1, self.global_step,
+                                            last_layer=self.get_last_layer(), split="val")
+
+        self.log("val/rec_loss", log_dict_ae["val/rec_loss"])
+        self.log_dict(log_dict_ae)
+        self.log_dict(log_dict_disc)
+        return self.log_dict
+
+    def configure_optimizers(self):
+        lr = self.learning_rate
+        opt_ae = torch.optim.Adam(list(self.encoder.parameters())+
+                                  list(self.decoder.parameters())+
+                                  list(self.quant_conv.parameters())+
+                                  list(self.post_quant_conv.parameters()),
+                                  lr=lr, betas=(0.5, 0.9))
+        opt_disc = torch.optim.Adam(self.loss.discriminator.parameters(),
+                                    lr=lr, betas=(0.5, 0.9))
+        return [opt_ae, opt_disc], []
+
+    def get_last_layer(self):
+        return self.decoder.conv_out.weight
+
+    @torch.no_grad()
+    def log_images(self, batch, only_inputs=False, **kwargs):
+        log = dict()
+        x = self.get_input(batch, self.image_key)
+        x = x.to(self.device)
+        if not only_inputs:
+            xrec, posterior = self(x)
+            if x.shape[1] > 3:
+                # colorize with random projection
+                assert xrec.shape[1] > 3
+                x = self.to_rgb(x)
+                xrec = self.to_rgb(xrec)
+            log["samples"] = self.decode(torch.randn_like(posterior.sample()))
+            log["reconstructions"] = xrec
+        log["inputs"] = x
+        return log
+
+    def to_rgb(self, x):
+        assert self.image_key == "segmentation"
+        if not hasattr(self, "colorize"):
+            self.register_buffer("colorize", torch.randn(3, x.shape[1], 1, 1).to(x))
+        x = F.conv2d(x, weight=self.colorize)
+        x = 2.*(x-x.min())/(x.max()-x.min()) - 1.
+        return x
+
+
+class IdentityFirstStage(torch.nn.Module):
+    def __init__(self, *args, vq_interface=False, **kwargs):
+        self.vq_interface = vq_interface  # TODO: Should be true by default but check to not break older stuff
+        super().__init__()
+
+    def encode(self, x, *args, **kwargs):
+        return x
+
+    def decode(self, x, *args, **kwargs):
+        return x
+
+    def quantize(self, x, *args, **kwargs):
+        if self.vq_interface:
+            return x, None, [None, None, None]
+        return x
+
+    def forward(self, x, *args, **kwargs):
+        return x

ldm/models/diffusion/classifier.py

@@ -0,0 +1,267 @@
+import os
+import torch
+import pytorch_lightning as pl
+from omegaconf import OmegaConf
+from torch.nn import functional as F
+from torch.optim import AdamW
+from torch.optim.lr_scheduler import LambdaLR
+from copy import deepcopy
+from einops import rearrange
+from glob import glob
+from natsort import natsorted
+
+from ldm.modules.diffusionmodules.openaimodel import EncoderUNetModel, UNetModel
+from ldm.util import log_txt_as_img, default, ismap, instantiate_from_config
+
+__models__ = {
+    'class_label': EncoderUNetModel,
+    'segmentation': UNetModel
+}
+
+
+def disabled_train(self, mode=True):
+    """Overwrite model.train with this function to make sure train/eval mode
+    does not change anymore."""
+    return self
+
+
+class NoisyLatentImageClassifier(pl.LightningModule):
+
+    def __init__(self,
+                 diffusion_path,
+                 num_classes,
+                 ckpt_path=None,
+                 pool='attention',
+                 label_key=None,
+                 diffusion_ckpt_path=None,
+                 scheduler_config=None,
+                 weight_decay=1.e-2,
+                 log_steps=10,
+                 monitor='val/loss',
+                 *args,
+                 **kwargs):
+        super().__init__(*args, **kwargs)
+        self.num_classes = num_classes
+        # get latest config of diffusion model
+        diffusion_config = natsorted(glob(os.path.join(diffusion_path, 'configs', '*-project.yaml')))[-1]
+        self.diffusion_config = OmegaConf.load(diffusion_config).model
+        self.diffusion_config.params.ckpt_path = diffusion_ckpt_path
+        self.load_diffusion()
+
+        self.monitor = monitor
+        self.numd = self.diffusion_model.first_stage_model.encoder.num_resolutions - 1
+        self.log_time_interval = self.diffusion_model.num_timesteps // log_steps
+        self.log_steps = log_steps
+
+        self.label_key = label_key if not hasattr(self.diffusion_model, 'cond_stage_key') \
+            else self.diffusion_model.cond_stage_key
+
+        assert self.label_key is not None, 'label_key neither in diffusion model nor in model.params'
+
+        if self.label_key not in __models__:
+            raise NotImplementedError()
+
+        self.load_classifier(ckpt_path, pool)
+
+        self.scheduler_config = scheduler_config
+        self.use_scheduler = self.scheduler_config is not None
+        self.weight_decay = weight_decay
+
+    def init_from_ckpt(self, path, ignore_keys=list(), only_model=False):
+        sd = torch.load(path, map_location="cpu")
+        if "state_dict" in list(sd.keys()):
+            sd = sd["state_dict"]
+        keys = list(sd.keys())
+        for k in keys:
+            for ik in ignore_keys:
+                if k.startswith(ik):
+                    print("Deleting key {} from state_dict.".format(k))
+                    del sd[k]
+        missing, unexpected = self.load_state_dict(sd, strict=False) if not only_model else self.model.load_state_dict(
+            sd, strict=False)
+        print(f"Restored from {path} with {len(missing)} missing and {len(unexpected)} unexpected keys")
+        if len(missing) > 0:
+            print(f"Missing Keys: {missing}")
+        if len(unexpected) > 0:
+            print(f"Unexpected Keys: {unexpected}")
+
+    def load_diffusion(self):
+        model = instantiate_from_config(self.diffusion_config)
+        self.diffusion_model = model.eval()
+        self.diffusion_model.train = disabled_train
+        for param in self.diffusion_model.parameters():
+            param.requires_grad = False
+
+    def load_classifier(self, ckpt_path, pool):
+        model_config = deepcopy(self.diffusion_config.params.unet_config.params)
+        model_config.in_channels = self.diffusion_config.params.unet_config.params.out_channels
+        model_config.out_channels = self.num_classes
+        if self.label_key == 'class_label':
+            model_config.pool = pool
+
+        self.model = __models__[self.label_key](**model_config)
+        if ckpt_path is not None:
+            print('#####################################################################')
+            print(f'load from ckpt "{ckpt_path}"')
+            print('#####################################################################')
+            self.init_from_ckpt(ckpt_path)
+
+    @torch.no_grad()
+    def get_x_noisy(self, x, t, noise=None):
+        noise = default(noise, lambda: torch.randn_like(x))
+        continuous_sqrt_alpha_cumprod = None
+        if self.diffusion_model.use_continuous_noise:
+            continuous_sqrt_alpha_cumprod = self.diffusion_model.sample_continuous_noise_level(x.shape[0], t + 1)
+            # todo: make sure t+1 is correct here
+
+        return self.diffusion_model.q_sample(x_start=x, t=t, noise=noise,
+                                             continuous_sqrt_alpha_cumprod=continuous_sqrt_alpha_cumprod)
+
+    def forward(self, x_noisy, t, *args, **kwargs):
+        return self.model(x_noisy, t)
+
+    @torch.no_grad()
+    def get_input(self, batch, k):
+        x = batch[k]
+        if len(x.shape) == 3:
+            x = x[..., None]
+        x = rearrange(x, 'b h w c -> b c h w')
+        x = x.to(memory_format=torch.contiguous_format).float()
+        return x
+
+    @torch.no_grad()
+    def get_conditioning(self, batch, k=None):
+        if k is None:
+            k = self.label_key
+        assert k is not None, 'Needs to provide label key'
+
+        targets = batch[k].to(self.device)
+
+        if self.label_key == 'segmentation':
+            targets = rearrange(targets, 'b h w c -> b c h w')
+            for down in range(self.numd):
+                h, w = targets.shape[-2:]
+                targets = F.interpolate(targets, size=(h // 2, w // 2), mode='nearest')
+
+            # targets = rearrange(targets,'b c h w -> b h w c')
+
+        return targets
+
+    def compute_top_k(self, logits, labels, k, reduction="mean"):
+        _, top_ks = torch.topk(logits, k, dim=1)
+        if reduction == "mean":
+            return (top_ks == labels[:, None]).float().sum(dim=-1).mean().item()
+        elif reduction == "none":
+            return (top_ks == labels[:, None]).float().sum(dim=-1)
+
+    def on_train_epoch_start(self):
+        # save some memory
+        self.diffusion_model.model.to('cpu')
+
+    @torch.no_grad()
+    def write_logs(self, loss, logits, targets):
+        log_prefix = 'train' if self.training else 'val'
+        log = {}
+        log[f"{log_prefix}/loss"] = loss.mean()
+        log[f"{log_prefix}/acc@1"] = self.compute_top_k(
+            logits, targets, k=1, reduction="mean"
+        )
+        log[f"{log_prefix}/acc@5"] = self.compute_top_k(
+            logits, targets, k=5, reduction="mean"
+        )
+
+        self.log_dict(log, prog_bar=False, logger=True, on_step=self.training, on_epoch=True)
+        self.log('loss', log[f"{log_prefix}/loss"], prog_bar=True, logger=False)
+        self.log('global_step', self.global_step, logger=False, on_epoch=False, prog_bar=True)
+        lr = self.optimizers().param_groups[0]['lr']
+        self.log('lr_abs', lr, on_step=True, logger=True, on_epoch=False, prog_bar=True)
+
+    def shared_step(self, batch, t=None):
+        x, *_ = self.diffusion_model.get_input(batch, k=self.diffusion_model.first_stage_key)
+        targets = self.get_conditioning(batch)
+        if targets.dim() == 4:
+            targets = targets.argmax(dim=1)
+        if t is None:
+            t = torch.randint(0, self.diffusion_model.num_timesteps, (x.shape[0],), device=self.device).long()
+        else:
+            t = torch.full(size=(x.shape[0],), fill_value=t, device=self.device).long()
+        x_noisy = self.get_x_noisy(x, t)
+        logits = self(x_noisy, t)
+
+        loss = F.cross_entropy(logits, targets, reduction='none')
+
+        self.write_logs(loss.detach(), logits.detach(), targets.detach())
+
+        loss = loss.mean()
+        return loss, logits, x_noisy, targets
+
+    def training_step(self, batch, batch_idx):
+        loss, *_ = self.shared_step(batch)
+        return loss
+
+    def reset_noise_accs(self):
+        self.noisy_acc = {t: {'acc@1': [], 'acc@5': []} for t in
+                          range(0, self.diffusion_model.num_timesteps, self.diffusion_model.log_every_t)}
+
+    def on_validation_start(self):
+        self.reset_noise_accs()
+
+    @torch.no_grad()
+    def validation_step(self, batch, batch_idx):
+        loss, *_ = self.shared_step(batch)
+
+        for t in self.noisy_acc:
+            _, logits, _, targets = self.shared_step(batch, t)
+            self.noisy_acc[t]['acc@1'].append(self.compute_top_k(logits, targets, k=1, reduction='mean'))
+            self.noisy_acc[t]['acc@5'].append(self.compute_top_k(logits, targets, k=5, reduction='mean'))
+
+        return loss
+
+    def configure_optimizers(self):
+        optimizer = AdamW(self.model.parameters(), lr=self.learning_rate, weight_decay=self.weight_decay)
+
+        if self.use_scheduler:
+            scheduler = instantiate_from_config(self.scheduler_config)
+
+            print("Setting up LambdaLR scheduler...")
+            scheduler = [
+                {
+                    'scheduler': LambdaLR(optimizer, lr_lambda=scheduler.schedule),
+                    'interval': 'step',
+                    'frequency': 1
+                }]
+            return [optimizer], scheduler
+
+        return optimizer
+
+    @torch.no_grad()
+    def log_images(self, batch, N=8, *args, **kwargs):
+        log = dict()
+        x = self.get_input(batch, self.diffusion_model.first_stage_key)
+        log['inputs'] = x
+
+        y = self.get_conditioning(batch)
+
+        if self.label_key == 'class_label':
+            y = log_txt_as_img((x.shape[2], x.shape[3]), batch["human_label"])
+            log['labels'] = y
+
+        if ismap(y):
+            log['labels'] = self.diffusion_model.to_rgb(y)
+
+            for step in range(self.log_steps):
+                current_time = step * self.log_time_interval
+
+                _, logits, x_noisy, _ = self.shared_step(batch, t=current_time)
+
+                log[f'inputs@t{current_time}'] = x_noisy
+
+                pred = F.one_hot(logits.argmax(dim=1), num_classes=self.num_classes)
+                pred = rearrange(pred, 'b h w c -> b c h w')
+
+                log[f'pred@t{current_time}'] = self.diffusion_model.to_rgb(pred)
+
+        for key in log:
+            log[key] = log[key][:N]
+
+        return log

ldm/models/diffusion/ddim.py

@@ -0,0 +1,324 @@
+"""SAMPLING ONLY."""
+
+import torch
+import numpy as np
+from tqdm import tqdm
+from functools import partial
+from einops import rearrange
+
+from ldm.modules.diffusionmodules.util import make_ddim_sampling_parameters, make_ddim_timesteps, noise_like, extract_into_tensor
+from ldm.models.diffusion.sampling_util import renorm_thresholding, norm_thresholding, spatial_norm_thresholding
+
+
+class DDIMSampler(object):
+    def __init__(self, model, schedule="linear", **kwargs):
+        super().__init__()
+        self.model = model
+        self.ddpm_num_timesteps = model.num_timesteps
+        self.schedule = schedule
+
+    def to(self, device):
+        """Same as to in torch module
+        Don't really underestand why this isn't a module in the first place"""
+        for k, v in self.__dict__.items():
+            if isinstance(v, torch.Tensor):
+                new_v = getattr(self, k).to(device)
+                setattr(self, k, new_v)
+
+
+    def register_buffer(self, name, attr):
+        if type(attr) == torch.Tensor:
+            if attr.device != torch.device("cuda"):
+                attr = attr.to(torch.device("cuda"))
+        setattr(self, name, attr)
+
+    def make_schedule(self, ddim_num_steps, ddim_discretize="uniform", ddim_eta=0., verbose=True):
+        self.ddim_timesteps = make_ddim_timesteps(ddim_discr_method=ddim_discretize, num_ddim_timesteps=ddim_num_steps,
+                                                  num_ddpm_timesteps=self.ddpm_num_timesteps,verbose=verbose)
+        alphas_cumprod = self.model.alphas_cumprod
+        assert alphas_cumprod.shape[0] == self.ddpm_num_timesteps, 'alphas have to be defined for each timestep'
+        to_torch = lambda x: x.clone().detach().to(torch.float32).to(self.model.device)
+
+        self.register_buffer('betas', to_torch(self.model.betas))
+        self.register_buffer('alphas_cumprod', to_torch(alphas_cumprod))
+        self.register_buffer('alphas_cumprod_prev', to_torch(self.model.alphas_cumprod_prev))
+
+        # calculations for diffusion q(x_t | x_{t-1}) and others
+        self.register_buffer('sqrt_alphas_cumprod', to_torch(np.sqrt(alphas_cumprod.cpu())))
+        self.register_buffer('sqrt_one_minus_alphas_cumprod', to_torch(np.sqrt(1. - alphas_cumprod.cpu())))
+        self.register_buffer('log_one_minus_alphas_cumprod', to_torch(np.log(1. - alphas_cumprod.cpu())))
+        self.register_buffer('sqrt_recip_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod.cpu())))
+        self.register_buffer('sqrt_recipm1_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod.cpu() - 1)))
+
+        # ddim sampling parameters
+        ddim_sigmas, ddim_alphas, ddim_alphas_prev = make_ddim_sampling_parameters(alphacums=alphas_cumprod.cpu(),
+                                                                                   ddim_timesteps=self.ddim_timesteps,
+                                                                                   eta=ddim_eta,verbose=verbose)
+        self.register_buffer('ddim_sigmas', ddim_sigmas)
+        self.register_buffer('ddim_alphas', ddim_alphas)
+        self.register_buffer('ddim_alphas_prev', ddim_alphas_prev)
+        self.register_buffer('ddim_sqrt_one_minus_alphas', np.sqrt(1. - ddim_alphas))
+        sigmas_for_original_sampling_steps = ddim_eta * torch.sqrt(
+            (1 - self.alphas_cumprod_prev) / (1 - self.alphas_cumprod) * (
+                        1 - self.alphas_cumprod / self.alphas_cumprod_prev))
+        self.register_buffer('ddim_sigmas_for_original_num_steps', sigmas_for_original_sampling_steps)
+
+    @torch.no_grad()
+    def sample(self,
+               S,
+               batch_size,
+               shape,
+               conditioning=None,
+               callback=None,
+               normals_sequence=None,
+               img_callback=None,
+               quantize_x0=False,
+               eta=0.,
+               mask=None,
+               x0=None,
+               temperature=1.,
+               noise_dropout=0.,
+               score_corrector=None,
+               corrector_kwargs=None,
+               verbose=True,
+               x_T=None,
+               log_every_t=100,
+               unconditional_guidance_scale=1.,
+               unconditional_conditioning=None, # this has to come in the same format as the conditioning, # e.g. as encoded tokens, ...
+               dynamic_threshold=None,
+               **kwargs
+               ):
+        if conditioning is not None:
+            if isinstance(conditioning, dict):
+                ctmp = conditioning[list(conditioning.keys())[0]]
+                while isinstance(ctmp, list): ctmp = ctmp[0]
+                cbs = ctmp.shape[0]
+                if cbs != batch_size:
+                    print(f"Warning: Got {cbs} conditionings but batch-size is {batch_size}")
+
+            else:
+                if conditioning.shape[0] != batch_size:
+                    print(f"Warning: Got {conditioning.shape[0]} conditionings but batch-size is {batch_size}")
+
+        self.make_schedule(ddim_num_steps=S, ddim_eta=eta, verbose=verbose)
+        # sampling
+        C, H, W = shape
+        size = (batch_size, C, H, W)
+        print(f'Data shape for DDIM sampling is {size}, eta {eta}')
+
+        samples, intermediates = self.ddim_sampling(conditioning, size,
+                                                    callback=callback,
+                                                    img_callback=img_callback,
+                                                    quantize_denoised=quantize_x0,
+                                                    mask=mask, x0=x0,
+                                                    ddim_use_original_steps=False,
+                                                    noise_dropout=noise_dropout,
+                                                    temperature=temperature,
+                                                    score_corrector=score_corrector,
+                                                    corrector_kwargs=corrector_kwargs,
+                                                    x_T=x_T,
+                                                    log_every_t=log_every_t,
+                                                    unconditional_guidance_scale=unconditional_guidance_scale,
+                                                    unconditional_conditioning=unconditional_conditioning,
+                                                    dynamic_threshold=dynamic_threshold,
+                                                    )
+        return samples, intermediates
+
+    @torch.no_grad()
+    def ddim_sampling(self, cond, shape,
+                      x_T=None, ddim_use_original_steps=False,
+                      callback=None, timesteps=None, quantize_denoised=False,
+                      mask=None, x0=None, img_callback=None, log_every_t=100,
+                      temperature=1., noise_dropout=0., score_corrector=None, corrector_kwargs=None,
+                      unconditional_guidance_scale=1., unconditional_conditioning=None, dynamic_threshold=None,
+                      t_start=-1):
+        device = self.model.betas.device
+        b = shape[0]
+        if x_T is None:
+            img = torch.randn(shape, device=device)
+        else:
+            img = x_T
+
+        if timesteps is None:
+            timesteps = self.ddpm_num_timesteps if ddim_use_original_steps else self.ddim_timesteps
+        elif timesteps is not None and not ddim_use_original_steps:
+            subset_end = int(min(timesteps / self.ddim_timesteps.shape[0], 1) * self.ddim_timesteps.shape[0]) - 1
+            timesteps = self.ddim_timesteps[:subset_end]
+
+        timesteps = timesteps[:t_start]
+
+        intermediates = {'x_inter': [img], 'pred_x0': [img]}
+        time_range = reversed(range(0,timesteps)) if ddim_use_original_steps else np.flip(timesteps)
+        total_steps = timesteps if ddim_use_original_steps else timesteps.shape[0]
+        print(f"Running DDIM Sampling with {total_steps} timesteps")
+
+        iterator = tqdm(time_range, desc='DDIM Sampler', total=total_steps)
+
+        for i, step in enumerate(iterator):
+            index = total_steps - i - 1
+            ts = torch.full((b,), step, device=device, dtype=torch.long)
+
+            if mask is not None:
+                assert x0 is not None
+                img_orig = self.model.q_sample(x0, ts)  # TODO: deterministic forward pass?
+                img = img_orig * mask + (1. - mask) * img
+
+            outs = self.p_sample_ddim(img, cond, ts, index=index, use_original_steps=ddim_use_original_steps,
+                                      quantize_denoised=quantize_denoised, temperature=temperature,
+                                      noise_dropout=noise_dropout, score_corrector=score_corrector,
+                                      corrector_kwargs=corrector_kwargs,
+                                      unconditional_guidance_scale=unconditional_guidance_scale,
+                                      unconditional_conditioning=unconditional_conditioning,
+                                      dynamic_threshold=dynamic_threshold)
+            img, pred_x0 = outs
+            if callback:
+                img = callback(i, img, pred_x0)
+            if img_callback: img_callback(pred_x0, i)
+
+            if index % log_every_t == 0 or index == total_steps - 1:
+                intermediates['x_inter'].append(img)
+                intermediates['pred_x0'].append(pred_x0)
+
+        return img, intermediates
+
+    @torch.no_grad()
+    def p_sample_ddim(self, x, c, t, index, repeat_noise=False, use_original_steps=False, quantize_denoised=False,
+                      temperature=1., noise_dropout=0., score_corrector=None, corrector_kwargs=None,
+                      unconditional_guidance_scale=1., unconditional_conditioning=None,
+                      dynamic_threshold=None):
+        b, *_, device = *x.shape, x.device
+
+        if unconditional_conditioning is None or unconditional_guidance_scale == 1.:
+            e_t = self.model.apply_model(x, t, c)
+        else:
+            x_in = torch.cat([x] * 2)
+            t_in = torch.cat([t] * 2)
+            if isinstance(c, dict):
+                assert isinstance(unconditional_conditioning, dict)
+                c_in = dict()
+                for k in c:
+                    if isinstance(c[k], list):
+                        c_in[k] = [torch.cat([
+                            unconditional_conditioning[k][i],
+                            c[k][i]]) for i in range(len(c[k]))]
+                    else:
+                        c_in[k] = torch.cat([
+                                unconditional_conditioning[k],
+                                c[k]])
+            else:
+                c_in = torch.cat([unconditional_conditioning, c])
+            e_t_uncond, e_t = self.model.apply_model(x_in, t_in, c_in).chunk(2)
+            e_t = e_t_uncond + unconditional_guidance_scale * (e_t - e_t_uncond)
+
+        if score_corrector is not None:
+            assert self.model.parameterization == "eps"
+            e_t = score_corrector.modify_score(self.model, e_t, x, t, c, **corrector_kwargs)
+
+        alphas = self.model.alphas_cumprod if use_original_steps else self.ddim_alphas
+        alphas_prev = self.model.alphas_cumprod_prev if use_original_steps else self.ddim_alphas_prev
+        sqrt_one_minus_alphas = self.model.sqrt_one_minus_alphas_cumprod if use_original_steps else self.ddim_sqrt_one_minus_alphas
+        sigmas = self.model.ddim_sigmas_for_original_num_steps if use_original_steps else self.ddim_sigmas
+        # select parameters corresponding to the currently considered timestep
+        a_t = torch.full((b, 1, 1, 1), alphas[index], device=device)
+        a_prev = torch.full((b, 1, 1, 1), alphas_prev[index], device=device)
+        sigma_t = torch.full((b, 1, 1, 1), sigmas[index], device=device)
+        sqrt_one_minus_at = torch.full((b, 1, 1, 1), sqrt_one_minus_alphas[index],device=device)
+
+        # current prediction for x_0
+        pred_x0 = (x - sqrt_one_minus_at * e_t) / a_t.sqrt()
+        if quantize_denoised:
+            pred_x0, _, *_ = self.model.first_stage_model.quantize(pred_x0)
+
+        if dynamic_threshold is not None:
+            pred_x0 = norm_thresholding(pred_x0, dynamic_threshold)
+
+        # direction pointing to x_t
+        dir_xt = (1. - a_prev - sigma_t**2).sqrt() * e_t
+        noise = sigma_t * noise_like(x.shape, device, repeat_noise) * temperature
+        if noise_dropout > 0.:
+            noise = torch.nn.functional.dropout(noise, p=noise_dropout)
+        x_prev = a_prev.sqrt() * pred_x0 + dir_xt + noise
+        return x_prev, pred_x0
+
+    @torch.no_grad()
+    def encode(self, x0, c, t_enc, use_original_steps=False, return_intermediates=None,
+               unconditional_guidance_scale=1.0, unconditional_conditioning=None):
+        num_reference_steps = self.ddpm_num_timesteps if use_original_steps else self.ddim_timesteps.shape[0]
+
+        assert t_enc <= num_reference_steps
+        num_steps = t_enc
+
+        if use_original_steps:
+            alphas_next = self.alphas_cumprod[:num_steps]
+            alphas = self.alphas_cumprod_prev[:num_steps]
+        else:
+            alphas_next = self.ddim_alphas[:num_steps]
+            alphas = torch.tensor(self.ddim_alphas_prev[:num_steps])
+
+        x_next = x0
+        intermediates = []
+        inter_steps = []
+        for i in tqdm(range(num_steps), desc='Encoding Image'):
+            t = torch.full((x0.shape[0],), i, device=self.model.device, dtype=torch.long)
+            if unconditional_guidance_scale == 1.:
+                noise_pred = self.model.apply_model(x_next, t, c)
+            else:
+                assert unconditional_conditioning is not None
+                e_t_uncond, noise_pred = torch.chunk(
+                    self.model.apply_model(torch.cat((x_next, x_next)), torch.cat((t, t)),
+                                           torch.cat((unconditional_conditioning, c))), 2)
+                noise_pred = e_t_uncond + unconditional_guidance_scale * (noise_pred - e_t_uncond)
+
+            xt_weighted = (alphas_next[i] / alphas[i]).sqrt() * x_next
+            weighted_noise_pred = alphas_next[i].sqrt() * (
+                    (1 / alphas_next[i] - 1).sqrt() - (1 / alphas[i] - 1).sqrt()) * noise_pred
+            x_next = xt_weighted + weighted_noise_pred
+            if return_intermediates and i % (
+                    num_steps // return_intermediates) == 0 and i < num_steps - 1:
+                intermediates.append(x_next)
+                inter_steps.append(i)
+            elif return_intermediates and i >= num_steps - 2:
+                intermediates.append(x_next)
+                inter_steps.append(i)
+
+        out = {'x_encoded': x_next, 'intermediate_steps': inter_steps}
+        if return_intermediates:
+            out.update({'intermediates': intermediates})
+        return x_next, out
+
+    @torch.no_grad()
+    def stochastic_encode(self, x0, t, use_original_steps=False, noise=None):
+        # fast, but does not allow for exact reconstruction
+        # t serves as an index to gather the correct alphas
+        if use_original_steps:
+            sqrt_alphas_cumprod = self.sqrt_alphas_cumprod
+            sqrt_one_minus_alphas_cumprod = self.sqrt_one_minus_alphas_cumprod
+        else:
+            sqrt_alphas_cumprod = torch.sqrt(self.ddim_alphas)
+            sqrt_one_minus_alphas_cumprod = self.ddim_sqrt_one_minus_alphas
+
+        if noise is None:
+            noise = torch.randn_like(x0)
+        return (extract_into_tensor(sqrt_alphas_cumprod, t, x0.shape) * x0 +
+                extract_into_tensor(sqrt_one_minus_alphas_cumprod, t, x0.shape) * noise)
+
+    @torch.no_grad()
+    def decode(self, x_latent, cond, t_start, unconditional_guidance_scale=1.0, unconditional_conditioning=None,
+               use_original_steps=False):
+
+        timesteps = np.arange(self.ddpm_num_timesteps) if use_original_steps else self.ddim_timesteps
+        timesteps = timesteps[:t_start]
+
+        time_range = np.flip(timesteps)
+        total_steps = timesteps.shape[0]
+        print(f"Running DDIM Sampling with {total_steps} timesteps")
+
+        iterator = tqdm(time_range, desc='Decoding image', total=total_steps)
+        x_dec = x_latent
+        for i, step in enumerate(iterator):
+            index = total_steps - i - 1
+            ts = torch.full((x_latent.shape[0],), step, device=x_latent.device, dtype=torch.long)
+            x_dec, _ = self.p_sample_ddim(x_dec, cond, ts, index=index, use_original_steps=use_original_steps,
+                                          unconditional_guidance_scale=unconditional_guidance_scale,
+                                          unconditional_conditioning=unconditional_conditioning)
+        return x_dec

ldm/models/diffusion/ddpm.py

@@ -0,0 +1,1994 @@
+"""
+wild mixture of
+https://github.com/lucidrains/denoising-diffusion-pytorch/blob/7706bdfc6f527f58d33f84b7b522e61e6e3164b3/denoising_diffusion_pytorch/denoising_diffusion_pytorch.py
+https://github.com/openai/improved-diffusion/blob/e94489283bb876ac1477d5dd7709bbbd2d9902ce/improved_diffusion/gaussian_diffusion.py
+https://github.com/CompVis/taming-transformers
+-- merci
+"""
+
+import torch
+import torch.nn as nn
+import numpy as np
+import pytorch_lightning as pl
+from torch.optim.lr_scheduler import LambdaLR
+from einops import rearrange, repeat
+from contextlib import contextmanager, nullcontext
+from functools import partial
+import itertools
+from tqdm import tqdm
+from torchvision.utils import make_grid
+from pytorch_lightning.utilities.distributed import rank_zero_only
+from omegaconf import ListConfig
+
+from ldm.util import log_txt_as_img, exists, default, ismap, isimage, mean_flat, count_params, instantiate_from_config
+from ldm.modules.ema import LitEma
+from ldm.modules.distributions.distributions import normal_kl, DiagonalGaussianDistribution
+from ldm.models.autoencoder import VQModelInterface, IdentityFirstStage, AutoencoderKL
+from ldm.modules.diffusionmodules.util import make_beta_schedule, extract_into_tensor, noise_like
+from ldm.models.diffusion.ddim import DDIMSampler
+from ldm.modules.attention import CrossAttention
+
+
+__conditioning_keys__ = {'concat': 'c_concat',
+                         'crossattn': 'c_crossattn',
+                         'adm': 'y'}
+
+
+def disabled_train(self, mode=True):
+    """Overwrite model.train with this function to make sure train/eval mode
+    does not change anymore."""
+    return self
+
+
+def uniform_on_device(r1, r2, shape, device):
+    return (r1 - r2) * torch.rand(*shape, device=device) + r2
+
+
+class DDPM(pl.LightningModule):
+    # classic DDPM with Gaussian diffusion, in image space
+    def __init__(self,
+                 unet_config,
+                 timesteps=1000,
+                 beta_schedule="linear",
+                 loss_type="l2",
+                 ckpt_path=None,
+                 ignore_keys=[],
+                 load_only_unet=False,
+                 monitor="val/loss",
+                 use_ema=True,
+                 first_stage_key="image",
+                 image_size=256,
+                 channels=3,
+                 log_every_t=100,
+                 clip_denoised=True,
+                 linear_start=1e-4,
+                 linear_end=2e-2,
+                 cosine_s=8e-3,
+                 given_betas=None,
+                 original_elbo_weight=0.,
+                 v_posterior=0.,  # weight for choosing posterior variance as sigma = (1-v) * beta_tilde + v * beta
+                 l_simple_weight=1.,
+                 conditioning_key=None,
+                 parameterization="eps",  # all assuming fixed variance schedules
+                 scheduler_config=None,
+                 use_positional_encodings=False,
+                 learn_logvar=False,
+                 logvar_init=0.,
+                 make_it_fit=False,
+                 ucg_training=None,
+                 ):
+        super().__init__()
+        assert parameterization in ["eps", "x0"], 'currently only supporting "eps" and "x0"'
+        self.parameterization = parameterization
+        print(f"{self.__class__.__name__}: Running in {self.parameterization}-prediction mode")
+        self.cond_stage_model = None
+        self.clip_denoised = clip_denoised
+        self.log_every_t = log_every_t
+        self.first_stage_key = first_stage_key
+        self.image_size = image_size  # try conv?
+        self.channels = channels
+        self.use_positional_encodings = use_positional_encodings
+        self.model = DiffusionWrapper(unet_config, conditioning_key)
+        count_params(self.model, verbose=True)
+        self.use_ema = use_ema
+        if self.use_ema:
+            self.model_ema = LitEma(self.model)
+            print(f"Keeping EMAs of {len(list(self.model_ema.buffers()))}.")
+
+        self.use_scheduler = scheduler_config is not None
+        if self.use_scheduler:
+            self.scheduler_config = scheduler_config
+
+        self.v_posterior = v_posterior
+        self.original_elbo_weight = original_elbo_weight
+        self.l_simple_weight = l_simple_weight
+
+        if monitor is not None:
+            self.monitor = monitor
+        self.make_it_fit = make_it_fit
+        if ckpt_path is not None:
+            self.init_from_ckpt(ckpt_path, ignore_keys=ignore_keys, only_model=load_only_unet)
+
+        self.register_schedule(given_betas=given_betas, beta_schedule=beta_schedule, timesteps=timesteps,
+                               linear_start=linear_start, linear_end=linear_end, cosine_s=cosine_s)
+
+        self.loss_type = loss_type
+
+        self.learn_logvar = learn_logvar
+        self.logvar = torch.full(fill_value=logvar_init, size=(self.num_timesteps,))
+        if self.learn_logvar:
+            self.logvar = nn.Parameter(self.logvar, requires_grad=True)
+
+        self.ucg_training = ucg_training or dict()
+        if self.ucg_training:
+            self.ucg_prng = np.random.RandomState()
+
+    def register_schedule(self, given_betas=None, beta_schedule="linear", timesteps=1000,
+                          linear_start=1e-4, linear_end=2e-2, cosine_s=8e-3):
+        if exists(given_betas):
+            betas = given_betas
+        else:
+            betas = make_beta_schedule(beta_schedule, timesteps, linear_start=linear_start, linear_end=linear_end,
+                                       cosine_s=cosine_s)
+        alphas = 1. - betas
+        alphas_cumprod = np.cumprod(alphas, axis=0)
+        alphas_cumprod_prev = np.append(1., alphas_cumprod[:-1])
+
+        timesteps, = betas.shape
+        self.num_timesteps = int(timesteps)
+        self.linear_start = linear_start
+        self.linear_end = linear_end
+        assert alphas_cumprod.shape[0] == self.num_timesteps, 'alphas have to be defined for each timestep'
+
+        to_torch = partial(torch.tensor, dtype=torch.float32)
+
+        self.register_buffer('betas', to_torch(betas))
+        self.register_buffer('alphas_cumprod', to_torch(alphas_cumprod))
+        self.register_buffer('alphas_cumprod_prev', to_torch(alphas_cumprod_prev))
+
+        # calculations for diffusion q(x_t | x_{t-1}) and others
+        self.register_buffer('sqrt_alphas_cumprod', to_torch(np.sqrt(alphas_cumprod)))
+        self.register_buffer('sqrt_one_minus_alphas_cumprod', to_torch(np.sqrt(1. - alphas_cumprod)))
+        self.register_buffer('log_one_minus_alphas_cumprod', to_torch(np.log(1. - alphas_cumprod)))
+        self.register_buffer('sqrt_recip_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod)))
+        self.register_buffer('sqrt_recipm1_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod - 1)))
+
+        # calculations for posterior q(x_{t-1} | x_t, x_0)
+        posterior_variance = (1 - self.v_posterior) * betas * (1. - alphas_cumprod_prev) / (
+                    1. - alphas_cumprod) + self.v_posterior * betas
+        # above: equal to 1. / (1. / (1. - alpha_cumprod_tm1) + alpha_t / beta_t)
+        self.register_buffer('posterior_variance', to_torch(posterior_variance))
+        # below: log calculation clipped because the posterior variance is 0 at the beginning of the diffusion chain
+        self.register_buffer('posterior_log_variance_clipped', to_torch(np.log(np.maximum(posterior_variance, 1e-20))))
+        self.register_buffer('posterior_mean_coef1', to_torch(
+            betas * np.sqrt(alphas_cumprod_prev) / (1. - alphas_cumprod)))
+        self.register_buffer('posterior_mean_coef2', to_torch(
+            (1. - alphas_cumprod_prev) * np.sqrt(alphas) / (1. - alphas_cumprod)))
+
+        if self.parameterization == "eps":
+            lvlb_weights = self.betas ** 2 / (
+                        2 * self.posterior_variance * to_torch(alphas) * (1 - self.alphas_cumprod))
+        elif self.parameterization == "x0":
+            lvlb_weights = 0.5 * np.sqrt(torch.Tensor(alphas_cumprod)) / (2. * 1 - torch.Tensor(alphas_cumprod))
+        else:
+            raise NotImplementedError("mu not supported")
+        # TODO how to choose this term
+        lvlb_weights[0] = lvlb_weights[1]
+        self.register_buffer('lvlb_weights', lvlb_weights, persistent=False)
+        assert not torch.isnan(self.lvlb_weights).all()
+
+    @contextmanager
+    def ema_scope(self, context=None):
+        if self.use_ema:
+            self.model_ema.store(self.model.parameters())
+            self.model_ema.copy_to(self.model)
+            if context is not None:
+                print(f"{context}: Switched to EMA weights")
+        try:
+            yield None
+        finally:
+            if self.use_ema:
+                self.model_ema.restore(self.model.parameters())
+                if context is not None:
+                    print(f"{context}: Restored training weights")
+
+    @torch.no_grad()
+    def init_from_ckpt(self, path, ignore_keys=list(), only_model=False):
+        sd = torch.load(path, map_location="cpu")
+        if "state_dict" in list(sd.keys()):
+            sd = sd["state_dict"]
+        keys = list(sd.keys())
+
+        if self.make_it_fit:
+            n_params = len([name for name, _ in
+                            itertools.chain(self.named_parameters(),
+                                            self.named_buffers())])
+            for name, param in tqdm(
+                    itertools.chain(self.named_parameters(),
+                                    self.named_buffers()),
+                    desc="Fitting old weights to new weights",
+                    total=n_params
+            ):
+                if not name in sd:
+                    continue
+                old_shape = sd[name].shape
+                new_shape = param.shape
+                assert len(old_shape)==len(new_shape)
+                if len(new_shape) > 2:
+                    # we only modify first two axes
+                    assert new_shape[2:] == old_shape[2:]
+                # assumes first axis corresponds to output dim
+                if not new_shape == old_shape:
+                    new_param = param.clone()
+                    old_param = sd[name]
+                    if len(new_shape) == 1:
+                        for i in range(new_param.shape[0]):
+                            new_param[i] = old_param[i % old_shape[0]]
+                    elif len(new_shape) >= 2:
+                        for i in range(new_param.shape[0]):
+                            for j in range(new_param.shape[1]):
+                                new_param[i, j] = old_param[i % old_shape[0], j % old_shape[1]]
+
+                        n_used_old = torch.ones(old_shape[1])
+                        for j in range(new_param.shape[1]):
+                            n_used_old[j % old_shape[1]] += 1
+                        n_used_new = torch.zeros(new_shape[1])
+                        for j in range(new_param.shape[1]):
+                            n_used_new[j] = n_used_old[j % old_shape[1]]
+
+                        n_used_new = n_used_new[None, :]
+                        while len(n_used_new.shape) < len(new_shape):
+                            n_used_new = n_used_new.unsqueeze(-1)
+                        new_param /= n_used_new
+
+                    sd[name] = new_param
+
+        missing, unexpected = self.load_state_dict(sd, strict=False) if not only_model else self.model.load_state_dict(
+            sd, strict=False)
+        print(f"Restored from {path} with {len(missing)} missing and {len(unexpected)} unexpected keys")
+        if len(missing) > 0:
+            print(f"Missing Keys: {missing}")
+        if len(unexpected) > 0:
+            print(f"Unexpected Keys: {unexpected}")
+
+    def q_mean_variance(self, x_start, t):
+        """
+        Get the distribution q(x_t | x_0).
+        :param x_start: the [N x C x ...] tensor of noiseless inputs.
+        :param t: the number of diffusion steps (minus 1). Here, 0 means one step.
+        :return: A tuple (mean, variance, log_variance), all of x_start's shape.
+        """
+        mean = (extract_into_tensor(self.sqrt_alphas_cumprod, t, x_start.shape) * x_start)
+        variance = extract_into_tensor(1.0 - self.alphas_cumprod, t, x_start.shape)
+        log_variance = extract_into_tensor(self.log_one_minus_alphas_cumprod, t, x_start.shape)
+        return mean, variance, log_variance
+
+    def predict_start_from_noise(self, x_t, t, noise):
+        return (
+                extract_into_tensor(self.sqrt_recip_alphas_cumprod, t, x_t.shape) * x_t -
+                extract_into_tensor(self.sqrt_recipm1_alphas_cumprod, t, x_t.shape) * noise
+        )
+
+    def q_posterior(self, x_start, x_t, t):
+        posterior_mean = (
+                extract_into_tensor(self.posterior_mean_coef1, t, x_t.shape) * x_start +
+                extract_into_tensor(self.posterior_mean_coef2, t, x_t.shape) * x_t
+        )
+        posterior_variance = extract_into_tensor(self.posterior_variance, t, x_t.shape)
+        posterior_log_variance_clipped = extract_into_tensor(self.posterior_log_variance_clipped, t, x_t.shape)
+        return posterior_mean, posterior_variance, posterior_log_variance_clipped
+
+    def p_mean_variance(self, x, t, clip_denoised: bool):
+        model_out = self.model(x, t)
+        if self.parameterization == "eps":
+            x_recon = self.predict_start_from_noise(x, t=t, noise=model_out)
+        elif self.parameterization == "x0":
+            x_recon = model_out
+        if clip_denoised:
+            x_recon.clamp_(-1., 1.)
+
+        model_mean, posterior_variance, posterior_log_variance = self.q_posterior(x_start=x_recon, x_t=x, t=t)
+        return model_mean, posterior_variance, posterior_log_variance
+
+    @torch.no_grad()
+    def p_sample(self, x, t, clip_denoised=True, repeat_noise=False):
+        b, *_, device = *x.shape, x.device
+        model_mean, _, model_log_variance = self.p_mean_variance(x=x, t=t, clip_denoised=clip_denoised)
+        noise = noise_like(x.shape, device, repeat_noise)
+        # no noise when t == 0
+        nonzero_mask = (1 - (t == 0).float()).reshape(b, *((1,) * (len(x.shape) - 1)))
+        return model_mean + nonzero_mask * (0.5 * model_log_variance).exp() * noise
+
+    @torch.no_grad()
+    def p_sample_loop(self, shape, return_intermediates=False):
+        device = self.betas.device
+        b = shape[0]
+        img = torch.randn(shape, device=device)
+        intermediates = [img]
+        for i in tqdm(reversed(range(0, self.num_timesteps)), desc='Sampling t', total=self.num_timesteps):
+            img = self.p_sample(img, torch.full((b,), i, device=device, dtype=torch.long),
+                                clip_denoised=self.clip_denoised)
+            if i % self.log_every_t == 0 or i == self.num_timesteps - 1:
+                intermediates.append(img)
+        if return_intermediates:
+            return img, intermediates
+        return img
+
+    @torch.no_grad()
+    def sample(self, batch_size=16, return_intermediates=False):
+        image_size = self.image_size
+        channels = self.channels
+        return self.p_sample_loop((batch_size, channels, image_size, image_size),
+                                  return_intermediates=return_intermediates)
+
+    def q_sample(self, x_start, t, noise=None):
+        noise = default(noise, lambda: torch.randn_like(x_start))
+        return (extract_into_tensor(self.sqrt_alphas_cumprod, t, x_start.shape) * x_start +
+                extract_into_tensor(self.sqrt_one_minus_alphas_cumprod, t, x_start.shape) * noise)
+
+    def get_loss(self, pred, target, mean=True):
+        if self.loss_type == 'l1':
+            loss = (target - pred).abs()
+            if mean:
+                loss = loss.mean()
+        elif self.loss_type == 'l2':
+            if mean:
+                loss = torch.nn.functional.mse_loss(target, pred)
+            else:
+                loss = torch.nn.functional.mse_loss(target, pred, reduction='none')
+        else:
+            raise NotImplementedError("unknown loss type '{loss_type}'")
+
+        return loss
+
+    def p_losses(self, x_start, t, noise=None):
+        noise = default(noise, lambda: torch.randn_like(x_start))
+        x_noisy = self.q_sample(x_start=x_start, t=t, noise=noise)
+        model_out = self.model(x_noisy, t)
+
+        loss_dict = {}
+        if self.parameterization == "eps":
+            target = noise
+        elif self.parameterization == "x0":
+            target = x_start
+        else:
+            raise NotImplementedError(f"Paramterization {self.parameterization} not yet supported")
+
+        loss = self.get_loss(model_out, target, mean=False).mean(dim=[1, 2, 3])
+
+        log_prefix = 'train' if self.training else 'val'
+
+        loss_dict.update({f'{log_prefix}/loss_simple': loss.mean()})
+        loss_simple = loss.mean() * self.l_simple_weight
+
+        loss_vlb = (self.lvlb_weights[t] * loss).mean()
+        loss_dict.update({f'{log_prefix}/loss_vlb': loss_vlb})
+
+        loss = loss_simple + self.original_elbo_weight * loss_vlb
+
+        loss_dict.update({f'{log_prefix}/loss': loss})
+
+        return loss, loss_dict
+
+    def forward(self, x, *args, **kwargs):
+        # b, c, h, w, device, img_size, = *x.shape, x.device, self.image_size
+        # assert h == img_size and w == img_size, f'height and width of image must be {img_size}'
+        t = torch.randint(0, self.num_timesteps, (x.shape[0],), device=self.device).long()
+        return self.p_losses(x, t, *args, **kwargs)
+
+    def get_input(self, batch, k):
+        x = batch[k]
+        if len(x.shape) == 3:
+            x = x[..., None]
+        x = rearrange(x, 'b h w c -> b c h w')
+        x = x.to(memory_format=torch.contiguous_format).float()
+        return x
+
+    def shared_step(self, batch):
+        x = self.get_input(batch, self.first_stage_key)
+        loss, loss_dict = self(x)
+        return loss, loss_dict
+
+    def training_step(self, batch, batch_idx):
+        for k in self.ucg_training:
+            p = self.ucg_training[k]["p"]
+            val = self.ucg_training[k]["val"]
+            if val is None:
+                val = ""
+            for i in range(len(batch[k])):
+                if self.ucg_prng.choice(2, p=[1-p, p]):
+                    batch[k][i] = val
+
+        loss, loss_dict = self.shared_step(batch)
+
+        self.log_dict(loss_dict, prog_bar=True,
+                      logger=True, on_step=True, on_epoch=True)
+
+        self.log("global_step", self.global_step,
+                 prog_bar=True, logger=True, on_step=True, on_epoch=False)
+
+        if self.use_scheduler:
+            lr = self.optimizers().param_groups[0]['lr']
+            self.log('lr_abs', lr, prog_bar=True, logger=True, on_step=True, on_epoch=False)
+
+        return loss
+
+    @torch.no_grad()
+    def validation_step(self, batch, batch_idx):
+        _, loss_dict_no_ema = self.shared_step(batch)
+        with self.ema_scope():
+            _, loss_dict_ema = self.shared_step(batch)
+            loss_dict_ema = {key + '_ema': loss_dict_ema[key] for key in loss_dict_ema}
+        self.log_dict(loss_dict_no_ema, prog_bar=False, logger=True, on_step=False, on_epoch=True)
+        self.log_dict(loss_dict_ema, prog_bar=False, logger=True, on_step=False, on_epoch=True)
+
+    def on_train_batch_end(self, *args, **kwargs):
+        if self.use_ema:
+            self.model_ema(self.model)
+
+    def _get_rows_from_list(self, samples):
+        n_imgs_per_row = len(samples)
+        denoise_grid = rearrange(samples, 'n b c h w -> b n c h w')
+        denoise_grid = rearrange(denoise_grid, 'b n c h w -> (b n) c h w')
+        denoise_grid = make_grid(denoise_grid, nrow=n_imgs_per_row)
+        return denoise_grid
+
+    @torch.no_grad()
+    def log_images(self, batch, N=8, n_row=2, sample=True, return_keys=None, **kwargs):
+        log = dict()
+        x = self.get_input(batch, self.first_stage_key)
+        N = min(x.shape[0], N)
+        n_row = min(x.shape[0], n_row)
+        x = x.to(self.device)[:N]
+        log["inputs"] = x
+
+        # get diffusion row
+        diffusion_row = list()
+        x_start = x[:n_row]
+
+        for t in range(self.num_timesteps):
+            if t % self.log_every_t == 0 or t == self.num_timesteps - 1:
+                t = repeat(torch.tensor([t]), '1 -> b', b=n_row)
+                t = t.to(self.device).long()
+                noise = torch.randn_like(x_start)
+                x_noisy = self.q_sample(x_start=x_start, t=t, noise=noise)
+                diffusion_row.append(x_noisy)
+
+        log["diffusion_row"] = self._get_rows_from_list(diffusion_row)
+
+        if sample:
+            # get denoise row
+            with self.ema_scope("Plotting"):
+                samples, denoise_row = self.sample(batch_size=N, return_intermediates=True)
+
+            log["samples"] = samples
+            log["denoise_row"] = self._get_rows_from_list(denoise_row)
+
+        if return_keys:
+            if np.intersect1d(list(log.keys()), return_keys).shape[0] == 0:
+                return log
+            else:
+                return {key: log[key] for key in return_keys}
+        return log
+
+    def configure_optimizers(self):
+        lr = self.learning_rate
+        params = list(self.model.parameters())
+        if self.learn_logvar:
+            params = params + [self.logvar]
+        opt = torch.optim.AdamW(params, lr=lr)
+        return opt
+
+
+class LatentDiffusion(DDPM):
+    """main class"""
+    def __init__(self,
+                 first_stage_config,
+                 cond_stage_config,
+                 num_timesteps_cond=None,
+                 cond_stage_key="image",
+                 cond_stage_trainable=False,
+                 concat_mode=True,
+                 cond_stage_forward=None,
+                 conditioning_key=None,
+                 scale_factor=1.0,
+                 scale_by_std=False,
+                 unet_trainable=True,
+                 *args, **kwargs):
+        self.num_timesteps_cond = default(num_timesteps_cond, 1)
+        self.scale_by_std = scale_by_std
+        assert self.num_timesteps_cond <= kwargs['timesteps']
+        # for backwards compatibility after implementation of DiffusionWrapper
+        if conditioning_key is None:
+            conditioning_key = 'concat' if concat_mode else 'crossattn'
+        if cond_stage_config == '__is_unconditional__':
+            conditioning_key = None
+        ckpt_path = kwargs.pop("ckpt_path", None)
+        ignore_keys = kwargs.pop("ignore_keys", [])
+        super().__init__(conditioning_key=conditioning_key, *args, **kwargs)
+        self.concat_mode = concat_mode
+        self.cond_stage_trainable = cond_stage_trainable
+        self.unet_trainable = unet_trainable
+        self.cond_stage_key = cond_stage_key
+        try:
+            self.num_downs = len(first_stage_config.params.ddconfig.ch_mult) - 1
+        except:
+            self.num_downs = 0
+        if not scale_by_std:
+            self.scale_factor = scale_factor
+        else:
+            self.register_buffer('scale_factor', torch.tensor(scale_factor))
+        self.instantiate_first_stage(first_stage_config)
+        self.instantiate_cond_stage(cond_stage_config)
+        self.cond_stage_forward = cond_stage_forward
+
+        # construct linear projection layer for concatenating image CLIP embedding and RT
+        self.cc_projection = nn.Linear(772, 768)
+        nn.init.eye_(list(self.cc_projection.parameters())[0][:768, :768])
+        nn.init.zeros_(list(self.cc_projection.parameters())[1])
+        self.cc_projection.requires_grad_(True)
+        
+        self.clip_denoised = False
+        self.bbox_tokenizer = None
+
+        self.restarted_from_ckpt = False
+        if ckpt_path is not None:
+            self.init_from_ckpt(ckpt_path, ignore_keys)
+            self.restarted_from_ckpt = True
+
+    def make_cond_schedule(self, ):
+        self.cond_ids = torch.full(size=(self.num_timesteps,), fill_value=self.num_timesteps - 1, dtype=torch.long)
+        ids = torch.round(torch.linspace(0, self.num_timesteps - 1, self.num_timesteps_cond)).long()
+        self.cond_ids[:self.num_timesteps_cond] = ids
+
+    @rank_zero_only
+    @torch.no_grad()
+    def on_train_batch_start(self, batch, batch_idx, dataloader_idx):
+        # only for very first batch
+        if self.scale_by_std and self.current_epoch == 0 and self.global_step == 0 and batch_idx == 0 and not self.restarted_from_ckpt:
+            assert self.scale_factor == 1., 'rather not use custom rescaling and std-rescaling simultaneously'
+            # set rescale weight to 1./std of encodings
+            print("### USING STD-RESCALING ###")
+            x = super().get_input(batch, self.first_stage_key)
+            x = x.to(self.device)
+            encoder_posterior = self.encode_first_stage(x)
+            z = self.get_first_stage_encoding(encoder_posterior).detach()
+            del self.scale_factor
+            self.register_buffer('scale_factor', 1. / z.flatten().std())
+            print(f"setting self.scale_factor to {self.scale_factor}")
+            print("### USING STD-RESCALING ###")
+
+    def register_schedule(self,
+                          given_betas=None, beta_schedule="linear", timesteps=1000,
+                          linear_start=1e-4, linear_end=2e-2, cosine_s=8e-3):
+        super().register_schedule(given_betas, beta_schedule, timesteps, linear_start, linear_end, cosine_s)
+
+        self.shorten_cond_schedule = self.num_timesteps_cond > 1
+        if self.shorten_cond_schedule:
+            self.make_cond_schedule()
+
+    def instantiate_first_stage(self, config):
+        model = instantiate_from_config(config)
+        self.first_stage_model = model.eval()
+        self.first_stage_model.train = disabled_train
+        for param in self.first_stage_model.parameters():
+            param.requires_grad = False
+
+    def instantiate_cond_stage(self, config):
+        if not self.cond_stage_trainable:
+            if config == "__is_first_stage__":
+                print("Using first stage also as cond stage.")
+                self.cond_stage_model = self.first_stage_model
+            elif config == "__is_unconditional__":
+                print(f"Training {self.__class__.__name__} as an unconditional model.")
+                self.cond_stage_model = None
+                # self.be_unconditional = True
+            else:
+                model = instantiate_from_config(config)
+                self.cond_stage_model = model.eval()
+                self.cond_stage_model.train = disabled_train
+                for param in self.cond_stage_model.parameters():
+                    param.requires_grad = False
+        else:
+            assert config != '__is_first_stage__'
+            assert config != '__is_unconditional__'
+            model = instantiate_from_config(config)
+            self.cond_stage_model = model
+
+    def _get_denoise_row_from_list(self, samples, desc='', force_no_decoder_quantization=False):
+        denoise_row = []
+        for zd in tqdm(samples, desc=desc):
+            denoise_row.append(self.decode_first_stage(zd.to(self.device),
+                                                            force_not_quantize=force_no_decoder_quantization))
+        n_imgs_per_row = len(denoise_row)
+        denoise_row = torch.stack(denoise_row)  # n_log_step, n_row, C, H, W
+        denoise_grid = rearrange(denoise_row, 'n b c h w -> b n c h w')
+        denoise_grid = rearrange(denoise_grid, 'b n c h w -> (b n) c h w')
+        denoise_grid = make_grid(denoise_grid, nrow=n_imgs_per_row)
+        return denoise_grid
+
+    def get_first_stage_encoding(self, encoder_posterior):
+        if isinstance(encoder_posterior, DiagonalGaussianDistribution):
+            z = encoder_posterior.sample()
+        elif isinstance(encoder_posterior, torch.Tensor):
+            z = encoder_posterior
+        else:
+            raise NotImplementedError(f"encoder_posterior of type '{type(encoder_posterior)}' not yet implemented")
+        return self.scale_factor * z
+
+    def get_learned_conditioning(self, c):
+        if self.cond_stage_forward is None:
+            if hasattr(self.cond_stage_model, 'encode') and callable(self.cond_stage_model.encode):
+                c = self.cond_stage_model.encode(c)
+                if isinstance(c, DiagonalGaussianDistribution):
+                    c = c.mode()
+            else:
+                c = self.cond_stage_model(c)
+        else:
+            assert hasattr(self.cond_stage_model, self.cond_stage_forward)
+            c = getattr(self.cond_stage_model, self.cond_stage_forward)(c)
+        return c
+
+    def meshgrid(self, h, w):
+        y = torch.arange(0, h).view(h, 1, 1).repeat(1, w, 1)
+        x = torch.arange(0, w).view(1, w, 1).repeat(h, 1, 1)
+
+        arr = torch.cat([y, x], dim=-1)
+        return arr
+
+    def delta_border(self, h, w):
+        """
+        :param h: height
+        :param w: width
+        :return: normalized distance to image border,
+         wtith min distance = 0 at border and max dist = 0.5 at image center
+        """
+        lower_right_corner = torch.tensor([h - 1, w - 1]).view(1, 1, 2)
+        arr = self.meshgrid(h, w) / lower_right_corner
+        dist_left_up = torch.min(arr, dim=-1, keepdims=True)[0]
+        dist_right_down = torch.min(1 - arr, dim=-1, keepdims=True)[0]
+        edge_dist = torch.min(torch.cat([dist_left_up, dist_right_down], dim=-1), dim=-1)[0]
+        return edge_dist
+
+    def get_weighting(self, h, w, Ly, Lx, device):
+        weighting = self.delta_border(h, w)
+        weighting = torch.clip(weighting, self.split_input_params["clip_min_weight"],
+                               self.split_input_params["clip_max_weight"], )
+        weighting = weighting.view(1, h * w, 1).repeat(1, 1, Ly * Lx).to(device)
+
+        if self.split_input_params["tie_braker"]:
+            L_weighting = self.delta_border(Ly, Lx)
+            L_weighting = torch.clip(L_weighting,
+                                     self.split_input_params["clip_min_tie_weight"],
+                                     self.split_input_params["clip_max_tie_weight"])
+
+            L_weighting = L_weighting.view(1, 1, Ly * Lx).to(device)
+            weighting = weighting * L_weighting
+        return weighting
+
+    def get_fold_unfold(self, x, kernel_size, stride, uf=1, df=1):  # todo load once not every time, shorten code
+        """
+        :param x: img of size (bs, c, h, w)
+        :return: n img crops of size (n, bs, c, kernel_size[0], kernel_size[1])
+        """
+        bs, nc, h, w = x.shape
+
+        # number of crops in image
+        Ly = (h - kernel_size[0]) // stride[0] + 1
+        Lx = (w - kernel_size[1]) // stride[1] + 1
+
+        if uf == 1 and df == 1:
+            fold_params = dict(kernel_size=kernel_size, dilation=1, padding=0, stride=stride)
+            unfold = torch.nn.Unfold(**fold_params)
+
+            fold = torch.nn.Fold(output_size=x.shape[2:], **fold_params)
+
+            weighting = self.get_weighting(kernel_size[0], kernel_size[1], Ly, Lx, x.device).to(x.dtype)
+            normalization = fold(weighting).view(1, 1, h, w)  # normalizes the overlap
+            weighting = weighting.view((1, 1, kernel_size[0], kernel_size[1], Ly * Lx))
+
+        elif uf > 1 and df == 1:
+            fold_params = dict(kernel_size=kernel_size, dilation=1, padding=0, stride=stride)
+            unfold = torch.nn.Unfold(**fold_params)
+
+            fold_params2 = dict(kernel_size=(kernel_size[0] * uf, kernel_size[0] * uf),
+                                dilation=1, padding=0,
+                                stride=(stride[0] * uf, stride[1] * uf))
+            fold = torch.nn.Fold(output_size=(x.shape[2] * uf, x.shape[3] * uf), **fold_params2)
+
+            weighting = self.get_weighting(kernel_size[0] * uf, kernel_size[1] * uf, Ly, Lx, x.device).to(x.dtype)
+            normalization = fold(weighting).view(1, 1, h * uf, w * uf)  # normalizes the overlap
+            weighting = weighting.view((1, 1, kernel_size[0] * uf, kernel_size[1] * uf, Ly * Lx))
+
+        elif df > 1 and uf == 1:
+            fold_params = dict(kernel_size=kernel_size, dilation=1, padding=0, stride=stride)
+            unfold = torch.nn.Unfold(**fold_params)
+
+            fold_params2 = dict(kernel_size=(kernel_size[0] // df, kernel_size[0] // df),
+                                dilation=1, padding=0,
+                                stride=(stride[0] // df, stride[1] // df))
+            fold = torch.nn.Fold(output_size=(x.shape[2] // df, x.shape[3] // df), **fold_params2)
+
+            weighting = self.get_weighting(kernel_size[0] // df, kernel_size[1] // df, Ly, Lx, x.device).to(x.dtype)
+            normalization = fold(weighting).view(1, 1, h // df, w // df)  # normalizes the overlap
+            weighting = weighting.view((1, 1, kernel_size[0] // df, kernel_size[1] // df, Ly * Lx))
+
+        else:
+            raise NotImplementedError
+
+        return fold, unfold, normalization, weighting
+
+    
+    @torch.no_grad()
+    def get_input(self, batch, k, return_first_stage_outputs=False, force_c_encode=False,
+                  cond_key=None, return_original_cond=False, bs=None, uncond=0.05):
+        x = super().get_input(batch, k)
+        T = batch['T'].to(memory_format=torch.contiguous_format).float()
+        
+        if bs is not None:
+            x = x[:bs]
+            T = T[:bs].to(self.device)
+
+        x = x.to(self.device)
+        encoder_posterior = self.encode_first_stage(x)
+        z = self.get_first_stage_encoding(encoder_posterior).detach()
+        cond_key = cond_key or self.cond_stage_key
+        xc = super().get_input(batch, cond_key).to(self.device)
+        if bs is not None:
+            xc = xc[:bs]
+        cond = {}
+
+        # To support classifier-free guidance, randomly drop out only text conditioning 5%, only image conditioning 5%, and both 5%.
+        random = torch.rand(x.size(0), device=x.device)
+        prompt_mask = rearrange(random < 2 * uncond, "n -> n 1 1")
+        input_mask = 1 - rearrange((random >= uncond).float() * (random < 3 * uncond).float(), "n -> n 1 1 1")
+        null_prompt = self.get_learned_conditioning([""])
+
+        # z.shape: [8, 4, 64, 64]; c.shape: [8, 1, 768]
+        # print('=========== xc shape ===========', xc.shape)
+        with torch.enable_grad():
+            clip_emb = self.get_learned_conditioning(xc).detach()
+            null_prompt = self.get_learned_conditioning([""]).detach()
+            cond["c_crossattn"] = [self.cc_projection(torch.cat([torch.where(prompt_mask, null_prompt, clip_emb), T[:, None, :]], dim=-1))]
+        cond["c_concat"] = [input_mask * self.encode_first_stage((xc.to(self.device))).mode().detach()]
+        out = [z, cond]
+        if return_first_stage_outputs:
+            xrec = self.decode_first_stage(z)
+            out.extend([x, xrec])
+        if return_original_cond:
+            out.append(xc)
+        return out
+
+    # @torch.no_grad()
+    def decode_first_stage(self, z, predict_cids=False, force_not_quantize=False):
+        if predict_cids:
+            if z.dim() == 4:
+                z = torch.argmax(z.exp(), dim=1).long()
+            z = self.first_stage_model.quantize.get_codebook_entry(z, shape=None)
+            z = rearrange(z, 'b h w c -> b c h w').contiguous()
+
+        z = 1. / self.scale_factor * z
+
+        if hasattr(self, "split_input_params"):
+            if self.split_input_params["patch_distributed_vq"]:
+                ks = self.split_input_params["ks"]  # eg. (128, 128)
+                stride = self.split_input_params["stride"]  # eg. (64, 64)
+                uf = self.split_input_params["vqf"]
+                bs, nc, h, w = z.shape
+                if ks[0] > h or ks[1] > w:
+                    ks = (min(ks[0], h), min(ks[1], w))
+                    print("reducing Kernel")
+
+                if stride[0] > h or stride[1] > w:
+                    stride = (min(stride[0], h), min(stride[1], w))
+                    print("reducing stride")
+
+                fold, unfold, normalization, weighting = self.get_fold_unfold(z, ks, stride, uf=uf)
+
+                z = unfold(z)  # (bn, nc * prod(**ks), L)
+                # 1. Reshape to img shape
+                z = z.view((z.shape[0], -1, ks[0], ks[1], z.shape[-1]))  # (bn, nc, ks[0], ks[1], L )
+
+                # 2. apply model loop over last dim
+                if isinstance(self.first_stage_model, VQModelInterface):
+                    output_list = [self.first_stage_model.decode(z[:, :, :, :, i],
+                                                                 force_not_quantize=predict_cids or force_not_quantize)
+                                   for i in range(z.shape[-1])]
+                else:
+
+                    output_list = [self.first_stage_model.decode(z[:, :, :, :, i])
+                                   for i in range(z.shape[-1])]
+
+                o = torch.stack(output_list, axis=-1)  # # (bn, nc, ks[0], ks[1], L)
+                o = o * weighting
+                # Reverse 1. reshape to img shape
+                o = o.view((o.shape[0], -1, o.shape[-1]))  # (bn, nc * ks[0] * ks[1], L)
+                # stitch crops together
+                decoded = fold(o)
+                decoded = decoded / normalization  # norm is shape (1, 1, h, w)
+                return decoded
+            else:
+                if isinstance(self.first_stage_model, VQModelInterface):
+                    return self.first_stage_model.decode(z, force_not_quantize=predict_cids or force_not_quantize)
+                else:
+                    return self.first_stage_model.decode(z)
+
+        else:
+            if isinstance(self.first_stage_model, VQModelInterface):
+                return self.first_stage_model.decode(z, force_not_quantize=predict_cids or force_not_quantize)
+            else:
+                return self.first_stage_model.decode(z)
+
+    @torch.no_grad()
+    def encode_first_stage(self, x):
+        if hasattr(self, "split_input_params"):
+            if self.split_input_params["patch_distributed_vq"]:
+                ks = self.split_input_params["ks"]  # eg. (128, 128)
+                stride = self.split_input_params["stride"]  # eg. (64, 64)
+                df = self.split_input_params["vqf"]
+                self.split_input_params['original_image_size'] = x.shape[-2:]
+                bs, nc, h, w = x.shape
+                if ks[0] > h or ks[1] > w:
+                    ks = (min(ks[0], h), min(ks[1], w))
+                    print("reducing Kernel")
+
+                if stride[0] > h or stride[1] > w:
+                    stride = (min(stride[0], h), min(stride[1], w))
+                    print("reducing stride")
+
+                fold, unfold, normalization, weighting = self.get_fold_unfold(x, ks, stride, df=df)
+                z = unfold(x)  # (bn, nc * prod(**ks), L)
+                # Reshape to img shape
+                z = z.view((z.shape[0], -1, ks[0], ks[1], z.shape[-1]))  # (bn, nc, ks[0], ks[1], L )
+
+                output_list = [self.first_stage_model.encode(z[:, :, :, :, i])
+                               for i in range(z.shape[-1])]
+
+                o = torch.stack(output_list, axis=-1)
+                o = o * weighting
+
+                # Reverse reshape to img shape
+                o = o.view((o.shape[0], -1, o.shape[-1]))  # (bn, nc * ks[0] * ks[1], L)
+                # stitch crops together
+                decoded = fold(o)
+                decoded = decoded / normalization
+                return decoded
+
+            else:
+                return self.first_stage_model.encode(x)
+        else:
+            return self.first_stage_model.encode(x)
+
+    def shared_step(self, batch, **kwargs):
+        x, c = self.get_input(batch, self.first_stage_key)
+        loss = self(x, c)
+        return loss
+
+    def forward(self, x, c, *args, **kwargs):
+        t = torch.randint(0, self.num_timesteps, (x.shape[0],), device=self.device).long()
+        if self.model.conditioning_key is not None:
+            assert c is not None
+            # if self.cond_stage_trainable:
+            #     c = self.get_learned_conditioning(c)
+            if self.shorten_cond_schedule:  # TODO: drop this option
+                tc = self.cond_ids[t].to(self.device)
+                c = self.q_sample(x_start=c, t=tc, noise=torch.randn_like(c.float()))
+        return self.p_losses(x, c, t, *args, **kwargs)
+
+    def _rescale_annotations(self, bboxes, crop_coordinates):  # TODO: move to dataset
+        def rescale_bbox(bbox):
+            x0 = clamp((bbox[0] - crop_coordinates[0]) / crop_coordinates[2])
+            y0 = clamp((bbox[1] - crop_coordinates[1]) / crop_coordinates[3])
+            w = min(bbox[2] / crop_coordinates[2], 1 - x0)
+            h = min(bbox[3] / crop_coordinates[3], 1 - y0)
+            return x0, y0, w, h
+
+        return [rescale_bbox(b) for b in bboxes]
+
+    def apply_model(self, x_noisy, t, cond, return_ids=False):
+
+        if isinstance(cond, dict):
+            # hybrid case, cond is exptected to be a dict
+            pass
+        else:
+            if not isinstance(cond, list):
+                cond = [cond]
+            key = 'c_concat' if self.model.conditioning_key == 'concat' else 'c_crossattn'
+            cond = {key: cond}
+
+        if hasattr(self, "split_input_params"):
+            assert len(cond) == 1  # todo can only deal with one conditioning atm
+            assert not return_ids
+            ks = self.split_input_params["ks"]  # eg. (128, 128)
+            stride = self.split_input_params["stride"]  # eg. (64, 64)
+
+            h, w = x_noisy.shape[-2:]
+
+            fold, unfold, normalization, weighting = self.get_fold_unfold(x_noisy, ks, stride)
+
+            z = unfold(x_noisy)  # (bn, nc * prod(**ks), L)
+            # Reshape to img shape
+            z = z.view((z.shape[0], -1, ks[0], ks[1], z.shape[-1]))  # (bn, nc, ks[0], ks[1], L )
+            z_list = [z[:, :, :, :, i] for i in range(z.shape[-1])]
+
+            if self.cond_stage_key in ["image", "LR_image", "segmentation",
+                                       'bbox_img'] and self.model.conditioning_key:  # todo check for completeness
+                c_key = next(iter(cond.keys()))  # get key
+                c = next(iter(cond.values()))  # get value
+                assert (len(c) == 1)  # todo extend to list with more than one elem
+                c = c[0]  # get element
+
+                c = unfold(c)
+                c = c.view((c.shape[0], -1, ks[0], ks[1], c.shape[-1]))  # (bn, nc, ks[0], ks[1], L )
+
+                cond_list = [{c_key: [c[:, :, :, :, i]]} for i in range(c.shape[-1])]
+
+            elif self.cond_stage_key == 'coordinates_bbox':
+                assert 'original_image_size' in self.split_input_params, 'BoudingBoxRescaling is missing original_image_size'
+
+                # assuming padding of unfold is always 0 and its dilation is always 1
+                n_patches_per_row = int((w - ks[0]) / stride[0] + 1)
+                full_img_h, full_img_w = self.split_input_params['original_image_size']
+                # as we are operating on latents, we need the factor from the original image size to the
+                # spatial latent size to properly rescale the crops for regenerating the bbox annotations
+                num_downs = self.first_stage_model.encoder.num_resolutions - 1
+                rescale_latent = 2 ** (num_downs)
+
+                # get top left postions of patches as conforming for the bbbox tokenizer, therefore we
+                # need to rescale the tl patch coordinates to be in between (0,1)
+                tl_patch_coordinates = [(rescale_latent * stride[0] * (patch_nr % n_patches_per_row) / full_img_w,
+                                         rescale_latent * stride[1] * (patch_nr // n_patches_per_row) / full_img_h)
+                                        for patch_nr in range(z.shape[-1])]
+
+                # patch_limits are tl_coord, width and height coordinates as (x_tl, y_tl, h, w)
+                patch_limits = [(x_tl, y_tl,
+                                 rescale_latent * ks[0] / full_img_w,
+                                 rescale_latent * ks[1] / full_img_h) for x_tl, y_tl in tl_patch_coordinates]
+                # patch_values = [(np.arange(x_tl,min(x_tl+ks, 1.)),np.arange(y_tl,min(y_tl+ks, 1.))) for x_tl, y_tl in tl_patch_coordinates]
+
+                # tokenize crop coordinates for the bounding boxes of the respective patches
+                patch_limits_tknzd = [torch.LongTensor(self.bbox_tokenizer._crop_encoder(bbox))[None].to(self.device)
+                                      for bbox in patch_limits]  # list of length l with tensors of shape (1, 2)
+                # cut tknzd crop position from conditioning
+                assert isinstance(cond, dict), 'cond must be dict to be fed into model'
+                cut_cond = cond['c_crossattn'][0][..., :-2].to(self.device)
+
+                adapted_cond = torch.stack([torch.cat([cut_cond, p], dim=1) for p in patch_limits_tknzd])
+                adapted_cond = rearrange(adapted_cond, 'l b n -> (l b) n')
+                adapted_cond = self.get_learned_conditioning(adapted_cond)
+                adapted_cond = rearrange(adapted_cond, '(l b) n d -> l b n d', l=z.shape[-1])
+
+                cond_list = [{'c_crossattn': [e]} for e in adapted_cond]
+
+            else:
+                cond_list = [cond for i in range(z.shape[-1])]  # Todo make this more efficient
+
+            # apply model by loop over crops
+            output_list = [self.model(z_list[i], t, **cond_list[i]) for i in range(z.shape[-1])]
+            assert not isinstance(output_list[0],
+                                  tuple)  # todo cant deal with multiple model outputs check this never happens
+
+            o = torch.stack(output_list, axis=-1)
+            o = o * weighting
+            # Reverse reshape to img shape
+            o = o.view((o.shape[0], -1, o.shape[-1]))  # (bn, nc * ks[0] * ks[1], L)
+            # stitch crops together
+            x_recon = fold(o) / normalization
+
+        else:
+            x_recon = self.model(x_noisy, t, **cond)
+
+        if isinstance(x_recon, tuple) and not return_ids:
+            return x_recon[0]
+        else:
+            return x_recon
+
+    def _predict_eps_from_xstart(self, x_t, t, pred_xstart):
+        return (extract_into_tensor(self.sqrt_recip_alphas_cumprod, t, x_t.shape) * x_t - pred_xstart) / \
+               extract_into_tensor(self.sqrt_recipm1_alphas_cumprod, t, x_t.shape)
+
+    def _prior_bpd(self, x_start):
+        """
+        Get the prior KL term for the variational lower-bound, measured in
+        bits-per-dim.
+        This term can't be optimized, as it only depends on the encoder.
+        :param x_start: the [N x C x ...] tensor of inputs.
+        :return: a batch of [N] KL values (in bits), one per batch element.
+        """
+        batch_size = x_start.shape[0]
+        t = torch.tensor([self.num_timesteps - 1] * batch_size, device=x_start.device)
+        qt_mean, _, qt_log_variance = self.q_mean_variance(x_start, t)
+        kl_prior = normal_kl(mean1=qt_mean, logvar1=qt_log_variance, mean2=0.0, logvar2=0.0)
+        return mean_flat(kl_prior) / np.log(2.0)
+
+    def p_losses(self, x_start, cond, t, noise=None):
+        noise = default(noise, lambda: torch.randn_like(x_start))
+        x_noisy = self.q_sample(x_start=x_start, t=t, noise=noise)
+        model_output = self.apply_model(x_noisy, t, cond)
+
+        loss_dict = {}
+        prefix = 'train' if self.training else 'val'
+
+        if self.parameterization == "x0":
+            target = x_start
+        elif self.parameterization == "eps":
+            target = noise
+        else:
+            raise NotImplementedError()
+
+        loss_simple = self.get_loss(model_output, target, mean=False).mean([1, 2, 3])
+        loss_dict.update({f'{prefix}/loss_simple': loss_simple.mean()})
+
+        logvar_t = self.logvar[t].to(self.device)
+        loss = loss_simple / torch.exp(logvar_t) + logvar_t
+        # loss = loss_simple / torch.exp(self.logvar) + self.logvar
+        if self.learn_logvar:
+            loss_dict.update({f'{prefix}/loss_gamma': loss.mean()})
+            loss_dict.update({'logvar': self.logvar.data.mean()})
+
+        loss = self.l_simple_weight * loss.mean()
+
+        loss_vlb = self.get_loss(model_output, target, mean=False).mean(dim=(1, 2, 3))
+        loss_vlb = (self.lvlb_weights[t] * loss_vlb).mean()
+        loss_dict.update({f'{prefix}/loss_vlb': loss_vlb})
+        loss += (self.original_elbo_weight * loss_vlb)
+        loss_dict.update({f'{prefix}/loss': loss})
+
+        return loss, loss_dict
+
+    def p_mean_variance(self, x, c, t, clip_denoised: bool, return_codebook_ids=False, quantize_denoised=False,
+                        return_x0=False, score_corrector=None, corrector_kwargs=None):
+        t_in = t
+        model_out = self.apply_model(x, t_in, c, return_ids=return_codebook_ids)
+
+        if score_corrector is not None:
+            assert self.parameterization == "eps"
+            model_out = score_corrector.modify_score(self, model_out, x, t, c, **corrector_kwargs)
+
+        if return_codebook_ids:
+            model_out, logits = model_out
+
+        if self.parameterization == "eps":
+            x_recon = self.predict_start_from_noise(x, t=t, noise=model_out)
+        elif self.parameterization == "x0":
+            x_recon = model_out
+        else:
+            raise NotImplementedError()
+
+        if clip_denoised:
+            x_recon.clamp_(-1., 1.)
+        if quantize_denoised:
+            x_recon, _, [_, _, indices] = self.first_stage_model.quantize(x_recon)
+        model_mean, posterior_variance, posterior_log_variance = self.q_posterior(x_start=x_recon, x_t=x, t=t)
+        if return_codebook_ids:
+            return model_mean, posterior_variance, posterior_log_variance, logits
+        elif return_x0:
+            return model_mean, posterior_variance, posterior_log_variance, x_recon
+        else:
+            return model_mean, posterior_variance, posterior_log_variance
+
+    @torch.no_grad()
+    def p_sample(self, x, c, t, clip_denoised=False, repeat_noise=False,
+                 return_codebook_ids=False, quantize_denoised=False, return_x0=False,
+                 temperature=1., noise_dropout=0., score_corrector=None, corrector_kwargs=None):
+        b, *_, device = *x.shape, x.device
+        outputs = self.p_mean_variance(x=x, c=c, t=t, clip_denoised=clip_denoised,
+                                       return_codebook_ids=return_codebook_ids,
+                                       quantize_denoised=quantize_denoised,
+                                       return_x0=return_x0,
+                                       score_corrector=score_corrector, corrector_kwargs=corrector_kwargs)
+        if return_codebook_ids:
+            raise DeprecationWarning("Support dropped.")
+            model_mean, _, model_log_variance, logits = outputs
+        elif return_x0:
+            model_mean, _, model_log_variance, x0 = outputs
+        else:
+            model_mean, _, model_log_variance = outputs
+
+        noise = noise_like(x.shape, device, repeat_noise) * temperature
+        if noise_dropout > 0.:
+            noise = torch.nn.functional.dropout(noise, p=noise_dropout)
+        # no noise when t == 0
+        nonzero_mask = (1 - (t == 0).float()).reshape(b, *((1,) * (len(x.shape) - 1)))
+
+        if return_codebook_ids:
+            return model_mean + nonzero_mask * (0.5 * model_log_variance).exp() * noise, logits.argmax(dim=1)
+        if return_x0:
+            return model_mean + nonzero_mask * (0.5 * model_log_variance).exp() * noise, x0
+        else:
+            return model_mean + nonzero_mask * (0.5 * model_log_variance).exp() * noise
+
+    @torch.no_grad()
+    def progressive_denoising(self, cond, shape, verbose=True, callback=None, quantize_denoised=False,
+                              img_callback=None, mask=None, x0=None, temperature=1., noise_dropout=0.,
+                              score_corrector=None, corrector_kwargs=None, batch_size=None, x_T=None, start_T=None,
+                              log_every_t=None):
+        if not log_every_t:
+            log_every_t = self.log_every_t
+        timesteps = self.num_timesteps
+        if batch_size is not None:
+            b = batch_size if batch_size is not None else shape[0]
+            shape = [batch_size] + list(shape)
+        else:
+            b = batch_size = shape[0]
+        if x_T is None:
+            img = torch.randn(shape, device=self.device)
+        else:
+            img = x_T
+        intermediates = []
+        if cond is not None:
+            if isinstance(cond, dict):
+                cond = {key: cond[key][:batch_size] if not isinstance(cond[key], list) else
+                list(map(lambda x: x[:batch_size], cond[key])) for key in cond}
+            else:
+                cond = [c[:batch_size] for c in cond] if isinstance(cond, list) else cond[:batch_size]
+
+        if start_T is not None:
+            timesteps = min(timesteps, start_T)
+        iterator = tqdm(reversed(range(0, timesteps)), desc='Progressive Generation',
+                        total=timesteps) if verbose else reversed(
+            range(0, timesteps))
+        if type(temperature) == float:
+            temperature = [temperature] * timesteps
+
+        for i in iterator:
+            ts = torch.full((b,), i, device=self.device, dtype=torch.long)
+            if self.shorten_cond_schedule:
+                assert self.model.conditioning_key != 'hybrid'
+                tc = self.cond_ids[ts].to(cond.device)
+                cond = self.q_sample(x_start=cond, t=tc, noise=torch.randn_like(cond))
+
+            img, x0_partial = self.p_sample(img, cond, ts,
+                                            clip_denoised=self.clip_denoised,
+                                            quantize_denoised=quantize_denoised, return_x0=True,
+                                            temperature=temperature[i], noise_dropout=noise_dropout,
+                                            score_corrector=score_corrector, corrector_kwargs=corrector_kwargs)
+            if mask is not None:
+                assert x0 is not None
+                img_orig = self.q_sample(x0, ts)
+                img = img_orig * mask + (1. - mask) * img
+
+            if i % log_every_t == 0 or i == timesteps - 1:
+                intermediates.append(x0_partial)
+            if callback: callback(i)
+            if img_callback: img_callback(img, i)
+        return img, intermediates
+
+    @torch.no_grad()
+    def p_sample_loop(self, cond, shape, return_intermediates=False,
+                      x_T=None, verbose=True, callback=None, timesteps=None, quantize_denoised=False,
+                      mask=None, x0=None, img_callback=None, start_T=None,
+                      log_every_t=None):
+
+        if not log_every_t:
+            log_every_t = self.log_every_t
+        device = self.betas.device
+        b = shape[0]
+        if x_T is None:
+            img = torch.randn(shape, device=device)
+        else:
+            img = x_T
+
+        intermediates = [img]
+        if timesteps is None:
+            timesteps = self.num_timesteps
+
+        if start_T is not None:
+            timesteps = min(timesteps, start_T)
+        iterator = tqdm(reversed(range(0, timesteps)), desc='Sampling t', total=timesteps) if verbose else reversed(
+            range(0, timesteps))
+
+        if mask is not None:
+            assert x0 is not None
+            assert x0.shape[2:3] == mask.shape[2:3]  # spatial size has to match
+
+        for i in iterator:
+            ts = torch.full((b,), i, device=device, dtype=torch.long)
+            if self.shorten_cond_schedule:
+                assert self.model.conditioning_key != 'hybrid'
+                tc = self.cond_ids[ts].to(cond.device)
+                cond = self.q_sample(x_start=cond, t=tc, noise=torch.randn_like(cond))
+
+            img = self.p_sample(img, cond, ts,
+                                clip_denoised=self.clip_denoised,
+                                quantize_denoised=quantize_denoised)
+            if mask is not None:
+                img_orig = self.q_sample(x0, ts)
+                img = img_orig * mask + (1. - mask) * img
+
+            if i % log_every_t == 0 or i == timesteps - 1:
+                intermediates.append(img)
+            if callback: callback(i)
+            if img_callback: img_callback(img, i)
+
+        if return_intermediates:
+            return img, intermediates
+        return img
+
+    @torch.no_grad()
+    def sample(self, cond, batch_size=16, return_intermediates=False, x_T=None,
+               verbose=True, timesteps=None, quantize_denoised=False,
+               mask=None, x0=None, shape=None,**kwargs):
+        if shape is None:
+            shape = (batch_size, self.channels, self.image_size, self.image_size)
+        if cond is not None:
+            if isinstance(cond, dict):
+                cond = {key: cond[key][:batch_size] if not isinstance(cond[key], list) else
+                list(map(lambda x: x[:batch_size], cond[key])) for key in cond}
+            else:
+                cond = [c[:batch_size] for c in cond] if isinstance(cond, list) else cond[:batch_size]
+        return self.p_sample_loop(cond,
+                                  shape,
+                                  return_intermediates=return_intermediates, x_T=x_T,
+                                  verbose=verbose, timesteps=timesteps, quantize_denoised=quantize_denoised,
+                                  mask=mask, x0=x0)
+
+    @torch.no_grad()
+    def sample_log(self, cond, batch_size, ddim, ddim_steps, **kwargs):
+        if ddim:
+            ddim_sampler = DDIMSampler(self)
+            shape = (self.channels, self.image_size, self.image_size)
+            samples, intermediates = ddim_sampler.sample(ddim_steps, batch_size,
+                                                         shape, cond, verbose=False, **kwargs)
+
+        else:
+            samples, intermediates = self.sample(cond=cond, batch_size=batch_size,
+                                                 return_intermediates=True, **kwargs)
+
+        return samples, intermediates
+
+    @torch.no_grad()
+    def get_unconditional_conditioning(self, batch_size, null_label=None, image_size=512):
+        if null_label is not None:
+            xc = null_label
+            if isinstance(xc, ListConfig):
+                xc = list(xc)
+            if isinstance(xc, dict) or isinstance(xc, list):
+                c = self.get_learned_conditioning(xc)
+            else:
+                if hasattr(xc, "to"):
+                    xc = xc.to(self.device)
+                c = self.get_learned_conditioning(xc)
+        else:
+            # todo: get null label from cond_stage_model
+            raise NotImplementedError()
+        c = repeat(c, '1 ... -> b ...', b=batch_size).to(self.device)
+        cond = {}
+        cond["c_crossattn"] = [c]
+        cond["c_concat"] = [torch.zeros([batch_size, 4, image_size // 8, image_size // 8]).to(self.device)]
+        return cond
+
+    @torch.no_grad()
+    def log_images(self, batch, N=8, n_row=4, sample=True, ddim_steps=200, ddim_eta=1., return_keys=None,
+                   quantize_denoised=True, inpaint=True, plot_denoise_rows=False, plot_progressive_rows=True,
+                   plot_diffusion_rows=True, unconditional_guidance_scale=1., unconditional_guidance_label=None,
+                   use_ema_scope=True,
+                   **kwargs):
+        ema_scope = self.ema_scope if use_ema_scope else nullcontext
+        use_ddim = ddim_steps is not None
+
+        log = dict()
+        z, c, x, xrec, xc = self.get_input(batch, self.first_stage_key,
+                                           return_first_stage_outputs=True,
+                                           force_c_encode=True,
+                                           return_original_cond=True,
+                                           bs=N)
+        N = min(x.shape[0], N)
+        n_row = min(x.shape[0], n_row)
+        log["inputs"] = x
+        log["reconstruction"] = xrec
+        if self.model.conditioning_key is not None:
+            if hasattr(self.cond_stage_model, "decode"):
+                xc = self.cond_stage_model.decode(c)
+                log["conditioning"] = xc
+            elif self.cond_stage_key in ["caption", "txt"]:
+                xc = log_txt_as_img((x.shape[2], x.shape[3]), batch[self.cond_stage_key], size=x.shape[2]//25)
+                log["conditioning"] = xc
+            elif self.cond_stage_key == 'class_label':
+                xc = log_txt_as_img((x.shape[2], x.shape[3]), batch["human_label"], size=x.shape[2]//25)
+                log['conditioning'] = xc
+            elif isimage(xc):
+                log["conditioning"] = xc
+            if ismap(xc):
+                log["original_conditioning"] = self.to_rgb(xc)
+
+        if plot_diffusion_rows:
+            # get diffusion row
+            diffusion_row = list()
+            z_start = z[:n_row]
+            for t in range(self.num_timesteps):
+                if t % self.log_every_t == 0 or t == self.num_timesteps - 1:
+                    t = repeat(torch.tensor([t]), '1 -> b', b=n_row)
+                    t = t.to(self.device).long()
+                    noise = torch.randn_like(z_start)
+                    z_noisy = self.q_sample(x_start=z_start, t=t, noise=noise)
+                    diffusion_row.append(self.decode_first_stage(z_noisy))
+
+            diffusion_row = torch.stack(diffusion_row)  # n_log_step, n_row, C, H, W
+            diffusion_grid = rearrange(diffusion_row, 'n b c h w -> b n c h w')
+            diffusion_grid = rearrange(diffusion_grid, 'b n c h w -> (b n) c h w')
+            diffusion_grid = make_grid(diffusion_grid, nrow=diffusion_row.shape[0])
+            log["diffusion_row"] = diffusion_grid
+
+        if sample:
+            # get denoise row
+            with ema_scope("Sampling"):
+                samples, z_denoise_row = self.sample_log(cond=c,batch_size=N,ddim=use_ddim,
+                                                         ddim_steps=ddim_steps,eta=ddim_eta)
+                # samples, z_denoise_row = self.sample(cond=c, batch_size=N, return_intermediates=True)
+            x_samples = self.decode_first_stage(samples)
+            log["samples"] = x_samples
+            if plot_denoise_rows:
+                denoise_grid = self._get_denoise_row_from_list(z_denoise_row)
+                log["denoise_row"] = denoise_grid
+
+            if quantize_denoised and not isinstance(self.first_stage_model, AutoencoderKL) and not isinstance(
+                    self.first_stage_model, IdentityFirstStage):
+                # also display when quantizing x0 while sampling
+                with ema_scope("Plotting Quantized Denoised"):
+                    samples, z_denoise_row = self.sample_log(cond=c,batch_size=N,ddim=use_ddim,
+                                                             ddim_steps=ddim_steps,eta=ddim_eta,
+                                                             quantize_denoised=True)
+                    # samples, z_denoise_row = self.sample(cond=c, batch_size=N, return_intermediates=True,
+                    #                                      quantize_denoised=True)
+                x_samples = self.decode_first_stage(samples.to(self.device))
+                log["samples_x0_quantized"] = x_samples
+
+        if unconditional_guidance_scale > 1.0:
+            uc = self.get_unconditional_conditioning(N, unconditional_guidance_label, image_size=x.shape[-1])
+            # uc = torch.zeros_like(c)
+            with ema_scope("Sampling with classifier-free guidance"):
+                samples_cfg, _ = self.sample_log(cond=c, batch_size=N, ddim=use_ddim,
+                                                 ddim_steps=ddim_steps, eta=ddim_eta,
+                                                 unconditional_guidance_scale=unconditional_guidance_scale,
+                                                 unconditional_conditioning=uc,
+                                                 )
+                x_samples_cfg = self.decode_first_stage(samples_cfg)
+                log[f"samples_cfg_scale_{unconditional_guidance_scale:.2f}"] = x_samples_cfg
+
+        if inpaint:
+            # make a simple center square
+            b, h, w = z.shape[0], z.shape[2], z.shape[3]
+            mask = torch.ones(N, h, w).to(self.device)
+            # zeros will be filled in
+            mask[:, h // 4:3 * h // 4, w // 4:3 * w // 4] = 0.
+            mask = mask[:, None, ...]
+            with ema_scope("Plotting Inpaint"):
+
+                samples, _ = self.sample_log(cond=c,batch_size=N,ddim=use_ddim, eta=ddim_eta,
+                                            ddim_steps=ddim_steps, x0=z[:N], mask=mask)
+            x_samples = self.decode_first_stage(samples.to(self.device))
+            log["samples_inpainting"] = x_samples
+            log["mask"] = mask
+
+            # outpaint
+            mask = 1. - mask
+            with ema_scope("Plotting Outpaint"):
+                samples, _ = self.sample_log(cond=c, batch_size=N, ddim=use_ddim,eta=ddim_eta,
+                                            ddim_steps=ddim_steps, x0=z[:N], mask=mask)
+            x_samples = self.decode_first_stage(samples.to(self.device))
+            log["samples_outpainting"] = x_samples
+
+        if plot_progressive_rows:
+            with ema_scope("Plotting Progressives"):
+                img, progressives = self.progressive_denoising(c,
+                                                               shape=(self.channels, self.image_size, self.image_size),
+                                                               batch_size=N)
+            prog_row = self._get_denoise_row_from_list(progressives, desc="Progressive Generation")
+            log["progressive_row"] = prog_row
+
+        if return_keys:
+            if np.intersect1d(list(log.keys()), return_keys).shape[0] == 0:
+                return log
+            else:
+                return {key: log[key] for key in return_keys}
+        return log
+
+    def configure_optimizers(self):
+        lr = self.learning_rate
+        params = []
+        if self.unet_trainable == "attn":
+            print("Training only unet attention layers")
+            for n, m in self.model.named_modules():
+                if isinstance(m, CrossAttention) and n.endswith('attn2'):
+                    params.extend(m.parameters())
+        if self.unet_trainable == "conv_in":
+            print("Training only unet input conv layers")
+            params = list(self.model.diffusion_model.input_blocks[0][0].parameters())
+        elif self.unet_trainable is True or self.unet_trainable == "all":
+            print("Training the full unet")
+            params = list(self.model.parameters())
+        else:
+            raise ValueError(f"Unrecognised setting for unet_trainable: {self.unet_trainable}")
+
+        if self.cond_stage_trainable:
+            print(f"{self.__class__.__name__}: Also optimizing conditioner params!")
+            params = params + list(self.cond_stage_model.parameters())
+        if self.learn_logvar:
+            print('Diffusion model optimizing logvar')
+            params.append(self.logvar)
+
+        if self.cc_projection is not None:
+            params = params + list(self.cc_projection.parameters())
+            print('========== optimizing for cc projection weight ==========')
+
+        opt = torch.optim.AdamW([{"params": self.model.parameters(), "lr": lr},
+                                {"params": self.cc_projection.parameters(), "lr": 10. * lr}], lr=lr)
+        if self.use_scheduler:
+            assert 'target' in self.scheduler_config
+            scheduler = instantiate_from_config(self.scheduler_config)
+
+            print("Setting up LambdaLR scheduler...")
+            scheduler = [
+                {
+                    'scheduler': LambdaLR(opt, lr_lambda=scheduler.schedule),
+                    'interval': 'step',
+                    'frequency': 1
+                }]
+            return [opt], scheduler
+        return opt
+
+    @torch.no_grad()
+    def to_rgb(self, x):
+        x = x.float()
+        if not hasattr(self, "colorize"):
+            self.colorize = torch.randn(3, x.shape[1], 1, 1).to(x)
+        x = nn.functional.conv2d(x, weight=self.colorize)
+        x = 2. * (x - x.min()) / (x.max() - x.min()) - 1.
+        return x
+
+
+class DiffusionWrapper(pl.LightningModule):
+    def __init__(self, diff_model_config, conditioning_key):
+        super().__init__()
+        self.diffusion_model = instantiate_from_config(diff_model_config)
+        self.conditioning_key = conditioning_key
+        assert self.conditioning_key in [None, 'concat', 'crossattn', 'hybrid', 'adm', 'hybrid-adm']
+
+    def forward(self, x, t, c_concat: list = None, c_crossattn: list = None, c_adm=None):
+        if self.conditioning_key is None:
+            out = self.diffusion_model(x, t)
+        elif self.conditioning_key == 'concat':
+            xc = torch.cat([x] + c_concat, dim=1)
+            out = self.diffusion_model(xc, t)
+        elif self.conditioning_key == 'crossattn':
+            # c_crossattn dimension:  torch.Size([8, 1, 768]) 1
+            # cc dimension:  torch.Size([8, 1, 768]
+            cc = torch.cat(c_crossattn, 1)
+            out = self.diffusion_model(x, t, context=cc)
+        elif self.conditioning_key == 'hybrid':
+            xc = torch.cat([x] + c_concat, dim=1)
+            cc = torch.cat(c_crossattn, 1)
+            out = self.diffusion_model(xc, t, context=cc)
+        elif self.conditioning_key == 'hybrid-adm':
+            assert c_adm is not None
+            xc = torch.cat([x] + c_concat, dim=1)
+            cc = torch.cat(c_crossattn, 1)
+            out = self.diffusion_model(xc, t, context=cc, y=c_adm)
+        elif self.conditioning_key == 'adm':
+            cc = c_crossattn[0]
+            out = self.diffusion_model(x, t, y=cc)
+        else:
+            raise NotImplementedError()
+
+        return out
+
+
+class LatentUpscaleDiffusion(LatentDiffusion):
+    def __init__(self, *args, low_scale_config, low_scale_key="LR", **kwargs):
+        super().__init__(*args, **kwargs)
+        # assumes that neither the cond_stage nor the low_scale_model contain trainable params
+        assert not self.cond_stage_trainable
+        self.instantiate_low_stage(low_scale_config)
+        self.low_scale_key = low_scale_key
+
+    def instantiate_low_stage(self, config):
+        model = instantiate_from_config(config)
+        self.low_scale_model = model.eval()
+        self.low_scale_model.train = disabled_train
+        for param in self.low_scale_model.parameters():
+            param.requires_grad = False
+
+    @torch.no_grad()
+    def get_input(self, batch, k, cond_key=None, bs=None, log_mode=False):
+        if not log_mode:
+            z, c = super().get_input(batch, k, force_c_encode=True, bs=bs)
+        else:
+            z, c, x, xrec, xc = super().get_input(batch, self.first_stage_key, return_first_stage_outputs=True,
+                                                  force_c_encode=True, return_original_cond=True, bs=bs)
+        x_low = batch[self.low_scale_key][:bs]
+        x_low = rearrange(x_low, 'b h w c -> b c h w')
+        x_low = x_low.to(memory_format=torch.contiguous_format).float()
+        zx, noise_level = self.low_scale_model(x_low)
+        all_conds = {"c_concat": [zx], "c_crossattn": [c], "c_adm": noise_level}
+        #import pudb; pu.db
+        if log_mode:
+            # TODO: maybe disable if too expensive
+            interpretability = False
+            if interpretability:
+                zx = zx[:, :, ::2, ::2]
+            x_low_rec = self.low_scale_model.decode(zx)
+            return z, all_conds, x, xrec, xc, x_low, x_low_rec, noise_level
+        return z, all_conds
+
+    @torch.no_grad()
+    def log_images(self, batch, N=8, n_row=4, sample=True, ddim_steps=200, ddim_eta=1., return_keys=None,
+                   plot_denoise_rows=False, plot_progressive_rows=True, plot_diffusion_rows=True,
+                   unconditional_guidance_scale=1., unconditional_guidance_label=None, use_ema_scope=True,
+                   **kwargs):
+        ema_scope = self.ema_scope if use_ema_scope else nullcontext
+        use_ddim = ddim_steps is not None
+
+        log = dict()
+        z, c, x, xrec, xc, x_low, x_low_rec, noise_level = self.get_input(batch, self.first_stage_key, bs=N,
+                                                                          log_mode=True)
+        N = min(x.shape[0], N)
+        n_row = min(x.shape[0], n_row)
+        log["inputs"] = x
+        log["reconstruction"] = xrec
+        log["x_lr"] = x_low
+        log[f"x_lr_rec_@noise_levels{'-'.join(map(lambda x: str(x), list(noise_level.cpu().numpy())))}"] = x_low_rec
+        if self.model.conditioning_key is not None:
+            if hasattr(self.cond_stage_model, "decode"):
+                xc = self.cond_stage_model.decode(c)
+                log["conditioning"] = xc
+            elif self.cond_stage_key in ["caption", "txt"]:
+                xc = log_txt_as_img((x.shape[2], x.shape[3]), batch[self.cond_stage_key], size=x.shape[2]//25)
+                log["conditioning"] = xc
+            elif self.cond_stage_key == 'class_label':
+                xc = log_txt_as_img((x.shape[2], x.shape[3]), batch["human_label"], size=x.shape[2]//25)
+                log['conditioning'] = xc
+            elif isimage(xc):
+                log["conditioning"] = xc
+            if ismap(xc):
+                log["original_conditioning"] = self.to_rgb(xc)
+
+        if plot_diffusion_rows:
+            # get diffusion row
+            diffusion_row = list()
+            z_start = z[:n_row]
+            for t in range(self.num_timesteps):
+                if t % self.log_every_t == 0 or t == self.num_timesteps - 1:
+                    t = repeat(torch.tensor([t]), '1 -> b', b=n_row)
+                    t = t.to(self.device).long()
+                    noise = torch.randn_like(z_start)
+                    z_noisy = self.q_sample(x_start=z_start, t=t, noise=noise)
+                    diffusion_row.append(self.decode_first_stage(z_noisy))
+
+            diffusion_row = torch.stack(diffusion_row)  # n_log_step, n_row, C, H, W
+            diffusion_grid = rearrange(diffusion_row, 'n b c h w -> b n c h w')
+            diffusion_grid = rearrange(diffusion_grid, 'b n c h w -> (b n) c h w')
+            diffusion_grid = make_grid(diffusion_grid, nrow=diffusion_row.shape[0])
+            log["diffusion_row"] = diffusion_grid
+
+        if sample:
+            # get denoise row
+            with ema_scope("Sampling"):
+                samples, z_denoise_row = self.sample_log(cond=c, batch_size=N, ddim=use_ddim,
+                                                         ddim_steps=ddim_steps, eta=ddim_eta)
+                # samples, z_denoise_row = self.sample(cond=c, batch_size=N, return_intermediates=True)
+            x_samples = self.decode_first_stage(samples)
+            log["samples"] = x_samples
+            if plot_denoise_rows:
+                denoise_grid = self._get_denoise_row_from_list(z_denoise_row)
+                log["denoise_row"] = denoise_grid
+
+        if unconditional_guidance_scale > 1.0:
+            uc_tmp = self.get_unconditional_conditioning(N, unconditional_guidance_label)
+            # TODO explore better "unconditional" choices for the other keys
+            # maybe guide away from empty text label and highest noise level and maximally degraded zx?
+            uc = dict()
+            for k in c:
+                if k == "c_crossattn":
+                    assert isinstance(c[k], list) and len(c[k]) == 1
+                    uc[k] = [uc_tmp]
+                elif k == "c_adm":  # todo: only run with text-based guidance?
+                    assert isinstance(c[k], torch.Tensor)
+                    uc[k] = torch.ones_like(c[k]) * self.low_scale_model.max_noise_level
+                elif isinstance(c[k], list):
+                    uc[k] = [c[k][i] for i in range(len(c[k]))]
+                else:
+                    uc[k] = c[k]
+
+            with ema_scope("Sampling with classifier-free guidance"):
+                samples_cfg, _ = self.sample_log(cond=c, batch_size=N, ddim=use_ddim,
+                                                 ddim_steps=ddim_steps, eta=ddim_eta,
+                                                 unconditional_guidance_scale=unconditional_guidance_scale,
+                                                 unconditional_conditioning=uc,
+                                                 )
+                x_samples_cfg = self.decode_first_stage(samples_cfg)
+                log[f"samples_cfg_scale_{unconditional_guidance_scale:.2f}"] = x_samples_cfg
+
+        if plot_progressive_rows:
+            with ema_scope("Plotting Progressives"):
+                img, progressives = self.progressive_denoising(c,
+                                                               shape=(self.channels, self.image_size, self.image_size),
+                                                               batch_size=N)
+            prog_row = self._get_denoise_row_from_list(progressives, desc="Progressive Generation")
+            log["progressive_row"] = prog_row
+
+        return log
+
+
+class LatentInpaintDiffusion(LatentDiffusion):
+    """
+    can either run as pure inpainting model (only concat mode) or with mixed conditionings,
+    e.g. mask as concat and text via cross-attn.
+    To disable finetuning mode, set finetune_keys to None
+     """
+    def __init__(self,
+                 finetune_keys=("model.diffusion_model.input_blocks.0.0.weight",
+                                "model_ema.diffusion_modelinput_blocks00weight"
+                                ),
+                 concat_keys=("mask", "masked_image"),
+                 masked_image_key="masked_image",
+                 keep_finetune_dims=4,  # if model was trained without concat mode before and we would like to keep these channels
+                 c_concat_log_start=None, # to log reconstruction of c_concat codes
+                 c_concat_log_end=None,
+                 *args, **kwargs
+                 ):
+        ckpt_path = kwargs.pop("ckpt_path", None)
+        ignore_keys = kwargs.pop("ignore_keys", list())
+        super().__init__(*args, **kwargs)
+        self.masked_image_key = masked_image_key
+        assert self.masked_image_key in concat_keys
+        self.finetune_keys = finetune_keys
+        self.concat_keys = concat_keys
+        self.keep_dims = keep_finetune_dims
+        self.c_concat_log_start = c_concat_log_start
+        self.c_concat_log_end = c_concat_log_end
+        if exists(self.finetune_keys): assert exists(ckpt_path), 'can only finetune from a given checkpoint'
+        if exists(ckpt_path):
+            self.init_from_ckpt(ckpt_path, ignore_keys)
+
+    def init_from_ckpt(self, path, ignore_keys=list(), only_model=False):
+        sd = torch.load(path, map_location="cpu")
+        if "state_dict" in list(sd.keys()):
+            sd = sd["state_dict"]
+        keys = list(sd.keys())
+        for k in keys:
+            for ik in ignore_keys:
+                if k.startswith(ik):
+                    print("Deleting key {} from state_dict.".format(k))
+                    del sd[k]
+
+            # make it explicit, finetune by including extra input channels
+            if exists(self.finetune_keys) and k in self.finetune_keys:
+                new_entry = None
+                for name, param in self.named_parameters():
+                    if name in self.finetune_keys:
+                        print(f"modifying key '{name}' and keeping its original {self.keep_dims} (channels) dimensions only")
+                        new_entry = torch.zeros_like(param)  # zero init
+                assert exists(new_entry), 'did not find matching parameter to modify'
+                new_entry[:, :self.keep_dims, ...] = sd[k]
+                sd[k] = new_entry
+
+        missing, unexpected = self.load_state_dict(sd, strict=False) if not only_model else self.model.load_state_dict(sd, strict=False)
+        print(f"Restored from {path} with {len(missing)} missing and {len(unexpected)} unexpected keys")
+        if len(missing) > 0:
+            print(f"Missing Keys: {missing}")
+        if len(unexpected) > 0:
+            print(f"Unexpected Keys: {unexpected}")
+
+    @torch.no_grad()
+    def get_input(self, batch, k, cond_key=None, bs=None, return_first_stage_outputs=False):
+        # note: restricted to non-trainable encoders currently
+        assert not self.cond_stage_trainable, 'trainable cond stages not yet supported for inpainting'
+        z, c, x, xrec, xc = super().get_input(batch, self.first_stage_key, return_first_stage_outputs=True,
+                                              force_c_encode=True, return_original_cond=True, bs=bs)
+
+        assert exists(self.concat_keys)
+        c_cat = list()
+        for ck in self.concat_keys:
+            cc = rearrange(batch[ck], 'b h w c -> b c h w').to(memory_format=torch.contiguous_format).float()
+            if bs is not None:
+                cc = cc[:bs]
+                cc = cc.to(self.device)
+            bchw = z.shape
+            if ck != self.masked_image_key:
+                cc = torch.nn.functional.interpolate(cc, size=bchw[-2:])
+            else:
+                cc = self.get_first_stage_encoding(self.encode_first_stage(cc))
+            c_cat.append(cc)
+        c_cat = torch.cat(c_cat, dim=1)
+        all_conds = {"c_concat": [c_cat], "c_crossattn": [c]}
+        if return_first_stage_outputs:
+            return z, all_conds, x, xrec, xc
+        return z, all_conds
+
+    @torch.no_grad()
+    def log_images(self, batch, N=8, n_row=4, sample=True, ddim_steps=200, ddim_eta=1., return_keys=None,
+                   quantize_denoised=True, inpaint=True, plot_denoise_rows=False, plot_progressive_rows=True,
+                   plot_diffusion_rows=True, unconditional_guidance_scale=1., unconditional_guidance_label=None,
+                   use_ema_scope=True,
+                   **kwargs):
+        ema_scope = self.ema_scope if use_ema_scope else nullcontext
+        use_ddim = ddim_steps is not None
+
+        log = dict()
+        z, c, x, xrec, xc = self.get_input(batch, self.first_stage_key, bs=N, return_first_stage_outputs=True)
+        c_cat, c = c["c_concat"][0], c["c_crossattn"][0]
+        N = min(x.shape[0], N)
+        n_row = min(x.shape[0], n_row)
+        log["inputs"] = x
+        log["reconstruction"] = xrec
+        if self.model.conditioning_key is not None:
+            if hasattr(self.cond_stage_model, "decode"):
+                xc = self.cond_stage_model.decode(c)
+                log["conditioning"] = xc
+            elif self.cond_stage_key in ["caption", "txt"]:
+                xc = log_txt_as_img((x.shape[2], x.shape[3]), batch[self.cond_stage_key], size=x.shape[2] // 25)
+                log["conditioning"] = xc
+            elif self.cond_stage_key == 'class_label':
+                xc = log_txt_as_img((x.shape[2], x.shape[3]), batch["human_label"], size=x.shape[2] // 25)
+                log['conditioning'] = xc
+            elif isimage(xc):
+                log["conditioning"] = xc
+            if ismap(xc):
+                log["original_conditioning"] = self.to_rgb(xc)
+
+        if not (self.c_concat_log_start is None and self.c_concat_log_end is None):
+            log["c_concat_decoded"] = self.decode_first_stage(c_cat[:,self.c_concat_log_start:self.c_concat_log_end])
+
+        if plot_diffusion_rows:
+            # get diffusion row
+            diffusion_row = list()
+            z_start = z[:n_row]
+            for t in range(self.num_timesteps):
+                if t % self.log_every_t == 0 or t == self.num_timesteps - 1:
+                    t = repeat(torch.tensor([t]), '1 -> b', b=n_row)
+                    t = t.to(self.device).long()
+                    noise = torch.randn_like(z_start)
+                    z_noisy = self.q_sample(x_start=z_start, t=t, noise=noise)
+                    diffusion_row.append(self.decode_first_stage(z_noisy))
+
+            diffusion_row = torch.stack(diffusion_row)  # n_log_step, n_row, C, H, W
+            diffusion_grid = rearrange(diffusion_row, 'n b c h w -> b n c h w')
+            diffusion_grid = rearrange(diffusion_grid, 'b n c h w -> (b n) c h w')
+            diffusion_grid = make_grid(diffusion_grid, nrow=diffusion_row.shape[0])
+            log["diffusion_row"] = diffusion_grid
+
+        if sample:
+            # get denoise row
+            with ema_scope("Sampling"):
+                samples, z_denoise_row = self.sample_log(cond={"c_concat": [c_cat], "c_crossattn": [c]},
+                                                         batch_size=N, ddim=use_ddim,
+                                                         ddim_steps=ddim_steps, eta=ddim_eta)
+                # samples, z_denoise_row = self.sample(cond=c, batch_size=N, return_intermediates=True)
+            x_samples = self.decode_first_stage(samples)
+            log["samples"] = x_samples
+            if plot_denoise_rows:
+                denoise_grid = self._get_denoise_row_from_list(z_denoise_row)
+                log["denoise_row"] = denoise_grid
+
+        if unconditional_guidance_scale > 1.0:
+            uc_cross = self.get_unconditional_conditioning(N, unconditional_guidance_label)
+            uc_cat = c_cat
+            uc_full = {"c_concat": [uc_cat], "c_crossattn": [uc_cross]}
+            with ema_scope("Sampling with classifier-free guidance"):
+                samples_cfg, _ = self.sample_log(cond={"c_concat": [c_cat], "c_crossattn": [c]},
+                                                 batch_size=N, ddim=use_ddim,
+                                                 ddim_steps=ddim_steps, eta=ddim_eta,
+                                                 unconditional_guidance_scale=unconditional_guidance_scale,
+                                                 unconditional_conditioning=uc_full,
+                                                 )
+                x_samples_cfg = self.decode_first_stage(samples_cfg)
+                log[f"samples_cfg_scale_{unconditional_guidance_scale:.2f}"] = x_samples_cfg
+
+        log["masked_image"] = rearrange(batch["masked_image"],
+                                        'b h w c -> b c h w').to(memory_format=torch.contiguous_format).float()
+        return log
+
+
+class Layout2ImgDiffusion(LatentDiffusion):
+    # TODO: move all layout-specific hacks to this class
+    def __init__(self, cond_stage_key, *args, **kwargs):
+        assert cond_stage_key == 'coordinates_bbox', 'Layout2ImgDiffusion only for cond_stage_key="coordinates_bbox"'
+        super().__init__(cond_stage_key=cond_stage_key, *args, **kwargs)
+
+    def log_images(self, batch, N=8, *args, **kwargs):
+        logs = super().log_images(batch=batch, N=N, *args, **kwargs)
+
+        key = 'train' if self.training else 'validation'
+        dset = self.trainer.datamodule.datasets[key]
+        mapper = dset.conditional_builders[self.cond_stage_key]
+
+        bbox_imgs = []
+        map_fn = lambda catno: dset.get_textual_label(dset.get_category_id(catno))
+        for tknzd_bbox in batch[self.cond_stage_key][:N]:
+            bboximg = mapper.plot(tknzd_bbox.detach().cpu(), map_fn, (256, 256))
+            bbox_imgs.append(bboximg)
+
+        cond_img = torch.stack(bbox_imgs, dim=0)
+        logs['bbox_image'] = cond_img
+        return logs
+
+
+class SimpleUpscaleDiffusion(LatentDiffusion):
+    def __init__(self, *args, low_scale_key="LR", **kwargs):
+        super().__init__(*args, **kwargs)
+        # assumes that neither the cond_stage nor the low_scale_model contain trainable params
+        assert not self.cond_stage_trainable
+        self.low_scale_key = low_scale_key
+
+    @torch.no_grad()
+    def get_input(self, batch, k, cond_key=None, bs=None, log_mode=False):
+        if not log_mode:
+            z, c = super().get_input(batch, k, force_c_encode=True, bs=bs)
+        else:
+            z, c, x, xrec, xc = super().get_input(batch, self.first_stage_key, return_first_stage_outputs=True,
+                                                  force_c_encode=True, return_original_cond=True, bs=bs)
+        x_low = batch[self.low_scale_key][:bs]
+        x_low = rearrange(x_low, 'b h w c -> b c h w')
+        x_low = x_low.to(memory_format=torch.contiguous_format).float()
+
+        encoder_posterior = self.encode_first_stage(x_low)
+        zx = self.get_first_stage_encoding(encoder_posterior).detach()
+        all_conds = {"c_concat": [zx], "c_crossattn": [c]}
+
+        if log_mode:
+            # TODO: maybe disable if too expensive
+            interpretability = False
+            if interpretability:
+                zx = zx[:, :, ::2, ::2]
+            return z, all_conds, x, xrec, xc, x_low
+        return z, all_conds
+
+    @torch.no_grad()
+    def log_images(self, batch, N=8, n_row=4, sample=True, ddim_steps=200, ddim_eta=1., return_keys=None,
+                   plot_denoise_rows=False, plot_progressive_rows=True, plot_diffusion_rows=True,
+                   unconditional_guidance_scale=1., unconditional_guidance_label=None, use_ema_scope=True,
+                   **kwargs):
+        ema_scope = self.ema_scope if use_ema_scope else nullcontext
+        use_ddim = ddim_steps is not None
+
+        log = dict()
+        z, c, x, xrec, xc, x_low = self.get_input(batch, self.first_stage_key, bs=N, log_mode=True)
+        N = min(x.shape[0], N)
+        n_row = min(x.shape[0], n_row)
+        log["inputs"] = x
+        log["reconstruction"] = xrec
+        log["x_lr"] = x_low
+
+        if self.model.conditioning_key is not None:
+            if hasattr(self.cond_stage_model, "decode"):
+                xc = self.cond_stage_model.decode(c)
+                log["conditioning"] = xc
+            elif self.cond_stage_key in ["caption", "txt"]:
+                xc = log_txt_as_img((x.shape[2], x.shape[3]), batch[self.cond_stage_key], size=x.shape[2]//25)
+                log["conditioning"] = xc
+            elif self.cond_stage_key == 'class_label':
+                xc = log_txt_as_img((x.shape[2], x.shape[3]), batch["human_label"], size=x.shape[2]//25)
+                log['conditioning'] = xc
+            elif isimage(xc):
+                log["conditioning"] = xc
+            if ismap(xc):
+                log["original_conditioning"] = self.to_rgb(xc)
+
+        if sample:
+            # get denoise row
+            with ema_scope("Sampling"):
+                samples, z_denoise_row = self.sample_log(cond=c, batch_size=N, ddim=use_ddim,
+                                                         ddim_steps=ddim_steps, eta=ddim_eta)
+                # samples, z_denoise_row = self.sample(cond=c, batch_size=N, return_intermediates=True)
+            x_samples = self.decode_first_stage(samples)
+            log["samples"] = x_samples
+
+        if unconditional_guidance_scale > 1.0:
+            uc_tmp = self.get_unconditional_conditioning(N, unconditional_guidance_label)
+            uc = dict()
+            for k in c:
+                if k == "c_crossattn":
+                    assert isinstance(c[k], list) and len(c[k]) == 1
+                    uc[k] = [uc_tmp]
+                elif isinstance(c[k], list):
+                    uc[k] = [c[k][i] for i in range(len(c[k]))]
+                else:
+                    uc[k] = c[k]
+
+            with ema_scope("Sampling with classifier-free guidance"):
+                samples_cfg, _ = self.sample_log(cond=c, batch_size=N, ddim=use_ddim,
+                                                 ddim_steps=ddim_steps, eta=ddim_eta,
+                                                 unconditional_guidance_scale=unconditional_guidance_scale,
+                                                 unconditional_conditioning=uc,
+                                                 )
+                x_samples_cfg = self.decode_first_stage(samples_cfg)
+                log[f"samples_cfg_scale_{unconditional_guidance_scale:.2f}"] = x_samples_cfg
+        return log
+
+class MultiCatFrameDiffusion(LatentDiffusion):
+    def __init__(self, *args, low_scale_key="LR", **kwargs):
+        super().__init__(*args, **kwargs)
+        # assumes that neither the cond_stage nor the low_scale_model contain trainable params
+        assert not self.cond_stage_trainable
+        self.low_scale_key = low_scale_key
+
+    @torch.no_grad()
+    def get_input(self, batch, k, cond_key=None, bs=None, log_mode=False):
+        n = 2
+        if not log_mode:
+            z, c = super().get_input(batch, k, force_c_encode=True, bs=bs)
+        else:
+            z, c, x, xrec, xc = super().get_input(batch, self.first_stage_key, return_first_stage_outputs=True,
+                                                  force_c_encode=True, return_original_cond=True, bs=bs)
+        cat_conds = batch[self.low_scale_key][:bs]
+        cats = []
+        for i in range(n):
+            x_low = cat_conds[:,:,:,3*i:3*(i+1)]
+            x_low = rearrange(x_low, 'b h w c -> b c h w')
+            x_low = x_low.to(memory_format=torch.contiguous_format).float()
+            encoder_posterior = self.encode_first_stage(x_low)
+            zx = self.get_first_stage_encoding(encoder_posterior).detach()
+            cats.append(zx)
+
+        all_conds = {"c_concat": [torch.cat(cats, dim=1)], "c_crossattn": [c]}
+
+        if log_mode:
+            # TODO: maybe disable if too expensive
+            interpretability = False
+            if interpretability:
+                zx = zx[:, :, ::2, ::2]
+            return z, all_conds, x, xrec, xc, x_low
+        return z, all_conds
+
+    @torch.no_grad()
+    def log_images(self, batch, N=8, n_row=4, sample=True, ddim_steps=200, ddim_eta=1., return_keys=None,
+                   plot_denoise_rows=False, plot_progressive_rows=True, plot_diffusion_rows=True,
+                   unconditional_guidance_scale=1., unconditional_guidance_label=None, use_ema_scope=True,
+                   **kwargs):
+        ema_scope = self.ema_scope if use_ema_scope else nullcontext
+        use_ddim = ddim_steps is not None
+
+        log = dict()
+        z, c, x, xrec, xc, x_low = self.get_input(batch, self.first_stage_key, bs=N, log_mode=True)
+        N = min(x.shape[0], N)
+        n_row = min(x.shape[0], n_row)
+        log["inputs"] = x
+        log["reconstruction"] = xrec
+        log["x_lr"] = x_low
+
+        if self.model.conditioning_key is not None:
+            if hasattr(self.cond_stage_model, "decode"):
+                xc = self.cond_stage_model.decode(c)
+                log["conditioning"] = xc
+            elif self.cond_stage_key in ["caption", "txt"]:
+                xc = log_txt_as_img((x.shape[2], x.shape[3]), batch[self.cond_stage_key], size=x.shape[2]//25)
+                log["conditioning"] = xc
+            elif self.cond_stage_key == 'class_label':
+                xc = log_txt_as_img((x.shape[2], x.shape[3]), batch["human_label"], size=x.shape[2]//25)
+                log['conditioning'] = xc
+            elif isimage(xc):
+                log["conditioning"] = xc
+            if ismap(xc):
+                log["original_conditioning"] = self.to_rgb(xc)
+
+        if sample:
+            # get denoise row
+            with ema_scope("Sampling"):
+                samples, z_denoise_row = self.sample_log(cond=c, batch_size=N, ddim=use_ddim,
+                                                         ddim_steps=ddim_steps, eta=ddim_eta)
+                # samples, z_denoise_row = self.sample(cond=c, batch_size=N, return_intermediates=True)
+            x_samples = self.decode_first_stage(samples)
+            log["samples"] = x_samples
+
+        if unconditional_guidance_scale > 1.0:
+            uc_tmp = self.get_unconditional_conditioning(N, unconditional_guidance_label)
+            uc = dict()
+            for k in c:
+                if k == "c_crossattn":
+                    assert isinstance(c[k], list) and len(c[k]) == 1
+                    uc[k] = [uc_tmp]
+                elif isinstance(c[k], list):
+                    uc[k] = [c[k][i] for i in range(len(c[k]))]
+                else:
+                    uc[k] = c[k]
+
+            with ema_scope("Sampling with classifier-free guidance"):
+                samples_cfg, _ = self.sample_log(cond=c, batch_size=N, ddim=use_ddim,
+                                                 ddim_steps=ddim_steps, eta=ddim_eta,
+                                                 unconditional_guidance_scale=unconditional_guidance_scale,
+                                                 unconditional_conditioning=uc,
+                                                 )
+                x_samples_cfg = self.decode_first_stage(samples_cfg)
+                log[f"samples_cfg_scale_{unconditional_guidance_scale:.2f}"] = x_samples_cfg
+        return log

ldm/models/diffusion/plms.py

@@ -0,0 +1,259 @@
+"""SAMPLING ONLY."""
+
+import torch
+import numpy as np
+from tqdm import tqdm
+from functools import partial
+
+from ldm.modules.diffusionmodules.util import make_ddim_sampling_parameters, make_ddim_timesteps, noise_like
+from ldm.models.diffusion.sampling_util import norm_thresholding
+
+
+class PLMSSampler(object):
+    def __init__(self, model, schedule="linear", **kwargs):
+        super().__init__()
+        self.model = model
+        self.ddpm_num_timesteps = model.num_timesteps
+        self.schedule = schedule
+
+    def register_buffer(self, name, attr):
+        if type(attr) == torch.Tensor:
+            if attr.device != torch.device("cuda"):
+                attr = attr.to(torch.device("cuda"))
+        setattr(self, name, attr)
+
+    def make_schedule(self, ddim_num_steps, ddim_discretize="uniform", ddim_eta=0., verbose=True):
+        if ddim_eta != 0:
+            raise ValueError('ddim_eta must be 0 for PLMS')
+        self.ddim_timesteps = make_ddim_timesteps(ddim_discr_method=ddim_discretize, num_ddim_timesteps=ddim_num_steps,
+                                                  num_ddpm_timesteps=self.ddpm_num_timesteps,verbose=verbose)
+        alphas_cumprod = self.model.alphas_cumprod
+        assert alphas_cumprod.shape[0] == self.ddpm_num_timesteps, 'alphas have to be defined for each timestep'
+        to_torch = lambda x: x.clone().detach().to(torch.float32).to(self.model.device)
+
+        self.register_buffer('betas', to_torch(self.model.betas))
+        self.register_buffer('alphas_cumprod', to_torch(alphas_cumprod))
+        self.register_buffer('alphas_cumprod_prev', to_torch(self.model.alphas_cumprod_prev))
+
+        # calculations for diffusion q(x_t | x_{t-1}) and others
+        self.register_buffer('sqrt_alphas_cumprod', to_torch(np.sqrt(alphas_cumprod.cpu())))
+        self.register_buffer('sqrt_one_minus_alphas_cumprod', to_torch(np.sqrt(1. - alphas_cumprod.cpu())))
+        self.register_buffer('log_one_minus_alphas_cumprod', to_torch(np.log(1. - alphas_cumprod.cpu())))
+        self.register_buffer('sqrt_recip_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod.cpu())))
+        self.register_buffer('sqrt_recipm1_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod.cpu() - 1)))
+
+        # ddim sampling parameters
+        ddim_sigmas, ddim_alphas, ddim_alphas_prev = make_ddim_sampling_parameters(alphacums=alphas_cumprod.cpu(),
+                                                                                   ddim_timesteps=self.ddim_timesteps,
+                                                                                   eta=ddim_eta,verbose=verbose)
+        self.register_buffer('ddim_sigmas', ddim_sigmas)
+        self.register_buffer('ddim_alphas', ddim_alphas)
+        self.register_buffer('ddim_alphas_prev', ddim_alphas_prev)
+        self.register_buffer('ddim_sqrt_one_minus_alphas', np.sqrt(1. - ddim_alphas))
+        sigmas_for_original_sampling_steps = ddim_eta * torch.sqrt(
+            (1 - self.alphas_cumprod_prev) / (1 - self.alphas_cumprod) * (
+                        1 - self.alphas_cumprod / self.alphas_cumprod_prev))
+        self.register_buffer('ddim_sigmas_for_original_num_steps', sigmas_for_original_sampling_steps)
+
+    @torch.no_grad()
+    def sample(self,
+               S,
+               batch_size,
+               shape,
+               conditioning=None,
+               callback=None,
+               normals_sequence=None,
+               img_callback=None,
+               quantize_x0=False,
+               eta=0.,
+               mask=None,
+               x0=None,
+               temperature=1.,
+               noise_dropout=0.,
+               score_corrector=None,
+               corrector_kwargs=None,
+               verbose=True,
+               x_T=None,
+               log_every_t=100,
+               unconditional_guidance_scale=1.,
+               unconditional_conditioning=None,
+               # this has to come in the same format as the conditioning, # e.g. as encoded tokens, ...
+               dynamic_threshold=None,
+               **kwargs
+               ):
+        if conditioning is not None:
+            if isinstance(conditioning, dict):
+                ctmp = conditioning[list(conditioning.keys())[0]]
+                while isinstance(ctmp, list): ctmp = ctmp[0]
+                cbs = ctmp.shape[0]
+                if cbs != batch_size:
+                    print(f"Warning: Got {cbs} conditionings but batch-size is {batch_size}")
+            else:
+                if conditioning.shape[0] != batch_size:
+                    print(f"Warning: Got {conditioning.shape[0]} conditionings but batch-size is {batch_size}")
+
+        self.make_schedule(ddim_num_steps=S, ddim_eta=eta, verbose=verbose)
+        # sampling
+        C, H, W = shape
+        size = (batch_size, C, H, W)
+        print(f'Data shape for PLMS sampling is {size}')
+
+        samples, intermediates = self.plms_sampling(conditioning, size,
+                                                    callback=callback,
+                                                    img_callback=img_callback,
+                                                    quantize_denoised=quantize_x0,
+                                                    mask=mask, x0=x0,
+                                                    ddim_use_original_steps=False,
+                                                    noise_dropout=noise_dropout,
+                                                    temperature=temperature,
+                                                    score_corrector=score_corrector,
+                                                    corrector_kwargs=corrector_kwargs,
+                                                    x_T=x_T,
+                                                    log_every_t=log_every_t,
+                                                    unconditional_guidance_scale=unconditional_guidance_scale,
+                                                    unconditional_conditioning=unconditional_conditioning,
+                                                    dynamic_threshold=dynamic_threshold,
+                                                    )
+        return samples, intermediates
+
+    @torch.no_grad()
+    def plms_sampling(self, cond, shape,
+                      x_T=None, ddim_use_original_steps=False,
+                      callback=None, timesteps=None, quantize_denoised=False,
+                      mask=None, x0=None, img_callback=None, log_every_t=100,
+                      temperature=1., noise_dropout=0., score_corrector=None, corrector_kwargs=None,
+                      unconditional_guidance_scale=1., unconditional_conditioning=None,
+                      dynamic_threshold=None):
+        device = self.model.betas.device
+        b = shape[0]
+        if x_T is None:
+            img = torch.randn(shape, device=device)
+        else:
+            img = x_T
+
+        if timesteps is None:
+            timesteps = self.ddpm_num_timesteps if ddim_use_original_steps else self.ddim_timesteps
+        elif timesteps is not None and not ddim_use_original_steps:
+            subset_end = int(min(timesteps / self.ddim_timesteps.shape[0], 1) * self.ddim_timesteps.shape[0]) - 1
+            timesteps = self.ddim_timesteps[:subset_end]
+
+        intermediates = {'x_inter': [img], 'pred_x0': [img]}
+        time_range = list(reversed(range(0,timesteps))) if ddim_use_original_steps else np.flip(timesteps)
+        total_steps = timesteps if ddim_use_original_steps else timesteps.shape[0]
+        print(f"Running PLMS Sampling with {total_steps} timesteps")
+
+        iterator = tqdm(time_range, desc='PLMS Sampler', total=total_steps)
+        old_eps = []
+
+        for i, step in enumerate(iterator):
+            index = total_steps - i - 1
+            ts = torch.full((b,), step, device=device, dtype=torch.long)
+            ts_next = torch.full((b,), time_range[min(i + 1, len(time_range) - 1)], device=device, dtype=torch.long)
+
+            if mask is not None:
+                assert x0 is not None
+                img_orig = self.model.q_sample(x0, ts)  # TODO: deterministic forward pass?
+                img = img_orig * mask + (1. - mask) * img
+
+            outs = self.p_sample_plms(img, cond, ts, index=index, use_original_steps=ddim_use_original_steps,
+                                      quantize_denoised=quantize_denoised, temperature=temperature,
+                                      noise_dropout=noise_dropout, score_corrector=score_corrector,
+                                      corrector_kwargs=corrector_kwargs,
+                                      unconditional_guidance_scale=unconditional_guidance_scale,
+                                      unconditional_conditioning=unconditional_conditioning,
+                                      old_eps=old_eps, t_next=ts_next,
+                                      dynamic_threshold=dynamic_threshold)
+            img, pred_x0, e_t = outs
+            old_eps.append(e_t)
+            if len(old_eps) >= 4:
+                old_eps.pop(0)
+            if callback: callback(i)
+            if img_callback: img_callback(pred_x0, i)
+
+            if index % log_every_t == 0 or index == total_steps - 1:
+                intermediates['x_inter'].append(img)
+                intermediates['pred_x0'].append(pred_x0)
+
+        return img, intermediates
+
+    @torch.no_grad()
+    def p_sample_plms(self, x, c, t, index, repeat_noise=False, use_original_steps=False, quantize_denoised=False,
+                      temperature=1., noise_dropout=0., score_corrector=None, corrector_kwargs=None,
+                      unconditional_guidance_scale=1., unconditional_conditioning=None, old_eps=None, t_next=None,
+                      dynamic_threshold=None):
+        b, *_, device = *x.shape, x.device
+
+        def get_model_output(x, t):
+            if unconditional_conditioning is None or unconditional_guidance_scale == 1.:
+                e_t = self.model.apply_model(x, t, c)
+            else:
+                x_in = torch.cat([x] * 2)
+                t_in = torch.cat([t] * 2)
+                if isinstance(c, dict):
+                    assert isinstance(unconditional_conditioning, dict)
+                    c_in = dict()
+                    for k in c:
+                        if isinstance(c[k], list):
+                            c_in[k] = [torch.cat([
+                                unconditional_conditioning[k][i],
+                                c[k][i]]) for i in range(len(c[k]))]
+                        else:
+                            c_in[k] = torch.cat([
+                                    unconditional_conditioning[k],
+                                    c[k]])
+                else:
+                    c_in = torch.cat([unconditional_conditioning, c])
+                e_t_uncond, e_t = self.model.apply_model(x_in, t_in, c_in).chunk(2)
+                e_t = e_t_uncond + unconditional_guidance_scale * (e_t - e_t_uncond)
+
+            if score_corrector is not None:
+                assert self.model.parameterization == "eps"
+                e_t = score_corrector.modify_score(self.model, e_t, x, t, c, **corrector_kwargs)
+
+            return e_t
+
+        alphas = self.model.alphas_cumprod if use_original_steps else self.ddim_alphas
+        alphas_prev = self.model.alphas_cumprod_prev if use_original_steps else self.ddim_alphas_prev
+        sqrt_one_minus_alphas = self.model.sqrt_one_minus_alphas_cumprod if use_original_steps else self.ddim_sqrt_one_minus_alphas
+        sigmas = self.model.ddim_sigmas_for_original_num_steps if use_original_steps else self.ddim_sigmas
+
+        def get_x_prev_and_pred_x0(e_t, index):
+            # select parameters corresponding to the currently considered timestep
+            a_t = torch.full((b, 1, 1, 1), alphas[index], device=device)
+            a_prev = torch.full((b, 1, 1, 1), alphas_prev[index], device=device)
+            sigma_t = torch.full((b, 1, 1, 1), sigmas[index], device=device)
+            sqrt_one_minus_at = torch.full((b, 1, 1, 1), sqrt_one_minus_alphas[index],device=device)
+
+            # current prediction for x_0
+            pred_x0 = (x - sqrt_one_minus_at * e_t) / a_t.sqrt()
+            if quantize_denoised:
+                pred_x0, _, *_ = self.model.first_stage_model.quantize(pred_x0)
+            if dynamic_threshold is not None:
+                pred_x0 = norm_thresholding(pred_x0, dynamic_threshold)
+            # direction pointing to x_t
+            dir_xt = (1. - a_prev - sigma_t**2).sqrt() * e_t
+            noise = sigma_t * noise_like(x.shape, device, repeat_noise) * temperature
+            if noise_dropout > 0.:
+                noise = torch.nn.functional.dropout(noise, p=noise_dropout)
+            x_prev = a_prev.sqrt() * pred_x0 + dir_xt + noise
+            return x_prev, pred_x0
+
+        e_t = get_model_output(x, t)
+        if len(old_eps) == 0:
+            # Pseudo Improved Euler (2nd order)
+            x_prev, pred_x0 = get_x_prev_and_pred_x0(e_t, index)
+            e_t_next = get_model_output(x_prev, t_next)
+            e_t_prime = (e_t + e_t_next) / 2
+        elif len(old_eps) == 1:
+            # 2nd order Pseudo Linear Multistep (Adams-Bashforth)
+            e_t_prime = (3 * e_t - old_eps[-1]) / 2
+        elif len(old_eps) == 2:
+            # 3nd order Pseudo Linear Multistep (Adams-Bashforth)
+            e_t_prime = (23 * e_t - 16 * old_eps[-1] + 5 * old_eps[-2]) / 12
+        elif len(old_eps) >= 3:
+            # 4nd order Pseudo Linear Multistep (Adams-Bashforth)
+            e_t_prime = (55 * e_t - 59 * old_eps[-1] + 37 * old_eps[-2] - 9 * old_eps[-3]) / 24
+
+        x_prev, pred_x0 = get_x_prev_and_pred_x0(e_t_prime, index)
+
+        return x_prev, pred_x0, e_t

ldm/models/diffusion/sampling_util.py

@@ -0,0 +1,50 @@
+import torch
+import numpy as np
+
+
+def append_dims(x, target_dims):
+    """Appends dimensions to the end of a tensor until it has target_dims dimensions.
+    From https://github.com/crowsonkb/k-diffusion/blob/master/k_diffusion/utils.py"""
+    dims_to_append = target_dims - x.ndim
+    if dims_to_append < 0:
+        raise ValueError(f'input has {x.ndim} dims but target_dims is {target_dims}, which is less')
+    return x[(...,) + (None,) * dims_to_append]
+
+
+def renorm_thresholding(x0, value):
+    # renorm
+    pred_max = x0.max()
+    pred_min = x0.min()
+    pred_x0 = (x0 - pred_min) / (pred_max - pred_min)  # 0 ... 1
+    pred_x0 = 2 * pred_x0 - 1.  # -1 ... 1
+
+    s = torch.quantile(
+        rearrange(pred_x0, 'b ... -> b (...)').abs(),
+        value,
+        dim=-1
+    )
+    s.clamp_(min=1.0)
+    s = s.view(-1, *((1,) * (pred_x0.ndim - 1)))
+
+    # clip by threshold
+    # pred_x0 = pred_x0.clamp(-s, s) / s  # needs newer pytorch  # TODO bring back to pure-gpu with min/max
+
+    # temporary hack: numpy on cpu
+    pred_x0 = np.clip(pred_x0.cpu().numpy(), -s.cpu().numpy(), s.cpu().numpy()) / s.cpu().numpy()
+    pred_x0 = torch.tensor(pred_x0).to(self.model.device)
+
+    # re.renorm
+    pred_x0 = (pred_x0 + 1.) / 2.  # 0 ... 1
+    pred_x0 = (pred_max - pred_min) * pred_x0 + pred_min  # orig range
+    return pred_x0
+
+
+def norm_thresholding(x0, value):
+    s = append_dims(x0.pow(2).flatten(1).mean(1).sqrt().clamp(min=value), x0.ndim)
+    return x0 * (value / s)
+
+
+def spatial_norm_thresholding(x0, value):
+    # b c h w
+    s = x0.pow(2).mean(1, keepdim=True).sqrt().clamp(min=value)
+    return x0 * (value / s)

ldm/modules/attention.py

@@ -0,0 +1,266 @@
+from inspect import isfunction
+import math
+import torch
+import torch.nn.functional as F
+from torch import nn, einsum
+from einops import rearrange, repeat
+
+from ldm.modules.diffusionmodules.util import checkpoint
+
+
+def exists(val):
+    return val is not None
+
+
+def uniq(arr):
+    return{el: True for el in arr}.keys()
+
+
+def default(val, d):
+    if exists(val):
+        return val
+    return d() if isfunction(d) else d
+
+
+def max_neg_value(t):
+    return -torch.finfo(t.dtype).max
+
+
+def init_(tensor):
+    dim = tensor.shape[-1]
+    std = 1 / math.sqrt(dim)
+    tensor.uniform_(-std, std)
+    return tensor
+
+
+# feedforward
+class GEGLU(nn.Module):
+    def __init__(self, dim_in, dim_out):
+        super().__init__()
+        self.proj = nn.Linear(dim_in, dim_out * 2)
+
+    def forward(self, x):
+        x, gate = self.proj(x).chunk(2, dim=-1)
+        return x * F.gelu(gate)
+
+
+class FeedForward(nn.Module):
+    def __init__(self, dim, dim_out=None, mult=4, glu=False, dropout=0.):
+        super().__init__()
+        inner_dim = int(dim * mult)
+        dim_out = default(dim_out, dim)
+        project_in = nn.Sequential(
+            nn.Linear(dim, inner_dim),
+            nn.GELU()
+        ) if not glu else GEGLU(dim, inner_dim)
+
+        self.net = nn.Sequential(
+            project_in,
+            nn.Dropout(dropout),
+            nn.Linear(inner_dim, dim_out)
+        )
+
+    def forward(self, x):
+        return self.net(x)
+
+
+def zero_module(module):
+    """
+    Zero out the parameters of a module and return it.
+    """
+    for p in module.parameters():
+        p.detach().zero_()
+    return module
+
+
+def Normalize(in_channels):
+    return torch.nn.GroupNorm(num_groups=32, num_channels=in_channels, eps=1e-6, affine=True)
+
+
+class LinearAttention(nn.Module):
+    def __init__(self, dim, heads=4, dim_head=32):
+        super().__init__()
+        self.heads = heads
+        hidden_dim = dim_head * heads
+        self.to_qkv = nn.Conv2d(dim, hidden_dim * 3, 1, bias = False)
+        self.to_out = nn.Conv2d(hidden_dim, dim, 1)
+
+    def forward(self, x):
+        b, c, h, w = x.shape
+        qkv = self.to_qkv(x)
+        q, k, v = rearrange(qkv, 'b (qkv heads c) h w -> qkv b heads c (h w)', heads = self.heads, qkv=3)
+        k = k.softmax(dim=-1)  
+        context = torch.einsum('bhdn,bhen->bhde', k, v)
+        out = torch.einsum('bhde,bhdn->bhen', context, q)
+        out = rearrange(out, 'b heads c (h w) -> b (heads c) h w', heads=self.heads, h=h, w=w)
+        return self.to_out(out)
+
+
+class SpatialSelfAttention(nn.Module):
+    def __init__(self, in_channels):
+        super().__init__()
+        self.in_channels = in_channels
+
+        self.norm = Normalize(in_channels)
+        self.q = torch.nn.Conv2d(in_channels,
+                                 in_channels,
+                                 kernel_size=1,
+                                 stride=1,
+                                 padding=0)
+        self.k = torch.nn.Conv2d(in_channels,
+                                 in_channels,
+                                 kernel_size=1,
+                                 stride=1,
+                                 padding=0)
+        self.v = torch.nn.Conv2d(in_channels,
+                                 in_channels,
+                                 kernel_size=1,
+                                 stride=1,
+                                 padding=0)
+        self.proj_out = torch.nn.Conv2d(in_channels,
+                                        in_channels,
+                                        kernel_size=1,
+                                        stride=1,
+                                        padding=0)
+
+    def forward(self, x):
+        h_ = x
+        h_ = self.norm(h_)
+        q = self.q(h_)
+        k = self.k(h_)
+        v = self.v(h_)
+
+        # compute attention
+        b,c,h,w = q.shape
+        q = rearrange(q, 'b c h w -> b (h w) c')
+        k = rearrange(k, 'b c h w -> b c (h w)')
+        w_ = torch.einsum('bij,bjk->bik', q, k)
+
+        w_ = w_ * (int(c)**(-0.5))
+        w_ = torch.nn.functional.softmax(w_, dim=2)
+
+        # attend to values
+        v = rearrange(v, 'b c h w -> b c (h w)')
+        w_ = rearrange(w_, 'b i j -> b j i')
+        h_ = torch.einsum('bij,bjk->bik', v, w_)
+        h_ = rearrange(h_, 'b c (h w) -> b c h w', h=h)
+        h_ = self.proj_out(h_)
+
+        return x+h_
+
+
+class CrossAttention(nn.Module):
+    def __init__(self, query_dim, context_dim=None, heads=8, dim_head=64, dropout=0.):
+        super().__init__()
+        inner_dim = dim_head * heads
+        context_dim = default(context_dim, query_dim)
+
+        self.scale = dim_head ** -0.5
+        self.heads = heads
+
+        self.to_q = nn.Linear(query_dim, inner_dim, bias=False)
+        self.to_k = nn.Linear(context_dim, inner_dim, bias=False)
+        self.to_v = nn.Linear(context_dim, inner_dim, bias=False)
+
+        self.to_out = nn.Sequential(
+            nn.Linear(inner_dim, query_dim),
+            nn.Dropout(dropout)
+        )
+
+    def forward(self, x, context=None, mask=None):
+        h = self.heads
+
+        q = self.to_q(x)
+        context = default(context, x)
+        k = self.to_k(context)
+        v = self.to_v(context)
+
+        q, k, v = map(lambda t: rearrange(t, 'b n (h d) -> (b h) n d', h=h), (q, k, v))
+
+        sim = einsum('b i d, b j d -> b i j', q, k) * self.scale
+
+        if exists(mask):
+            mask = rearrange(mask, 'b ... -> b (...)')
+            max_neg_value = -torch.finfo(sim.dtype).max
+            mask = repeat(mask, 'b j -> (b h) () j', h=h)
+            sim.masked_fill_(~mask, max_neg_value)
+
+        # attention, what we cannot get enough of
+        attn = sim.softmax(dim=-1)
+
+        out = einsum('b i j, b j d -> b i d', attn, v)
+        out = rearrange(out, '(b h) n d -> b n (h d)', h=h)
+        return self.to_out(out)
+
+
+class BasicTransformerBlock(nn.Module):
+    def __init__(self, dim, n_heads, d_head, dropout=0., context_dim=None, gated_ff=True, checkpoint=True,
+                 disable_self_attn=False):
+        super().__init__()
+        self.disable_self_attn = disable_self_attn
+        self.attn1 = CrossAttention(query_dim=dim, heads=n_heads, dim_head=d_head, dropout=dropout,
+                                    context_dim=context_dim if self.disable_self_attn else None)  # is a self-attention if not self.disable_self_attn
+        self.ff = FeedForward(dim, dropout=dropout, glu=gated_ff)
+        self.attn2 = CrossAttention(query_dim=dim, context_dim=context_dim,
+                                    heads=n_heads, dim_head=d_head, dropout=dropout)  # is self-attn if context is none
+        self.norm1 = nn.LayerNorm(dim)
+        self.norm2 = nn.LayerNorm(dim)
+        self.norm3 = nn.LayerNorm(dim)
+        self.checkpoint = checkpoint
+
+    def forward(self, x, context=None):
+        return checkpoint(self._forward, (x, context), self.parameters(), self.checkpoint)
+
+    def _forward(self, x, context=None):
+        x = self.attn1(self.norm1(x), context=context if self.disable_self_attn else None) + x
+        x = self.attn2(self.norm2(x), context=context) + x
+        x = self.ff(self.norm3(x)) + x
+        return x
+
+
+class SpatialTransformer(nn.Module):
+    """
+    Transformer block for image-like data.
+    First, project the input (aka embedding)
+    and reshape to b, t, d.
+    Then apply standard transformer action.
+    Finally, reshape to image
+    """
+    def __init__(self, in_channels, n_heads, d_head,
+                 depth=1, dropout=0., context_dim=None,
+                 disable_self_attn=False):
+        super().__init__()
+        self.in_channels = in_channels
+        inner_dim = n_heads * d_head
+        self.norm = Normalize(in_channels)
+
+        self.proj_in = nn.Conv2d(in_channels,
+                                 inner_dim,
+                                 kernel_size=1,
+                                 stride=1,
+                                 padding=0)
+
+        self.transformer_blocks = nn.ModuleList(
+            [BasicTransformerBlock(inner_dim, n_heads, d_head, dropout=dropout, context_dim=context_dim,
+                                   disable_self_attn=disable_self_attn)
+                for d in range(depth)]
+        )
+
+        self.proj_out = zero_module(nn.Conv2d(inner_dim,
+                                              in_channels,
+                                              kernel_size=1,
+                                              stride=1,
+                                              padding=0))
+
+    def forward(self, x, context=None):
+        # note: if no context is given, cross-attention defaults to self-attention
+        b, c, h, w = x.shape
+        x_in = x
+        x = self.norm(x)
+        x = self.proj_in(x)
+        x = rearrange(x, 'b c h w -> b (h w) c').contiguous()
+        for block in self.transformer_blocks:
+            x = block(x, context=context)
+        x = rearrange(x, 'b (h w) c -> b c h w', h=h, w=w).contiguous()
+        x = self.proj_out(x)
+        return x + x_in

ldm/modules/diffusionmodules/model.py

@@ -0,0 +1,835 @@
+# pytorch_diffusion + derived encoder decoder
+import math
+import torch
+import torch.nn as nn
+import numpy as np
+from einops import rearrange
+
+from ldm.util import instantiate_from_config
+from ldm.modules.attention import LinearAttention
+
+
+def get_timestep_embedding(timesteps, embedding_dim):
+    """
+    This matches the implementation in Denoising Diffusion Probabilistic Models:
+    From Fairseq.
+    Build sinusoidal embeddings.
+    This matches the implementation in tensor2tensor, but differs slightly
+    from the description in Section 3.5 of "Attention Is All You Need".
+    """
+    assert len(timesteps.shape) == 1
+
+    half_dim = embedding_dim // 2
+    emb = math.log(10000) / (half_dim - 1)
+    emb = torch.exp(torch.arange(half_dim, dtype=torch.float32) * -emb)
+    emb = emb.to(device=timesteps.device)
+    emb = timesteps.float()[:, None] * emb[None, :]
+    emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1)
+    if embedding_dim % 2 == 1:  # zero pad
+        emb = torch.nn.functional.pad(emb, (0,1,0,0))
+    return emb
+
+
+def nonlinearity(x):
+    # swish
+    return x*torch.sigmoid(x)
+
+
+def Normalize(in_channels, num_groups=32):
+    return torch.nn.GroupNorm(num_groups=num_groups, num_channels=in_channels, eps=1e-6, affine=True)
+
+
+class Upsample(nn.Module):
+    def __init__(self, in_channels, with_conv):
+        super().__init__()
+        self.with_conv = with_conv
+        if self.with_conv:
+            self.conv = torch.nn.Conv2d(in_channels,
+                                        in_channels,
+                                        kernel_size=3,
+                                        stride=1,
+                                        padding=1)
+
+    def forward(self, x):
+        x = torch.nn.functional.interpolate(x, scale_factor=2.0, mode="nearest")
+        if self.with_conv:
+            x = self.conv(x)
+        return x
+
+
+class Downsample(nn.Module):
+    def __init__(self, in_channels, with_conv):
+        super().__init__()
+        self.with_conv = with_conv
+        if self.with_conv:
+            # no asymmetric padding in torch conv, must do it ourselves
+            self.conv = torch.nn.Conv2d(in_channels,
+                                        in_channels,
+                                        kernel_size=3,
+                                        stride=2,
+                                        padding=0)
+
+    def forward(self, x):
+        if self.with_conv:
+            pad = (0,1,0,1)
+            x = torch.nn.functional.pad(x, pad, mode="constant", value=0)
+            x = self.conv(x)
+        else:
+            x = torch.nn.functional.avg_pool2d(x, kernel_size=2, stride=2)
+        return x
+
+
+class ResnetBlock(nn.Module):
+    def __init__(self, *, in_channels, out_channels=None, conv_shortcut=False,
+                 dropout, temb_channels=512):
+        super().__init__()
+        self.in_channels = in_channels
+        out_channels = in_channels if out_channels is None else out_channels
+        self.out_channels = out_channels
+        self.use_conv_shortcut = conv_shortcut
+
+        self.norm1 = Normalize(in_channels)
+        self.conv1 = torch.nn.Conv2d(in_channels,
+                                     out_channels,
+                                     kernel_size=3,
+                                     stride=1,
+                                     padding=1)
+        if temb_channels > 0:
+            self.temb_proj = torch.nn.Linear(temb_channels,
+                                             out_channels)
+        self.norm2 = Normalize(out_channels)
+        self.dropout = torch.nn.Dropout(dropout)
+        self.conv2 = torch.nn.Conv2d(out_channels,
+                                     out_channels,
+                                     kernel_size=3,
+                                     stride=1,
+                                     padding=1)
+        if self.in_channels != self.out_channels:
+            if self.use_conv_shortcut:
+                self.conv_shortcut = torch.nn.Conv2d(in_channels,
+                                                     out_channels,
+                                                     kernel_size=3,
+                                                     stride=1,
+                                                     padding=1)
+            else:
+                self.nin_shortcut = torch.nn.Conv2d(in_channels,
+                                                    out_channels,
+                                                    kernel_size=1,
+                                                    stride=1,
+                                                    padding=0)
+
+    def forward(self, x, temb):
+        h = x
+        h = self.norm1(h)
+        h = nonlinearity(h)
+        h = self.conv1(h)
+
+        if temb is not None:
+            h = h + self.temb_proj(nonlinearity(temb))[:,:,None,None]
+
+        h = self.norm2(h)
+        h = nonlinearity(h)
+        h = self.dropout(h)
+        h = self.conv2(h)
+
+        if self.in_channels != self.out_channels:
+            if self.use_conv_shortcut:
+                x = self.conv_shortcut(x)
+            else:
+                x = self.nin_shortcut(x)
+
+        return x+h
+
+
+class LinAttnBlock(LinearAttention):
+    """to match AttnBlock usage"""
+    def __init__(self, in_channels):
+        super().__init__(dim=in_channels, heads=1, dim_head=in_channels)
+
+
+class AttnBlock(nn.Module):
+    def __init__(self, in_channels):
+        super().__init__()
+        self.in_channels = in_channels
+
+        self.norm = Normalize(in_channels)
+        self.q = torch.nn.Conv2d(in_channels,
+                                 in_channels,
+                                 kernel_size=1,
+                                 stride=1,
+                                 padding=0)
+        self.k = torch.nn.Conv2d(in_channels,
+                                 in_channels,
+                                 kernel_size=1,
+                                 stride=1,
+                                 padding=0)
+        self.v = torch.nn.Conv2d(in_channels,
+                                 in_channels,
+                                 kernel_size=1,
+                                 stride=1,
+                                 padding=0)
+        self.proj_out = torch.nn.Conv2d(in_channels,
+                                        in_channels,
+                                        kernel_size=1,
+                                        stride=1,
+                                        padding=0)
+
+
+    def forward(self, x):
+        h_ = x
+        h_ = self.norm(h_)
+        q = self.q(h_)
+        k = self.k(h_)
+        v = self.v(h_)
+
+        # compute attention
+        b,c,h,w = q.shape
+        q = q.reshape(b,c,h*w)
+        q = q.permute(0,2,1)   # b,hw,c
+        k = k.reshape(b,c,h*w) # b,c,hw
+        w_ = torch.bmm(q,k)     # b,hw,hw    w[b,i,j]=sum_c q[b,i,c]k[b,c,j]
+        w_ = w_ * (int(c)**(-0.5))
+        w_ = torch.nn.functional.softmax(w_, dim=2)
+
+        # attend to values
+        v = v.reshape(b,c,h*w)
+        w_ = w_.permute(0,2,1)   # b,hw,hw (first hw of k, second of q)
+        h_ = torch.bmm(v,w_)     # b, c,hw (hw of q) h_[b,c,j] = sum_i v[b,c,i] w_[b,i,j]
+        h_ = h_.reshape(b,c,h,w)
+
+        h_ = self.proj_out(h_)
+
+        return x+h_
+
+
+def make_attn(in_channels, attn_type="vanilla"):
+    assert attn_type in ["vanilla", "linear", "none"], f'attn_type {attn_type} unknown'
+    print(f"making attention of type '{attn_type}' with {in_channels} in_channels")
+    if attn_type == "vanilla":
+        return AttnBlock(in_channels)
+    elif attn_type == "none":
+        return nn.Identity(in_channels)
+    else:
+        return LinAttnBlock(in_channels)
+
+
+class Model(nn.Module):
+    def __init__(self, *, ch, out_ch, ch_mult=(1,2,4,8), num_res_blocks,
+                 attn_resolutions, dropout=0.0, resamp_with_conv=True, in_channels,
+                 resolution, use_timestep=True, use_linear_attn=False, attn_type="vanilla"):
+        super().__init__()
+        if use_linear_attn: attn_type = "linear"
+        self.ch = ch
+        self.temb_ch = self.ch*4
+        self.num_resolutions = len(ch_mult)
+        self.num_res_blocks = num_res_blocks
+        self.resolution = resolution
+        self.in_channels = in_channels
+
+        self.use_timestep = use_timestep
+        if self.use_timestep:
+            # timestep embedding
+            self.temb = nn.Module()
+            self.temb.dense = nn.ModuleList([
+                torch.nn.Linear(self.ch,
+                                self.temb_ch),
+                torch.nn.Linear(self.temb_ch,
+                                self.temb_ch),
+            ])
+
+        # downsampling
+        self.conv_in = torch.nn.Conv2d(in_channels,
+                                       self.ch,
+                                       kernel_size=3,
+                                       stride=1,
+                                       padding=1)
+
+        curr_res = resolution
+        in_ch_mult = (1,)+tuple(ch_mult)
+        self.down = nn.ModuleList()
+        for i_level in range(self.num_resolutions):
+            block = nn.ModuleList()
+            attn = nn.ModuleList()
+            block_in = ch*in_ch_mult[i_level]
+            block_out = ch*ch_mult[i_level]
+            for i_block in range(self.num_res_blocks):
+                block.append(ResnetBlock(in_channels=block_in,
+                                         out_channels=block_out,
+                                         temb_channels=self.temb_ch,
+                                         dropout=dropout))
+                block_in = block_out
+                if curr_res in attn_resolutions:
+                    attn.append(make_attn(block_in, attn_type=attn_type))
+            down = nn.Module()
+            down.block = block
+            down.attn = attn
+            if i_level != self.num_resolutions-1:
+                down.downsample = Downsample(block_in, resamp_with_conv)
+                curr_res = curr_res // 2
+            self.down.append(down)
+
+        # middle
+        self.mid = nn.Module()
+        self.mid.block_1 = ResnetBlock(in_channels=block_in,
+                                       out_channels=block_in,
+                                       temb_channels=self.temb_ch,
+                                       dropout=dropout)
+        self.mid.attn_1 = make_attn(block_in, attn_type=attn_type)
+        self.mid.block_2 = ResnetBlock(in_channels=block_in,
+                                       out_channels=block_in,
+                                       temb_channels=self.temb_ch,
+                                       dropout=dropout)
+
+        # upsampling
+        self.up = nn.ModuleList()
+        for i_level in reversed(range(self.num_resolutions)):
+            block = nn.ModuleList()
+            attn = nn.ModuleList()
+            block_out = ch*ch_mult[i_level]
+            skip_in = ch*ch_mult[i_level]
+            for i_block in range(self.num_res_blocks+1):
+                if i_block == self.num_res_blocks:
+                    skip_in = ch*in_ch_mult[i_level]
+                block.append(ResnetBlock(in_channels=block_in+skip_in,
+                                         out_channels=block_out,
+                                         temb_channels=self.temb_ch,
+                                         dropout=dropout))
+                block_in = block_out
+                if curr_res in attn_resolutions:
+                    attn.append(make_attn(block_in, attn_type=attn_type))
+            up = nn.Module()
+            up.block = block
+            up.attn = attn
+            if i_level != 0:
+                up.upsample = Upsample(block_in, resamp_with_conv)
+                curr_res = curr_res * 2
+            self.up.insert(0, up) # prepend to get consistent order
+
+        # end
+        self.norm_out = Normalize(block_in)
+        self.conv_out = torch.nn.Conv2d(block_in,
+                                        out_ch,
+                                        kernel_size=3,
+                                        stride=1,
+                                        padding=1)
+
+    def forward(self, x, t=None, context=None):
+        #assert x.shape[2] == x.shape[3] == self.resolution
+        if context is not None:
+            # assume aligned context, cat along channel axis
+            x = torch.cat((x, context), dim=1)
+        if self.use_timestep:
+            # timestep embedding
+            assert t is not None
+            temb = get_timestep_embedding(t, self.ch)
+            temb = self.temb.dense[0](temb)
+            temb = nonlinearity(temb)
+            temb = self.temb.dense[1](temb)
+        else:
+            temb = None
+
+        # downsampling
+        hs = [self.conv_in(x)]
+        for i_level in range(self.num_resolutions):
+            for i_block in range(self.num_res_blocks):
+                h = self.down[i_level].block[i_block](hs[-1], temb)
+                if len(self.down[i_level].attn) > 0:
+                    h = self.down[i_level].attn[i_block](h)
+                hs.append(h)
+            if i_level != self.num_resolutions-1:
+                hs.append(self.down[i_level].downsample(hs[-1]))
+
+        # middle
+        h = hs[-1]
+        h = self.mid.block_1(h, temb)
+        h = self.mid.attn_1(h)
+        h = self.mid.block_2(h, temb)
+
+        # upsampling
+        for i_level in reversed(range(self.num_resolutions)):
+            for i_block in range(self.num_res_blocks+1):
+                h = self.up[i_level].block[i_block](
+                    torch.cat([h, hs.pop()], dim=1), temb)
+                if len(self.up[i_level].attn) > 0:
+                    h = self.up[i_level].attn[i_block](h)
+            if i_level != 0:
+                h = self.up[i_level].upsample(h)
+
+        # end
+        h = self.norm_out(h)
+        h = nonlinearity(h)
+        h = self.conv_out(h)
+        return h
+
+    def get_last_layer(self):
+        return self.conv_out.weight
+
+
+class Encoder(nn.Module):
+    def __init__(self, *, ch, out_ch, ch_mult=(1,2,4,8), num_res_blocks,
+                 attn_resolutions, dropout=0.0, resamp_with_conv=True, in_channels,
+                 resolution, z_channels, double_z=True, use_linear_attn=False, attn_type="vanilla",
+                 **ignore_kwargs):
+        super().__init__()
+        if use_linear_attn: attn_type = "linear"
+        self.ch = ch
+        self.temb_ch = 0
+        self.num_resolutions = len(ch_mult)
+        self.num_res_blocks = num_res_blocks
+        self.resolution = resolution
+        self.in_channels = in_channels
+
+        # downsampling
+        self.conv_in = torch.nn.Conv2d(in_channels,
+                                       self.ch,
+                                       kernel_size=3,
+                                       stride=1,
+                                       padding=1)
+
+        curr_res = resolution
+        in_ch_mult = (1,)+tuple(ch_mult)
+        self.in_ch_mult = in_ch_mult
+        self.down = nn.ModuleList()
+        for i_level in range(self.num_resolutions):
+            block = nn.ModuleList()
+            attn = nn.ModuleList()
+            block_in = ch*in_ch_mult[i_level]
+            block_out = ch*ch_mult[i_level]
+            for i_block in range(self.num_res_blocks):
+                block.append(ResnetBlock(in_channels=block_in,
+                                         out_channels=block_out,
+                                         temb_channels=self.temb_ch,
+                                         dropout=dropout))
+                block_in = block_out
+                if curr_res in attn_resolutions:
+                    attn.append(make_attn(block_in, attn_type=attn_type))
+            down = nn.Module()
+            down.block = block
+            down.attn = attn
+            if i_level != self.num_resolutions-1:
+                down.downsample = Downsample(block_in, resamp_with_conv)
+                curr_res = curr_res // 2
+            self.down.append(down)
+
+        # middle
+        self.mid = nn.Module()
+        self.mid.block_1 = ResnetBlock(in_channels=block_in,
+                                       out_channels=block_in,
+                                       temb_channels=self.temb_ch,
+                                       dropout=dropout)
+        self.mid.attn_1 = make_attn(block_in, attn_type=attn_type)
+        self.mid.block_2 = ResnetBlock(in_channels=block_in,
+                                       out_channels=block_in,
+                                       temb_channels=self.temb_ch,
+                                       dropout=dropout)
+
+        # end
+        self.norm_out = Normalize(block_in)
+        self.conv_out = torch.nn.Conv2d(block_in,
+                                        2*z_channels if double_z else z_channels,
+                                        kernel_size=3,
+                                        stride=1,
+                                        padding=1)
+
+    def forward(self, x):
+        # timestep embedding
+        temb = None
+
+        # downsampling
+        hs = [self.conv_in(x)]
+        for i_level in range(self.num_resolutions):
+            for i_block in range(self.num_res_blocks):
+                h = self.down[i_level].block[i_block](hs[-1], temb)
+                if len(self.down[i_level].attn) > 0:
+                    h = self.down[i_level].attn[i_block](h)
+                hs.append(h)
+            if i_level != self.num_resolutions-1:
+                hs.append(self.down[i_level].downsample(hs[-1]))
+
+        # middle
+        h = hs[-1]
+        h = self.mid.block_1(h, temb)
+        h = self.mid.attn_1(h)
+        h = self.mid.block_2(h, temb)
+
+        # end
+        h = self.norm_out(h)
+        h = nonlinearity(h)
+        h = self.conv_out(h)
+        return h
+
+
+class Decoder(nn.Module):
+    def __init__(self, *, ch, out_ch, ch_mult=(1,2,4,8), num_res_blocks,
+                 attn_resolutions, dropout=0.0, resamp_with_conv=True, in_channels,
+                 resolution, z_channels, give_pre_end=False, tanh_out=False, use_linear_attn=False,
+                 attn_type="vanilla", **ignorekwargs):
+        super().__init__()
+        if use_linear_attn: attn_type = "linear"
+        self.ch = ch
+        self.temb_ch = 0
+        self.num_resolutions = len(ch_mult)
+        self.num_res_blocks = num_res_blocks
+        self.resolution = resolution
+        self.in_channels = in_channels
+        self.give_pre_end = give_pre_end
+        self.tanh_out = tanh_out
+
+        # compute in_ch_mult, block_in and curr_res at lowest res
+        in_ch_mult = (1,)+tuple(ch_mult)
+        block_in = ch*ch_mult[self.num_resolutions-1]
+        curr_res = resolution // 2**(self.num_resolutions-1)
+        self.z_shape = (1,z_channels,curr_res,curr_res)
+        print("Working with z of shape {} = {} dimensions.".format(
+            self.z_shape, np.prod(self.z_shape)))
+
+        # z to block_in
+        self.conv_in = torch.nn.Conv2d(z_channels,
+                                       block_in,
+                                       kernel_size=3,
+                                       stride=1,
+                                       padding=1)
+
+        # middle
+        self.mid = nn.Module()
+        self.mid.block_1 = ResnetBlock(in_channels=block_in,
+                                       out_channels=block_in,
+                                       temb_channels=self.temb_ch,
+                                       dropout=dropout)
+        self.mid.attn_1 = make_attn(block_in, attn_type=attn_type)
+        self.mid.block_2 = ResnetBlock(in_channels=block_in,
+                                       out_channels=block_in,
+                                       temb_channels=self.temb_ch,
+                                       dropout=dropout)
+
+        # upsampling
+        self.up = nn.ModuleList()
+        for i_level in reversed(range(self.num_resolutions)):
+            block = nn.ModuleList()
+            attn = nn.ModuleList()
+            block_out = ch*ch_mult[i_level]
+            for i_block in range(self.num_res_blocks+1):
+                block.append(ResnetBlock(in_channels=block_in,
+                                         out_channels=block_out,
+                                         temb_channels=self.temb_ch,
+                                         dropout=dropout))
+                block_in = block_out
+                if curr_res in attn_resolutions:
+                    attn.append(make_attn(block_in, attn_type=attn_type))
+            up = nn.Module()
+            up.block = block
+            up.attn = attn
+            if i_level != 0:
+                up.upsample = Upsample(block_in, resamp_with_conv)
+                curr_res = curr_res * 2
+            self.up.insert(0, up) # prepend to get consistent order
+
+        # end
+        self.norm_out = Normalize(block_in)
+        self.conv_out = torch.nn.Conv2d(block_in,
+                                        out_ch,
+                                        kernel_size=3,
+                                        stride=1,
+                                        padding=1)
+
+    def forward(self, z):
+        #assert z.shape[1:] == self.z_shape[1:]
+        self.last_z_shape = z.shape
+
+        # timestep embedding
+        temb = None
+
+        # z to block_in
+        h = self.conv_in(z)
+
+        # middle
+        h = self.mid.block_1(h, temb)
+        h = self.mid.attn_1(h)
+        h = self.mid.block_2(h, temb)
+
+        # upsampling
+        for i_level in reversed(range(self.num_resolutions)):
+            for i_block in range(self.num_res_blocks+1):
+                h = self.up[i_level].block[i_block](h, temb)
+                if len(self.up[i_level].attn) > 0:
+                    h = self.up[i_level].attn[i_block](h)
+            if i_level != 0:
+                h = self.up[i_level].upsample(h)
+
+        # end
+        if self.give_pre_end:
+            return h
+
+        h = self.norm_out(h)
+        h = nonlinearity(h)
+        h = self.conv_out(h)
+        if self.tanh_out:
+            h = torch.tanh(h)
+        return h
+
+
+class SimpleDecoder(nn.Module):
+    def __init__(self, in_channels, out_channels, *args, **kwargs):
+        super().__init__()
+        self.model = nn.ModuleList([nn.Conv2d(in_channels, in_channels, 1),
+                                     ResnetBlock(in_channels=in_channels,
+                                                 out_channels=2 * in_channels,
+                                                 temb_channels=0, dropout=0.0),
+                                     ResnetBlock(in_channels=2 * in_channels,
+                                                out_channels=4 * in_channels,
+                                                temb_channels=0, dropout=0.0),
+                                     ResnetBlock(in_channels=4 * in_channels,
+                                                out_channels=2 * in_channels,
+                                                temb_channels=0, dropout=0.0),
+                                     nn.Conv2d(2*in_channels, in_channels, 1),
+                                     Upsample(in_channels, with_conv=True)])
+        # end
+        self.norm_out = Normalize(in_channels)
+        self.conv_out = torch.nn.Conv2d(in_channels,
+                                        out_channels,
+                                        kernel_size=3,
+                                        stride=1,
+                                        padding=1)
+
+    def forward(self, x):
+        for i, layer in enumerate(self.model):
+            if i in [1,2,3]:
+                x = layer(x, None)
+            else:
+                x = layer(x)
+
+        h = self.norm_out(x)
+        h = nonlinearity(h)
+        x = self.conv_out(h)
+        return x
+
+
+class UpsampleDecoder(nn.Module):
+    def __init__(self, in_channels, out_channels, ch, num_res_blocks, resolution,
+                 ch_mult=(2,2), dropout=0.0):
+        super().__init__()
+        # upsampling
+        self.temb_ch = 0
+        self.num_resolutions = len(ch_mult)
+        self.num_res_blocks = num_res_blocks
+        block_in = in_channels
+        curr_res = resolution // 2 ** (self.num_resolutions - 1)
+        self.res_blocks = nn.ModuleList()
+        self.upsample_blocks = nn.ModuleList()
+        for i_level in range(self.num_resolutions):
+            res_block = []
+            block_out = ch * ch_mult[i_level]
+            for i_block in range(self.num_res_blocks + 1):
+                res_block.append(ResnetBlock(in_channels=block_in,
+                                         out_channels=block_out,
+                                         temb_channels=self.temb_ch,
+                                         dropout=dropout))
+                block_in = block_out
+            self.res_blocks.append(nn.ModuleList(res_block))
+            if i_level != self.num_resolutions - 1:
+                self.upsample_blocks.append(Upsample(block_in, True))
+                curr_res = curr_res * 2
+
+        # end
+        self.norm_out = Normalize(block_in)
+        self.conv_out = torch.nn.Conv2d(block_in,
+                                        out_channels,
+                                        kernel_size=3,
+                                        stride=1,
+                                        padding=1)
+
+    def forward(self, x):
+        # upsampling
+        h = x
+        for k, i_level in enumerate(range(self.num_resolutions)):
+            for i_block in range(self.num_res_blocks + 1):
+                h = self.res_blocks[i_level][i_block](h, None)
+            if i_level != self.num_resolutions - 1:
+                h = self.upsample_blocks[k](h)
+        h = self.norm_out(h)
+        h = nonlinearity(h)
+        h = self.conv_out(h)
+        return h
+
+
+class LatentRescaler(nn.Module):
+    def __init__(self, factor, in_channels, mid_channels, out_channels, depth=2):
+        super().__init__()
+        # residual block, interpolate, residual block
+        self.factor = factor
+        self.conv_in = nn.Conv2d(in_channels,
+                                 mid_channels,
+                                 kernel_size=3,
+                                 stride=1,
+                                 padding=1)
+        self.res_block1 = nn.ModuleList([ResnetBlock(in_channels=mid_channels,
+                                                     out_channels=mid_channels,
+                                                     temb_channels=0,
+                                                     dropout=0.0) for _ in range(depth)])
+        self.attn = AttnBlock(mid_channels)
+        self.res_block2 = nn.ModuleList([ResnetBlock(in_channels=mid_channels,
+                                                     out_channels=mid_channels,
+                                                     temb_channels=0,
+                                                     dropout=0.0) for _ in range(depth)])
+
+        self.conv_out = nn.Conv2d(mid_channels,
+                                  out_channels,
+                                  kernel_size=1,
+                                  )
+
+    def forward(self, x):
+        x = self.conv_in(x)
+        for block in self.res_block1:
+            x = block(x, None)
+        x = torch.nn.functional.interpolate(x, size=(int(round(x.shape[2]*self.factor)), int(round(x.shape[3]*self.factor))))
+        x = self.attn(x)
+        for block in self.res_block2:
+            x = block(x, None)
+        x = self.conv_out(x)
+        return x
+
+
+class MergedRescaleEncoder(nn.Module):
+    def __init__(self, in_channels, ch, resolution, out_ch, num_res_blocks,
+                 attn_resolutions, dropout=0.0, resamp_with_conv=True,
+                 ch_mult=(1,2,4,8), rescale_factor=1.0, rescale_module_depth=1):
+        super().__init__()
+        intermediate_chn = ch * ch_mult[-1]
+        self.encoder = Encoder(in_channels=in_channels, num_res_blocks=num_res_blocks, ch=ch, ch_mult=ch_mult,
+                               z_channels=intermediate_chn, double_z=False, resolution=resolution,
+                               attn_resolutions=attn_resolutions, dropout=dropout, resamp_with_conv=resamp_with_conv,
+                               out_ch=None)
+        self.rescaler = LatentRescaler(factor=rescale_factor, in_channels=intermediate_chn,
+                                       mid_channels=intermediate_chn, out_channels=out_ch, depth=rescale_module_depth)
+
+    def forward(self, x):
+        x = self.encoder(x)
+        x = self.rescaler(x)
+        return x
+
+
+class MergedRescaleDecoder(nn.Module):
+    def __init__(self, z_channels, out_ch, resolution, num_res_blocks, attn_resolutions, ch, ch_mult=(1,2,4,8),
+                 dropout=0.0, resamp_with_conv=True, rescale_factor=1.0, rescale_module_depth=1):
+        super().__init__()
+        tmp_chn = z_channels*ch_mult[-1]
+        self.decoder = Decoder(out_ch=out_ch, z_channels=tmp_chn, attn_resolutions=attn_resolutions, dropout=dropout,
+                               resamp_with_conv=resamp_with_conv, in_channels=None, num_res_blocks=num_res_blocks,
+                               ch_mult=ch_mult, resolution=resolution, ch=ch)
+        self.rescaler = LatentRescaler(factor=rescale_factor, in_channels=z_channels, mid_channels=tmp_chn,
+                                       out_channels=tmp_chn, depth=rescale_module_depth)
+
+    def forward(self, x):
+        x = self.rescaler(x)
+        x = self.decoder(x)
+        return x
+
+
+class Upsampler(nn.Module):
+    def __init__(self, in_size, out_size, in_channels, out_channels, ch_mult=2):
+        super().__init__()
+        assert out_size >= in_size
+        num_blocks = int(np.log2(out_size//in_size))+1
+        factor_up = 1.+ (out_size % in_size)
+        print(f"Building {self.__class__.__name__} with in_size: {in_size} --> out_size {out_size} and factor {factor_up}")
+        self.rescaler = LatentRescaler(factor=factor_up, in_channels=in_channels, mid_channels=2*in_channels,
+                                       out_channels=in_channels)
+        self.decoder = Decoder(out_ch=out_channels, resolution=out_size, z_channels=in_channels, num_res_blocks=2,
+                               attn_resolutions=[], in_channels=None, ch=in_channels,
+                               ch_mult=[ch_mult for _ in range(num_blocks)])
+
+    def forward(self, x):
+        x = self.rescaler(x)
+        x = self.decoder(x)
+        return x
+
+
+class Resize(nn.Module):
+    def __init__(self, in_channels=None, learned=False, mode="bilinear"):
+        super().__init__()
+        self.with_conv = learned
+        self.mode = mode
+        if self.with_conv:
+            print(f"Note: {self.__class__.__name} uses learned downsampling and will ignore the fixed {mode} mode")
+            raise NotImplementedError()
+            assert in_channels is not None
+            # no asymmetric padding in torch conv, must do it ourselves
+            self.conv = torch.nn.Conv2d(in_channels,
+                                        in_channels,
+                                        kernel_size=4,
+                                        stride=2,
+                                        padding=1)
+
+    def forward(self, x, scale_factor=1.0):
+        if scale_factor==1.0:
+            return x
+        else:
+            x = torch.nn.functional.interpolate(x, mode=self.mode, align_corners=False, scale_factor=scale_factor)
+        return x
+
+class FirstStagePostProcessor(nn.Module):
+
+    def __init__(self, ch_mult:list, in_channels,
+                 pretrained_model:nn.Module=None,
+                 reshape=False,
+                 n_channels=None,
+                 dropout=0.,
+                 pretrained_config=None):
+        super().__init__()
+        if pretrained_config is None:
+            assert pretrained_model is not None, 'Either "pretrained_model" or "pretrained_config" must not be None'
+            self.pretrained_model = pretrained_model
+        else:
+            assert pretrained_config is not None, 'Either "pretrained_model" or "pretrained_config" must not be None'
+            self.instantiate_pretrained(pretrained_config)
+
+        self.do_reshape = reshape
+
+        if n_channels is None:
+            n_channels = self.pretrained_model.encoder.ch
+
+        self.proj_norm = Normalize(in_channels,num_groups=in_channels//2)
+        self.proj = nn.Conv2d(in_channels,n_channels,kernel_size=3,
+                            stride=1,padding=1)
+
+        blocks = []
+        downs = []
+        ch_in = n_channels
+        for m in ch_mult:
+            blocks.append(ResnetBlock(in_channels=ch_in,out_channels=m*n_channels,dropout=dropout))
+            ch_in = m * n_channels
+            downs.append(Downsample(ch_in, with_conv=False))
+
+        self.model = nn.ModuleList(blocks)
+        self.downsampler = nn.ModuleList(downs)
+
+
+    def instantiate_pretrained(self, config):
+        model = instantiate_from_config(config)
+        self.pretrained_model = model.eval()
+        # self.pretrained_model.train = False
+        for param in self.pretrained_model.parameters():
+            param.requires_grad = False
+
+
+    @torch.no_grad()
+    def encode_with_pretrained(self,x):
+        c = self.pretrained_model.encode(x)
+        if isinstance(c, DiagonalGaussianDistribution):
+            c = c.mode()
+        return  c
+
+    def forward(self,x):
+        z_fs = self.encode_with_pretrained(x)
+        z = self.proj_norm(z_fs)
+        z = self.proj(z)
+        z = nonlinearity(z)
+
+        for submodel, downmodel in zip(self.model,self.downsampler):
+            z = submodel(z,temb=None)
+            z = downmodel(z)
+
+        if self.do_reshape:
+            z = rearrange(z,'b c h w -> b (h w) c')
+        return z
+

ldm/modules/diffusionmodules/openaimodel.py

@@ -0,0 +1,996 @@
+from abc import abstractmethod
+from functools import partial
+import math
+from typing import Iterable
+
+import numpy as np
+import torch as th
+import torch.nn as nn
+import torch.nn.functional as F
+
+from ldm.modules.diffusionmodules.util import (
+    checkpoint,
+    conv_nd,
+    linear,
+    avg_pool_nd,
+    zero_module,
+    normalization,
+    timestep_embedding,
+)
+from ldm.modules.attention import SpatialTransformer
+from ldm.util import exists
+
+
+# dummy replace
+def convert_module_to_f16(x):
+    pass
+
+def convert_module_to_f32(x):
+    pass
+
+
+## go
+class AttentionPool2d(nn.Module):
+    """
+    Adapted from CLIP: https://github.com/openai/CLIP/blob/main/clip/model.py
+    """
+
+    def __init__(
+        self,
+        spacial_dim: int,
+        embed_dim: int,
+        num_heads_channels: int,
+        output_dim: int = None,
+    ):
+        super().__init__()
+        self.positional_embedding = nn.Parameter(th.randn(embed_dim, spacial_dim ** 2 + 1) / embed_dim ** 0.5)
+        self.qkv_proj = conv_nd(1, embed_dim, 3 * embed_dim, 1)
+        self.c_proj = conv_nd(1, embed_dim, output_dim or embed_dim, 1)
+        self.num_heads = embed_dim // num_heads_channels
+        self.attention = QKVAttention(self.num_heads)
+
+    def forward(self, x):
+        b, c, *_spatial = x.shape
+        x = x.reshape(b, c, -1)  # NC(HW)
+        x = th.cat([x.mean(dim=-1, keepdim=True), x], dim=-1)  # NC(HW+1)
+        x = x + self.positional_embedding[None, :, :].to(x.dtype)  # NC(HW+1)
+        x = self.qkv_proj(x)
+        x = self.attention(x)
+        x = self.c_proj(x)
+        return x[:, :, 0]
+
+
+class TimestepBlock(nn.Module):
+    """
+    Any module where forward() takes timestep embeddings as a second argument.
+    """
+
+    @abstractmethod
+    def forward(self, x, emb):
+        """
+        Apply the module to `x` given `emb` timestep embeddings.
+        """
+
+
+class TimestepEmbedSequential(nn.Sequential, TimestepBlock):
+    """
+    A sequential module that passes timestep embeddings to the children that
+    support it as an extra input.
+    """
+
+    def forward(self, x, emb, context=None):
+        for layer in self:
+            if isinstance(layer, TimestepBlock):
+                x = layer(x, emb)
+            elif isinstance(layer, SpatialTransformer):
+                x = layer(x, context)
+            else:
+                x = layer(x)
+        return x
+
+
+class Upsample(nn.Module):
+    """
+    An upsampling layer with an optional convolution.
+    :param channels: channels in the inputs and outputs.
+    :param use_conv: a bool determining if a convolution is applied.
+    :param dims: determines if the signal is 1D, 2D, or 3D. If 3D, then
+                 upsampling occurs in the inner-two dimensions.
+    """
+
+    def __init__(self, channels, use_conv, dims=2, out_channels=None, padding=1):
+        super().__init__()
+        self.channels = channels
+        self.out_channels = out_channels or channels
+        self.use_conv = use_conv
+        self.dims = dims
+        if use_conv:
+            self.conv = conv_nd(dims, self.channels, self.out_channels, 3, padding=padding)
+
+    def forward(self, x):
+        assert x.shape[1] == self.channels
+        if self.dims == 3:
+            x = F.interpolate(
+                x, (x.shape[2], x.shape[3] * 2, x.shape[4] * 2), mode="nearest"
+            )
+        else:
+            x = F.interpolate(x, scale_factor=2, mode="nearest")
+        if self.use_conv:
+            x = self.conv(x)
+        return x
+
+class TransposedUpsample(nn.Module):
+    'Learned 2x upsampling without padding'
+    def __init__(self, channels, out_channels=None, ks=5):
+        super().__init__()
+        self.channels = channels
+        self.out_channels = out_channels or channels
+
+        self.up = nn.ConvTranspose2d(self.channels,self.out_channels,kernel_size=ks,stride=2)
+
+    def forward(self,x):
+        return self.up(x)
+
+
+class Downsample(nn.Module):
+    """
+    A downsampling layer with an optional convolution.
+    :param channels: channels in the inputs and outputs.
+    :param use_conv: a bool determining if a convolution is applied.
+    :param dims: determines if the signal is 1D, 2D, or 3D. If 3D, then
+                 downsampling occurs in the inner-two dimensions.
+    """
+
+    def __init__(self, channels, use_conv, dims=2, out_channels=None,padding=1):
+        super().__init__()
+        self.channels = channels
+        self.out_channels = out_channels or channels
+        self.use_conv = use_conv
+        self.dims = dims
+        stride = 2 if dims != 3 else (1, 2, 2)
+        if use_conv:
+            self.op = conv_nd(
+                dims, self.channels, self.out_channels, 3, stride=stride, padding=padding
+            )
+        else:
+            assert self.channels == self.out_channels
+            self.op = avg_pool_nd(dims, kernel_size=stride, stride=stride)
+
+    def forward(self, x):
+        assert x.shape[1] == self.channels
+        return self.op(x)
+
+
+class ResBlock(TimestepBlock):
+    """
+    A residual block that can optionally change the number of channels.
+    :param channels: the number of input channels.
+    :param emb_channels: the number of timestep embedding channels.
+    :param dropout: the rate of dropout.
+    :param out_channels: if specified, the number of out channels.
+    :param use_conv: if True and out_channels is specified, use a spatial
+        convolution instead of a smaller 1x1 convolution to change the
+        channels in the skip connection.
+    :param dims: determines if the signal is 1D, 2D, or 3D.
+    :param use_checkpoint: if True, use gradient checkpointing on this module.
+    :param up: if True, use this block for upsampling.
+    :param down: if True, use this block for downsampling.
+    """
+
+    def __init__(
+        self,
+        channels,
+        emb_channels,
+        dropout,
+        out_channels=None,
+        use_conv=False,
+        use_scale_shift_norm=False,
+        dims=2,
+        use_checkpoint=False,
+        up=False,
+        down=False,
+    ):
+        super().__init__()
+        self.channels = channels
+        self.emb_channels = emb_channels
+        self.dropout = dropout
+        self.out_channels = out_channels or channels
+        self.use_conv = use_conv
+        self.use_checkpoint = use_checkpoint
+        self.use_scale_shift_norm = use_scale_shift_norm
+
+        self.in_layers = nn.Sequential(
+            normalization(channels),
+            nn.SiLU(),
+            conv_nd(dims, channels, self.out_channels, 3, padding=1),
+        )
+
+        self.updown = up or down
+
+        if up:
+            self.h_upd = Upsample(channels, False, dims)
+            self.x_upd = Upsample(channels, False, dims)
+        elif down:
+            self.h_upd = Downsample(channels, False, dims)
+            self.x_upd = Downsample(channels, False, dims)
+        else:
+            self.h_upd = self.x_upd = nn.Identity()
+
+        self.emb_layers = nn.Sequential(
+            nn.SiLU(),
+            linear(
+                emb_channels,
+                2 * self.out_channels if use_scale_shift_norm else self.out_channels,
+            ),
+        )
+        self.out_layers = nn.Sequential(
+            normalization(self.out_channels),
+            nn.SiLU(),
+            nn.Dropout(p=dropout),
+            zero_module(
+                conv_nd(dims, self.out_channels, self.out_channels, 3, padding=1)
+            ),
+        )
+
+        if self.out_channels == channels:
+            self.skip_connection = nn.Identity()
+        elif use_conv:
+            self.skip_connection = conv_nd(
+                dims, channels, self.out_channels, 3, padding=1
+            )
+        else:
+            self.skip_connection = conv_nd(dims, channels, self.out_channels, 1)
+
+    def forward(self, x, emb):
+        """
+        Apply the block to a Tensor, conditioned on a timestep embedding.
+        :param x: an [N x C x ...] Tensor of features.
+        :param emb: an [N x emb_channels] Tensor of timestep embeddings.
+        :return: an [N x C x ...] Tensor of outputs.
+        """
+        return checkpoint(
+            self._forward, (x, emb), self.parameters(), self.use_checkpoint
+        )
+
+
+    def _forward(self, x, emb):
+        if self.updown:
+            in_rest, in_conv = self.in_layers[:-1], self.in_layers[-1]
+            h = in_rest(x)
+            h = self.h_upd(h)
+            x = self.x_upd(x)
+            h = in_conv(h)
+        else:
+            h = self.in_layers(x)
+        emb_out = self.emb_layers(emb).type(h.dtype)
+        while len(emb_out.shape) < len(h.shape):
+            emb_out = emb_out[..., None]
+        if self.use_scale_shift_norm:
+            out_norm, out_rest = self.out_layers[0], self.out_layers[1:]
+            scale, shift = th.chunk(emb_out, 2, dim=1)
+            h = out_norm(h) * (1 + scale) + shift
+            h = out_rest(h)
+        else:
+            h = h + emb_out
+            h = self.out_layers(h)
+        return self.skip_connection(x) + h
+
+
+class AttentionBlock(nn.Module):
+    """
+    An attention block that allows spatial positions to attend to each other.
+    Originally ported from here, but adapted to the N-d case.
+    https://github.com/hojonathanho/diffusion/blob/1e0dceb3b3495bbe19116a5e1b3596cd0706c543/diffusion_tf/models/unet.py#L66.
+    """
+
+    def __init__(
+        self,
+        channels,
+        num_heads=1,
+        num_head_channels=-1,
+        use_checkpoint=False,
+        use_new_attention_order=False,
+    ):
+        super().__init__()
+        self.channels = channels
+        if num_head_channels == -1:
+            self.num_heads = num_heads
+        else:
+            assert (
+                channels % num_head_channels == 0
+            ), f"q,k,v channels {channels} is not divisible by num_head_channels {num_head_channels}"
+            self.num_heads = channels // num_head_channels
+        self.use_checkpoint = use_checkpoint
+        self.norm = normalization(channels)
+        self.qkv = conv_nd(1, channels, channels * 3, 1)
+        if use_new_attention_order:
+            # split qkv before split heads
+            self.attention = QKVAttention(self.num_heads)
+        else:
+            # split heads before split qkv
+            self.attention = QKVAttentionLegacy(self.num_heads)
+
+        self.proj_out = zero_module(conv_nd(1, channels, channels, 1))
+
+    def forward(self, x):
+        return checkpoint(self._forward, (x,), self.parameters(), True)   # TODO: check checkpoint usage, is True # TODO: fix the .half call!!!
+        #return pt_checkpoint(self._forward, x)  # pytorch
+
+    def _forward(self, x):
+        b, c, *spatial = x.shape
+        x = x.reshape(b, c, -1)
+        qkv = self.qkv(self.norm(x))
+        h = self.attention(qkv)
+        h = self.proj_out(h)
+        return (x + h).reshape(b, c, *spatial)
+
+
+def count_flops_attn(model, _x, y):
+    """
+    A counter for the `thop` package to count the operations in an
+    attention operation.
+    Meant to be used like:
+        macs, params = thop.profile(
+            model,
+            inputs=(inputs, timestamps),
+            custom_ops={QKVAttention: QKVAttention.count_flops},
+        )
+    """
+    b, c, *spatial = y[0].shape
+    num_spatial = int(np.prod(spatial))
+    # We perform two matmuls with the same number of ops.
+    # The first computes the weight matrix, the second computes
+    # the combination of the value vectors.
+    matmul_ops = 2 * b * (num_spatial ** 2) * c
+    model.total_ops += th.DoubleTensor([matmul_ops])
+
+
+class QKVAttentionLegacy(nn.Module):
+    """
+    A module which performs QKV attention. Matches legacy QKVAttention + input/ouput heads shaping
+    """
+
+    def __init__(self, n_heads):
+        super().__init__()
+        self.n_heads = n_heads
+
+    def forward(self, qkv):
+        """
+        Apply QKV attention.
+        :param qkv: an [N x (H * 3 * C) x T] tensor of Qs, Ks, and Vs.
+        :return: an [N x (H * C) x T] tensor after attention.
+        """
+        bs, width, length = qkv.shape
+        assert width % (3 * self.n_heads) == 0
+        ch = width // (3 * self.n_heads)
+        q, k, v = qkv.reshape(bs * self.n_heads, ch * 3, length).split(ch, dim=1)
+        scale = 1 / math.sqrt(math.sqrt(ch))
+        weight = th.einsum(
+            "bct,bcs->bts", q * scale, k * scale
+        )  # More stable with f16 than dividing afterwards
+        weight = th.softmax(weight.float(), dim=-1).type(weight.dtype)
+        a = th.einsum("bts,bcs->bct", weight, v)
+        return a.reshape(bs, -1, length)
+
+    @staticmethod
+    def count_flops(model, _x, y):
+        return count_flops_attn(model, _x, y)
+
+
+class QKVAttention(nn.Module):
+    """
+    A module which performs QKV attention and splits in a different order.
+    """
+
+    def __init__(self, n_heads):
+        super().__init__()
+        self.n_heads = n_heads
+
+    def forward(self, qkv):
+        """
+        Apply QKV attention.
+        :param qkv: an [N x (3 * H * C) x T] tensor of Qs, Ks, and Vs.
+        :return: an [N x (H * C) x T] tensor after attention.
+        """
+        bs, width, length = qkv.shape
+        assert width % (3 * self.n_heads) == 0
+        ch = width // (3 * self.n_heads)
+        q, k, v = qkv.chunk(3, dim=1)
+        scale = 1 / math.sqrt(math.sqrt(ch))
+        weight = th.einsum(
+            "bct,bcs->bts",
+            (q * scale).view(bs * self.n_heads, ch, length),
+            (k * scale).view(bs * self.n_heads, ch, length),
+        )  # More stable with f16 than dividing afterwards
+        weight = th.softmax(weight.float(), dim=-1).type(weight.dtype)
+        a = th.einsum("bts,bcs->bct", weight, v.reshape(bs * self.n_heads, ch, length))
+        return a.reshape(bs, -1, length)
+
+    @staticmethod
+    def count_flops(model, _x, y):
+        return count_flops_attn(model, _x, y)
+
+
+class UNetModel(nn.Module):
+    """
+    The full UNet model with attention and timestep embedding.
+    :param in_channels: channels in the input Tensor.
+    :param model_channels: base channel count for the model.
+    :param out_channels: channels in the output Tensor.
+    :param num_res_blocks: number of residual blocks per downsample.
+    :param attention_resolutions: a collection of downsample rates at which
+        attention will take place. May be a set, list, or tuple.
+        For example, if this contains 4, then at 4x downsampling, attention
+        will be used.
+    :param dropout: the dropout probability.
+    :param channel_mult: channel multiplier for each level of the UNet.
+    :param conv_resample: if True, use learned convolutions for upsampling and
+        downsampling.
+    :param dims: determines if the signal is 1D, 2D, or 3D.
+    :param num_classes: if specified (as an int), then this model will be
+        class-conditional with `num_classes` classes.
+    :param use_checkpoint: use gradient checkpointing to reduce memory usage.
+    :param num_heads: the number of attention heads in each attention layer.
+    :param num_heads_channels: if specified, ignore num_heads and instead use
+                               a fixed channel width per attention head.
+    :param num_heads_upsample: works with num_heads to set a different number
+                               of heads for upsampling. Deprecated.
+    :param use_scale_shift_norm: use a FiLM-like conditioning mechanism.
+    :param resblock_updown: use residual blocks for up/downsampling.
+    :param use_new_attention_order: use a different attention pattern for potentially
+                                    increased efficiency.
+    """
+
+    def __init__(
+        self,
+        image_size,
+        in_channels,
+        model_channels,
+        out_channels,
+        num_res_blocks,
+        attention_resolutions,
+        dropout=0,
+        channel_mult=(1, 2, 4, 8),
+        conv_resample=True,
+        dims=2,
+        num_classes=None,
+        use_checkpoint=False,
+        use_fp16=False,
+        num_heads=-1,
+        num_head_channels=-1,
+        num_heads_upsample=-1,
+        use_scale_shift_norm=False,
+        resblock_updown=False,
+        use_new_attention_order=False,
+        use_spatial_transformer=False,    # custom transformer support
+        transformer_depth=1,              # custom transformer support
+        context_dim=None,                 # custom transformer support
+        n_embed=None,                     # custom support for prediction of discrete ids into codebook of first stage vq model
+        legacy=True,
+        disable_self_attentions=None,
+        num_attention_blocks=None
+    ):
+        super().__init__()
+        if use_spatial_transformer:
+            assert context_dim is not None, 'Fool!! You forgot to include the dimension of your cross-attention conditioning...'
+
+        if context_dim is not None:
+            assert use_spatial_transformer, 'Fool!! You forgot to use the spatial transformer for your cross-attention conditioning...'
+            from omegaconf.listconfig import ListConfig
+            if type(context_dim) == ListConfig:
+                context_dim = list(context_dim)
+
+        if num_heads_upsample == -1:
+            num_heads_upsample = num_heads
+
+        if num_heads == -1:
+            assert num_head_channels != -1, 'Either num_heads or num_head_channels has to be set'
+
+        if num_head_channels == -1:
+            assert num_heads != -1, 'Either num_heads or num_head_channels has to be set'
+
+        self.image_size = image_size
+        self.in_channels = in_channels
+        self.model_channels = model_channels
+        self.out_channels = out_channels
+        if isinstance(num_res_blocks, int):
+            self.num_res_blocks = len(channel_mult) * [num_res_blocks]
+        else:
+            if len(num_res_blocks) != len(channel_mult):
+                raise ValueError("provide num_res_blocks either as an int (globally constant) or "
+                                 "as a list/tuple (per-level) with the same length as channel_mult")
+            self.num_res_blocks = num_res_blocks
+        #self.num_res_blocks = num_res_blocks
+        if disable_self_attentions is not None:
+            # should be a list of booleans, indicating whether to disable self-attention in TransformerBlocks or not
+            assert len(disable_self_attentions) == len(channel_mult)
+        if num_attention_blocks is not None:
+            assert len(num_attention_blocks) == len(self.num_res_blocks)
+            assert all(map(lambda i: self.num_res_blocks[i] >= num_attention_blocks[i], range(len(num_attention_blocks))))
+            print(f"Constructor of UNetModel received num_attention_blocks={num_attention_blocks}. "
+                  f"This option has LESS priority than attention_resolutions {attention_resolutions}, "
+                  f"i.e., in cases where num_attention_blocks[i] > 0 but 2**i not in attention_resolutions, "
+                  f"attention will still not be set.")  # todo: convert to warning
+
+        self.attention_resolutions = attention_resolutions
+        self.dropout = dropout
+        self.channel_mult = channel_mult
+        self.conv_resample = conv_resample
+        self.num_classes = num_classes
+        self.use_checkpoint = use_checkpoint
+        self.dtype = th.float16 if use_fp16 else th.float32
+        self.num_heads = num_heads
+        self.num_head_channels = num_head_channels
+        self.num_heads_upsample = num_heads_upsample
+        self.predict_codebook_ids = n_embed is not None
+
+        time_embed_dim = model_channels * 4
+        self.time_embed = nn.Sequential(
+            linear(model_channels, time_embed_dim),
+            nn.SiLU(),
+            linear(time_embed_dim, time_embed_dim),
+        )
+
+        if self.num_classes is not None:
+            self.label_emb = nn.Embedding(num_classes, time_embed_dim)
+
+        self.input_blocks = nn.ModuleList(
+            [
+                TimestepEmbedSequential(
+                    conv_nd(dims, in_channels, model_channels, 3, padding=1)
+                )
+            ]
+        )
+        self._feature_size = model_channels
+        input_block_chans = [model_channels]
+        ch = model_channels
+        ds = 1
+        for level, mult in enumerate(channel_mult):
+            for nr in range(self.num_res_blocks[level]):
+                layers = [
+                    ResBlock(
+                        ch,
+                        time_embed_dim,
+                        dropout,
+                        out_channels=mult * model_channels,
+                        dims=dims,
+                        use_checkpoint=use_checkpoint,
+                        use_scale_shift_norm=use_scale_shift_norm,
+                    )
+                ]
+                ch = mult * model_channels
+                if ds in attention_resolutions:
+                    if num_head_channels == -1:
+                        dim_head = ch // num_heads
+                    else:
+                        num_heads = ch // num_head_channels
+                        dim_head = num_head_channels
+                    if legacy:
+                        #num_heads = 1
+                        dim_head = ch // num_heads if use_spatial_transformer else num_head_channels
+                    if exists(disable_self_attentions):
+                        disabled_sa = disable_self_attentions[level]
+                    else:
+                        disabled_sa = False
+
+                    if not exists(num_attention_blocks) or nr < num_attention_blocks[level]:
+                        layers.append(
+                            AttentionBlock(
+                                ch,
+                                use_checkpoint=use_checkpoint,
+                                num_heads=num_heads,
+                                num_head_channels=dim_head,
+                                use_new_attention_order=use_new_attention_order,
+                            ) if not use_spatial_transformer else SpatialTransformer(
+                                ch, num_heads, dim_head, depth=transformer_depth, context_dim=context_dim,
+                                disable_self_attn=disabled_sa
+                            )
+                        )
+                self.input_blocks.append(TimestepEmbedSequential(*layers))
+                self._feature_size += ch
+                input_block_chans.append(ch)
+            if level != len(channel_mult) - 1:
+                out_ch = ch
+                self.input_blocks.append(
+                    TimestepEmbedSequential(
+                        ResBlock(
+                            ch,
+                            time_embed_dim,
+                            dropout,
+                            out_channels=out_ch,
+                            dims=dims,
+                            use_checkpoint=use_checkpoint,
+                            use_scale_shift_norm=use_scale_shift_norm,
+                            down=True,
+                        )
+                        if resblock_updown
+                        else Downsample(
+                            ch, conv_resample, dims=dims, out_channels=out_ch
+                        )
+                    )
+                )
+                ch = out_ch
+                input_block_chans.append(ch)
+                ds *= 2
+                self._feature_size += ch
+
+        if num_head_channels == -1:
+            dim_head = ch // num_heads
+        else:
+            num_heads = ch // num_head_channels
+            dim_head = num_head_channels
+        if legacy:
+            #num_heads = 1
+            dim_head = ch // num_heads if use_spatial_transformer else num_head_channels
+        self.middle_block = TimestepEmbedSequential(
+            ResBlock(
+                ch,
+                time_embed_dim,
+                dropout,
+                dims=dims,
+                use_checkpoint=use_checkpoint,
+                use_scale_shift_norm=use_scale_shift_norm,
+            ),
+            AttentionBlock(
+                ch,
+                use_checkpoint=use_checkpoint,
+                num_heads=num_heads,
+                num_head_channels=dim_head,
+                use_new_attention_order=use_new_attention_order,
+            ) if not use_spatial_transformer else SpatialTransformer(  # always uses a self-attn
+                            ch, num_heads, dim_head, depth=transformer_depth, context_dim=context_dim
+                        ),
+            ResBlock(
+                ch,
+                time_embed_dim,
+                dropout,
+                dims=dims,
+                use_checkpoint=use_checkpoint,
+                use_scale_shift_norm=use_scale_shift_norm,
+            ),
+        )
+        self._feature_size += ch
+
+        self.output_blocks = nn.ModuleList([])
+        for level, mult in list(enumerate(channel_mult))[::-1]:
+            for i in range(self.num_res_blocks[level] + 1):
+                ich = input_block_chans.pop()
+                layers = [
+                    ResBlock(
+                        ch + ich,
+                        time_embed_dim,
+                        dropout,
+                        out_channels=model_channels * mult,
+                        dims=dims,
+                        use_checkpoint=use_checkpoint,
+                        use_scale_shift_norm=use_scale_shift_norm,
+                    )
+                ]
+                ch = model_channels * mult
+                if ds in attention_resolutions:
+                    if num_head_channels == -1:
+                        dim_head = ch // num_heads
+                    else:
+                        num_heads = ch // num_head_channels
+                        dim_head = num_head_channels
+                    if legacy:
+                        #num_heads = 1
+                        dim_head = ch // num_heads if use_spatial_transformer else num_head_channels
+                    if exists(disable_self_attentions):
+                        disabled_sa = disable_self_attentions[level]
+                    else:
+                        disabled_sa = False
+
+                    if not exists(num_attention_blocks) or i < num_attention_blocks[level]:
+                        layers.append(
+                            AttentionBlock(
+                                ch,
+                                use_checkpoint=use_checkpoint,
+                                num_heads=num_heads_upsample,
+                                num_head_channels=dim_head,
+                                use_new_attention_order=use_new_attention_order,
+                            ) if not use_spatial_transformer else SpatialTransformer(
+                                ch, num_heads, dim_head, depth=transformer_depth, context_dim=context_dim,
+                                disable_self_attn=disabled_sa
+                            )
+                        )
+                if level and i == self.num_res_blocks[level]:
+                    out_ch = ch
+                    layers.append(
+                        ResBlock(
+                            ch,
+                            time_embed_dim,
+                            dropout,
+                            out_channels=out_ch,
+                            dims=dims,
+                            use_checkpoint=use_checkpoint,
+                            use_scale_shift_norm=use_scale_shift_norm,
+                            up=True,
+                        )
+                        if resblock_updown
+                        else Upsample(ch, conv_resample, dims=dims, out_channels=out_ch)
+                    )
+                    ds //= 2
+                self.output_blocks.append(TimestepEmbedSequential(*layers))
+                self._feature_size += ch
+
+        self.out = nn.Sequential(
+            normalization(ch),
+            nn.SiLU(),
+            zero_module(conv_nd(dims, model_channels, out_channels, 3, padding=1)),
+        )
+        if self.predict_codebook_ids:
+            self.id_predictor = nn.Sequential(
+            normalization(ch),
+            conv_nd(dims, model_channels, n_embed, 1),
+            #nn.LogSoftmax(dim=1)  # change to cross_entropy and produce non-normalized logits
+        )
+
+    def convert_to_fp16(self):
+        """
+        Convert the torso of the model to float16.
+        """
+        self.input_blocks.apply(convert_module_to_f16)
+        self.middle_block.apply(convert_module_to_f16)
+        self.output_blocks.apply(convert_module_to_f16)
+
+    def convert_to_fp32(self):
+        """
+        Convert the torso of the model to float32.
+        """
+        self.input_blocks.apply(convert_module_to_f32)
+        self.middle_block.apply(convert_module_to_f32)
+        self.output_blocks.apply(convert_module_to_f32)
+
+    def forward(self, x, timesteps=None, context=None, y=None,**kwargs):
+        """
+        Apply the model to an input batch.
+        :param x: an [N x C x ...] Tensor of inputs.
+        :param timesteps: a 1-D batch of timesteps.
+        :param context: conditioning plugged in via crossattn
+        :param y: an [N] Tensor of labels, if class-conditional.
+        :return: an [N x C x ...] Tensor of outputs.
+        """
+        assert (y is not None) == (
+            self.num_classes is not None
+        ), "must specify y if and only if the model is class-conditional"
+        hs = []
+        t_emb = timestep_embedding(timesteps, self.model_channels, repeat_only=False)
+        emb = self.time_embed(t_emb)
+
+        if self.num_classes is not None:
+            assert y.shape == (x.shape[0],)
+            emb = emb + self.label_emb(y)
+
+        h = x.type(self.dtype)
+        for module in self.input_blocks:
+            h = module(h, emb, context)
+            hs.append(h)
+        h = self.middle_block(h, emb, context)
+        for module in self.output_blocks:
+            h = th.cat([h, hs.pop()], dim=1)
+            h = module(h, emb, context)
+        h = h.type(x.dtype)
+        if self.predict_codebook_ids:
+            return self.id_predictor(h)
+        else:
+            return self.out(h)
+
+
+class EncoderUNetModel(nn.Module):
+    """
+    The half UNet model with attention and timestep embedding.
+    For usage, see UNet.
+    """
+
+    def __init__(
+        self,
+        image_size,
+        in_channels,
+        model_channels,
+        out_channels,
+        num_res_blocks,
+        attention_resolutions,
+        dropout=0,
+        channel_mult=(1, 2, 4, 8),
+        conv_resample=True,
+        dims=2,
+        use_checkpoint=False,
+        use_fp16=False,
+        num_heads=1,
+        num_head_channels=-1,
+        num_heads_upsample=-1,
+        use_scale_shift_norm=False,
+        resblock_updown=False,
+        use_new_attention_order=False,
+        pool="adaptive",
+        *args,
+        **kwargs
+    ):
+        super().__init__()
+
+        if num_heads_upsample == -1:
+            num_heads_upsample = num_heads
+
+        self.in_channels = in_channels
+        self.model_channels = model_channels
+        self.out_channels = out_channels
+        self.num_res_blocks = num_res_blocks
+        self.attention_resolutions = attention_resolutions
+        self.dropout = dropout
+        self.channel_mult = channel_mult
+        self.conv_resample = conv_resample
+        self.use_checkpoint = use_checkpoint
+        self.dtype = th.float16 if use_fp16 else th.float32
+        self.num_heads = num_heads
+        self.num_head_channels = num_head_channels
+        self.num_heads_upsample = num_heads_upsample
+
+        time_embed_dim = model_channels * 4
+        self.time_embed = nn.Sequential(
+            linear(model_channels, time_embed_dim),
+            nn.SiLU(),
+            linear(time_embed_dim, time_embed_dim),
+        )
+
+        self.input_blocks = nn.ModuleList(
+            [
+                TimestepEmbedSequential(
+                    conv_nd(dims, in_channels, model_channels, 3, padding=1)
+                )
+            ]
+        )
+        self._feature_size = model_channels
+        input_block_chans = [model_channels]
+        ch = model_channels
+        ds = 1
+        for level, mult in enumerate(channel_mult):
+            for _ in range(num_res_blocks):
+                layers = [
+                    ResBlock(
+                        ch,
+                        time_embed_dim,
+                        dropout,
+                        out_channels=mult * model_channels,
+                        dims=dims,
+                        use_checkpoint=use_checkpoint,
+                        use_scale_shift_norm=use_scale_shift_norm,
+                    )
+                ]
+                ch = mult * model_channels
+                if ds in attention_resolutions:
+                    layers.append(
+                        AttentionBlock(
+                            ch,
+                            use_checkpoint=use_checkpoint,
+                            num_heads=num_heads,
+                            num_head_channels=num_head_channels,
+                            use_new_attention_order=use_new_attention_order,
+                        )
+                    )
+                self.input_blocks.append(TimestepEmbedSequential(*layers))
+                self._feature_size += ch
+                input_block_chans.append(ch)
+            if level != len(channel_mult) - 1:
+                out_ch = ch
+                self.input_blocks.append(
+                    TimestepEmbedSequential(
+                        ResBlock(
+                            ch,
+                            time_embed_dim,
+                            dropout,
+                            out_channels=out_ch,
+                            dims=dims,
+                            use_checkpoint=use_checkpoint,
+                            use_scale_shift_norm=use_scale_shift_norm,
+                            down=True,
+                        )
+                        if resblock_updown
+                        else Downsample(
+                            ch, conv_resample, dims=dims, out_channels=out_ch
+                        )
+                    )
+                )
+                ch = out_ch
+                input_block_chans.append(ch)
+                ds *= 2
+                self._feature_size += ch
+
+        self.middle_block = TimestepEmbedSequential(
+            ResBlock(
+                ch,
+                time_embed_dim,
+                dropout,
+                dims=dims,
+                use_checkpoint=use_checkpoint,
+                use_scale_shift_norm=use_scale_shift_norm,
+            ),
+            AttentionBlock(
+                ch,
+                use_checkpoint=use_checkpoint,
+                num_heads=num_heads,
+                num_head_channels=num_head_channels,
+                use_new_attention_order=use_new_attention_order,
+            ),
+            ResBlock(
+                ch,
+                time_embed_dim,
+                dropout,
+                dims=dims,
+                use_checkpoint=use_checkpoint,
+                use_scale_shift_norm=use_scale_shift_norm,
+            ),
+        )
+        self._feature_size += ch
+        self.pool = pool
+        if pool == "adaptive":
+            self.out = nn.Sequential(
+                normalization(ch),
+                nn.SiLU(),
+                nn.AdaptiveAvgPool2d((1, 1)),
+                zero_module(conv_nd(dims, ch, out_channels, 1)),
+                nn.Flatten(),
+            )
+        elif pool == "attention":
+            assert num_head_channels != -1
+            self.out = nn.Sequential(
+                normalization(ch),
+                nn.SiLU(),
+                AttentionPool2d(
+                    (image_size // ds), ch, num_head_channels, out_channels
+                ),
+            )
+        elif pool == "spatial":
+            self.out = nn.Sequential(
+                nn.Linear(self._feature_size, 2048),
+                nn.ReLU(),
+                nn.Linear(2048, self.out_channels),
+            )
+        elif pool == "spatial_v2":
+            self.out = nn.Sequential(
+                nn.Linear(self._feature_size, 2048),
+                normalization(2048),
+                nn.SiLU(),
+                nn.Linear(2048, self.out_channels),
+            )
+        else:
+            raise NotImplementedError(f"Unexpected {pool} pooling")
+
+    def convert_to_fp16(self):
+        """
+        Convert the torso of the model to float16.
+        """
+        self.input_blocks.apply(convert_module_to_f16)
+        self.middle_block.apply(convert_module_to_f16)
+
+    def convert_to_fp32(self):
+        """
+        Convert the torso of the model to float32.
+        """
+        self.input_blocks.apply(convert_module_to_f32)
+        self.middle_block.apply(convert_module_to_f32)
+
+    def forward(self, x, timesteps):
+        """
+        Apply the model to an input batch.
+        :param x: an [N x C x ...] Tensor of inputs.
+        :param timesteps: a 1-D batch of timesteps.
+        :return: an [N x K] Tensor of outputs.
+        """
+        emb = self.time_embed(timestep_embedding(timesteps, self.model_channels))
+
+        results = []
+        h = x.type(self.dtype)
+        for module in self.input_blocks:
+            h = module(h, emb)
+            if self.pool.startswith("spatial"):
+                results.append(h.type(x.dtype).mean(dim=(2, 3)))
+        h = self.middle_block(h, emb)
+        if self.pool.startswith("spatial"):
+            results.append(h.type(x.dtype).mean(dim=(2, 3)))
+            h = th.cat(results, axis=-1)
+            return self.out(h)
+        else:
+            h = h.type(x.dtype)
+            return self.out(h)
+

ldm/modules/diffusionmodules/util.py

@@ -0,0 +1,267 @@
+# adopted from
+# https://github.com/openai/improved-diffusion/blob/main/improved_diffusion/gaussian_diffusion.py
+# and
+# https://github.com/lucidrains/denoising-diffusion-pytorch/blob/7706bdfc6f527f58d33f84b7b522e61e6e3164b3/denoising_diffusion_pytorch/denoising_diffusion_pytorch.py
+# and
+# https://github.com/openai/guided-diffusion/blob/0ba878e517b276c45d1195eb29f6f5f72659a05b/guided_diffusion/nn.py
+#
+# thanks!
+
+
+import os
+import math
+import torch
+import torch.nn as nn
+import numpy as np
+from einops import repeat
+
+from ldm.util import instantiate_from_config
+
+
+def make_beta_schedule(schedule, n_timestep, linear_start=1e-4, linear_end=2e-2, cosine_s=8e-3):
+    if schedule == "linear":
+        betas = (
+                torch.linspace(linear_start ** 0.5, linear_end ** 0.5, n_timestep, dtype=torch.float64) ** 2
+        )
+
+    elif schedule == "cosine":
+        timesteps = (
+                torch.arange(n_timestep + 1, dtype=torch.float64) / n_timestep + cosine_s
+        )
+        alphas = timesteps / (1 + cosine_s) * np.pi / 2
+        alphas = torch.cos(alphas).pow(2)
+        alphas = alphas / alphas[0]
+        betas = 1 - alphas[1:] / alphas[:-1]
+        betas = np.clip(betas, a_min=0, a_max=0.999)
+
+    elif schedule == "sqrt_linear":
+        betas = torch.linspace(linear_start, linear_end, n_timestep, dtype=torch.float64)
+    elif schedule == "sqrt":
+        betas = torch.linspace(linear_start, linear_end, n_timestep, dtype=torch.float64) ** 0.5
+    else:
+        raise ValueError(f"schedule '{schedule}' unknown.")
+    return betas.numpy()
+
+
+def make_ddim_timesteps(ddim_discr_method, num_ddim_timesteps, num_ddpm_timesteps, verbose=True):
+    if ddim_discr_method == 'uniform':
+        c = num_ddpm_timesteps // num_ddim_timesteps
+        ddim_timesteps = np.asarray(list(range(0, num_ddpm_timesteps, c)))
+    elif ddim_discr_method == 'quad':
+        ddim_timesteps = ((np.linspace(0, np.sqrt(num_ddpm_timesteps * .8), num_ddim_timesteps)) ** 2).astype(int)
+    else:
+        raise NotImplementedError(f'There is no ddim discretization method called "{ddim_discr_method}"')
+
+    # assert ddim_timesteps.shape[0] == num_ddim_timesteps
+    # add one to get the final alpha values right (the ones from first scale to data during sampling)
+    steps_out = ddim_timesteps + 1
+    if verbose:
+        print(f'Selected timesteps for ddim sampler: {steps_out}')
+    return steps_out
+
+
+def make_ddim_sampling_parameters(alphacums, ddim_timesteps, eta, verbose=True):
+    # select alphas for computing the variance schedule
+    alphas = alphacums[ddim_timesteps]
+    alphas_prev = np.asarray([alphacums[0]] + alphacums[ddim_timesteps[:-1]].tolist())
+
+    # according the the formula provided in https://arxiv.org/abs/2010.02502
+    sigmas = eta * np.sqrt((1 - alphas_prev) / (1 - alphas) * (1 - alphas / alphas_prev))
+    if verbose:
+        print(f'Selected alphas for ddim sampler: a_t: {alphas}; a_(t-1): {alphas_prev}')
+        print(f'For the chosen value of eta, which is {eta}, '
+              f'this results in the following sigma_t schedule for ddim sampler {sigmas}')
+    return sigmas, alphas, alphas_prev
+
+
+def betas_for_alpha_bar(num_diffusion_timesteps, alpha_bar, max_beta=0.999):
+    """
+    Create a beta schedule that discretizes the given alpha_t_bar function,
+    which defines the cumulative product of (1-beta) over time from t = [0,1].
+    :param num_diffusion_timesteps: the number of betas to produce.
+    :param alpha_bar: a lambda that takes an argument t from 0 to 1 and
+                      produces the cumulative product of (1-beta) up to that
+                      part of the diffusion process.
+    :param max_beta: the maximum beta to use; use values lower than 1 to
+                     prevent singularities.
+    """
+    betas = []
+    for i in range(num_diffusion_timesteps):
+        t1 = i / num_diffusion_timesteps
+        t2 = (i + 1) / num_diffusion_timesteps
+        betas.append(min(1 - alpha_bar(t2) / alpha_bar(t1), max_beta))
+    return np.array(betas)
+
+
+def extract_into_tensor(a, t, x_shape):
+    b, *_ = t.shape
+    out = a.gather(-1, t)
+    return out.reshape(b, *((1,) * (len(x_shape) - 1)))
+
+
+def checkpoint(func, inputs, params, flag):
+    """
+    Evaluate a function without caching intermediate activations, allowing for
+    reduced memory at the expense of extra compute in the backward pass.
+    :param func: the function to evaluate.
+    :param inputs: the argument sequence to pass to `func`.
+    :param params: a sequence of parameters `func` depends on but does not
+                   explicitly take as arguments.
+    :param flag: if False, disable gradient checkpointing.
+    """
+    if flag:
+        args = tuple(inputs) + tuple(params)
+        return CheckpointFunction.apply(func, len(inputs), *args)
+    else:
+        return func(*inputs)
+
+
+class CheckpointFunction(torch.autograd.Function):
+    @staticmethod
+    def forward(ctx, run_function, length, *args):
+        ctx.run_function = run_function
+        ctx.input_tensors = list(args[:length])
+        ctx.input_params = list(args[length:])
+
+        with torch.no_grad():
+            output_tensors = ctx.run_function(*ctx.input_tensors)
+        return output_tensors
+
+    @staticmethod
+    def backward(ctx, *output_grads):
+        ctx.input_tensors = [x.detach().requires_grad_(True) for x in ctx.input_tensors]
+        with torch.enable_grad():
+            # Fixes a bug where the first op in run_function modifies the
+            # Tensor storage in place, which is not allowed for detach()'d
+            # Tensors.
+            shallow_copies = [x.view_as(x) for x in ctx.input_tensors]
+            output_tensors = ctx.run_function(*shallow_copies)
+        input_grads = torch.autograd.grad(
+            output_tensors,
+            ctx.input_tensors + ctx.input_params,
+            output_grads,
+            allow_unused=True,
+        )
+        del ctx.input_tensors
+        del ctx.input_params
+        del output_tensors
+        return (None, None) + input_grads
+
+
+def timestep_embedding(timesteps, dim, max_period=10000, repeat_only=False):
+    """
+    Create sinusoidal timestep embeddings.
+    :param timesteps: a 1-D Tensor of N indices, one per batch element.
+                      These may be fractional.
+    :param dim: the dimension of the output.
+    :param max_period: controls the minimum frequency of the embeddings.
+    :return: an [N x dim] Tensor of positional embeddings.
+    """
+    if not repeat_only:
+        half = dim // 2
+        freqs = torch.exp(
+            -math.log(max_period) * torch.arange(start=0, end=half, dtype=torch.float32) / half
+        ).to(device=timesteps.device)
+        args = timesteps[:, None].float() * freqs[None]
+        embedding = torch.cat([torch.cos(args), torch.sin(args)], dim=-1)
+        if dim % 2:
+            embedding = torch.cat([embedding, torch.zeros_like(embedding[:, :1])], dim=-1)
+    else:
+        embedding = repeat(timesteps, 'b -> b d', d=dim)
+    return embedding
+
+
+def zero_module(module):
+    """
+    Zero out the parameters of a module and return it.
+    """
+    for p in module.parameters():
+        p.detach().zero_()
+    return module
+
+
+def scale_module(module, scale):
+    """
+    Scale the parameters of a module and return it.
+    """
+    for p in module.parameters():
+        p.detach().mul_(scale)
+    return module
+
+
+def mean_flat(tensor):
+    """
+    Take the mean over all non-batch dimensions.
+    """
+    return tensor.mean(dim=list(range(1, len(tensor.shape))))
+
+
+def normalization(channels):
+    """
+    Make a standard normalization layer.
+    :param channels: number of input channels.
+    :return: an nn.Module for normalization.
+    """
+    return GroupNorm32(32, channels)
+
+
+# PyTorch 1.7 has SiLU, but we support PyTorch 1.5.
+class SiLU(nn.Module):
+    def forward(self, x):
+        return x * torch.sigmoid(x)
+
+
+class GroupNorm32(nn.GroupNorm):
+    def forward(self, x):
+        return super().forward(x.float()).type(x.dtype)
+
+def conv_nd(dims, *args, **kwargs):
+    """
+    Create a 1D, 2D, or 3D convolution module.
+    """
+    if dims == 1:
+        return nn.Conv1d(*args, **kwargs)
+    elif dims == 2:
+        return nn.Conv2d(*args, **kwargs)
+    elif dims == 3:
+        return nn.Conv3d(*args, **kwargs)
+    raise ValueError(f"unsupported dimensions: {dims}")
+
+
+def linear(*args, **kwargs):
+    """
+    Create a linear module.
+    """
+    return nn.Linear(*args, **kwargs)
+
+
+def avg_pool_nd(dims, *args, **kwargs):
+    """
+    Create a 1D, 2D, or 3D average pooling module.
+    """
+    if dims == 1:
+        return nn.AvgPool1d(*args, **kwargs)
+    elif dims == 2:
+        return nn.AvgPool2d(*args, **kwargs)
+    elif dims == 3:
+        return nn.AvgPool3d(*args, **kwargs)
+    raise ValueError(f"unsupported dimensions: {dims}")
+
+
+class HybridConditioner(nn.Module):
+
+    def __init__(self, c_concat_config, c_crossattn_config):
+        super().__init__()
+        self.concat_conditioner = instantiate_from_config(c_concat_config)
+        self.crossattn_conditioner = instantiate_from_config(c_crossattn_config)
+
+    def forward(self, c_concat, c_crossattn):
+        c_concat = self.concat_conditioner(c_concat)
+        c_crossattn = self.crossattn_conditioner(c_crossattn)
+        return {'c_concat': [c_concat], 'c_crossattn': [c_crossattn]}
+
+
+def noise_like(shape, device, repeat=False):
+    repeat_noise = lambda: torch.randn((1, *shape[1:]), device=device).repeat(shape[0], *((1,) * (len(shape) - 1)))
+    noise = lambda: torch.randn(shape, device=device)
+    return repeat_noise() if repeat else noise()

ldm/modules/distributions/distributions.py

@@ -0,0 +1,92 @@
+import torch
+import numpy as np
+
+
+class AbstractDistribution:
+    def sample(self):
+        raise NotImplementedError()
+
+    def mode(self):
+        raise NotImplementedError()
+
+
+class DiracDistribution(AbstractDistribution):
+    def __init__(self, value):
+        self.value = value
+
+    def sample(self):
+        return self.value
+
+    def mode(self):
+        return self.value
+
+
+class DiagonalGaussianDistribution(object):
+    def __init__(self, parameters, deterministic=False):
+        self.parameters = parameters
+        self.mean, self.logvar = torch.chunk(parameters, 2, dim=1)
+        self.logvar = torch.clamp(self.logvar, -30.0, 20.0)
+        self.deterministic = deterministic
+        self.std = torch.exp(0.5 * self.logvar)
+        self.var = torch.exp(self.logvar)
+        if self.deterministic:
+            self.var = self.std = torch.zeros_like(self.mean).to(device=self.parameters.device)
+
+    def sample(self):
+        x = self.mean + self.std * torch.randn(self.mean.shape).to(device=self.parameters.device)
+        return x
+
+    def kl(self, other=None):
+        if self.deterministic:
+            return torch.Tensor([0.])
+        else:
+            if other is None:
+                return 0.5 * torch.sum(torch.pow(self.mean, 2)
+                                       + self.var - 1.0 - self.logvar,
+                                       dim=[1, 2, 3])
+            else:
+                return 0.5 * torch.sum(
+                    torch.pow(self.mean - other.mean, 2) / other.var
+                    + self.var / other.var - 1.0 - self.logvar + other.logvar,
+                    dim=[1, 2, 3])
+
+    def nll(self, sample, dims=[1,2,3]):
+        if self.deterministic:
+            return torch.Tensor([0.])
+        logtwopi = np.log(2.0 * np.pi)
+        return 0.5 * torch.sum(
+            logtwopi + self.logvar + torch.pow(sample - self.mean, 2) / self.var,
+            dim=dims)
+
+    def mode(self):
+        return self.mean
+
+
+def normal_kl(mean1, logvar1, mean2, logvar2):
+    """
+    source: https://github.com/openai/guided-diffusion/blob/27c20a8fab9cb472df5d6bdd6c8d11c8f430b924/guided_diffusion/losses.py#L12
+    Compute the KL divergence between two gaussians.
+    Shapes are automatically broadcasted, so batches can be compared to
+    scalars, among other use cases.
+    """
+    tensor = None
+    for obj in (mean1, logvar1, mean2, logvar2):
+        if isinstance(obj, torch.Tensor):
+            tensor = obj
+            break
+    assert tensor is not None, "at least one argument must be a Tensor"
+
+    # Force variances to be Tensors. Broadcasting helps convert scalars to
+    # Tensors, but it does not work for torch.exp().
+    logvar1, logvar2 = [
+        x if isinstance(x, torch.Tensor) else torch.tensor(x).to(tensor)
+        for x in (logvar1, logvar2)
+    ]
+
+    return 0.5 * (
+        -1.0
+        + logvar2
+        - logvar1
+        + torch.exp(logvar1 - logvar2)
+        + ((mean1 - mean2) ** 2) * torch.exp(-logvar2)
+    )

ldm/modules/ema.py

@@ -0,0 +1,76 @@
+import torch
+from torch import nn
+
+
+class LitEma(nn.Module):
+    def __init__(self, model, decay=0.9999, use_num_upates=True):
+        super().__init__()
+        if decay < 0.0 or decay > 1.0:
+            raise ValueError('Decay must be between 0 and 1')
+
+        self.m_name2s_name = {}
+        self.register_buffer('decay', torch.tensor(decay, dtype=torch.float32))
+        self.register_buffer('num_updates', torch.tensor(0,dtype=torch.int) if use_num_upates
+                             else torch.tensor(-1,dtype=torch.int))
+
+        for name, p in model.named_parameters():
+            if p.requires_grad:
+                #remove as '.'-character is not allowed in buffers
+                s_name = name.replace('.','')
+                self.m_name2s_name.update({name:s_name})
+                self.register_buffer(s_name,p.clone().detach().data)
+
+        self.collected_params = []
+
+    def forward(self,model):
+        decay = self.decay
+
+        if self.num_updates >= 0:
+            self.num_updates += 1
+            decay = min(self.decay,(1 + self.num_updates) / (10 + self.num_updates))
+
+        one_minus_decay = 1.0 - decay
+
+        with torch.no_grad():
+            m_param = dict(model.named_parameters())
+            shadow_params = dict(self.named_buffers())
+
+            for key in m_param:
+                if m_param[key].requires_grad:
+                    sname = self.m_name2s_name[key]
+                    shadow_params[sname] = shadow_params[sname].type_as(m_param[key])
+                    shadow_params[sname].sub_(one_minus_decay * (shadow_params[sname] - m_param[key]))
+                else:
+                    assert not key in self.m_name2s_name
+
+    def copy_to(self, model):
+        m_param = dict(model.named_parameters())
+        shadow_params = dict(self.named_buffers())
+        for key in m_param:
+            if m_param[key].requires_grad:
+                m_param[key].data.copy_(shadow_params[self.m_name2s_name[key]].data)
+            else:
+                assert not key in self.m_name2s_name
+
+    def store(self, parameters):
+        """
+        Save the current parameters for restoring later.
+        Args:
+          parameters: Iterable of `torch.nn.Parameter`; the parameters to be
+            temporarily stored.
+        """
+        self.collected_params = [param.clone() for param in parameters]
+
+    def restore(self, parameters):
+        """
+        Restore the parameters stored with the `store` method.
+        Useful to validate the model with EMA parameters without affecting the
+        original optimization process. Store the parameters before the
+        `copy_to` method. After validation (or model saving), use this to
+        restore the former parameters.
+        Args:
+          parameters: Iterable of `torch.nn.Parameter`; the parameters to be
+            updated with the stored parameters.
+        """
+        for c_param, param in zip(self.collected_params, parameters):
+            param.data.copy_(c_param.data)

ldm/modules/encoders/modules.py

@@ -0,0 +1,550 @@
+import torch
+import torch.nn as nn
+import numpy as np
+from functools import partial
+import kornia
+
+from ldm.modules.x_transformer import Encoder, TransformerWrapper  # TODO: can we directly rely on lucidrains code and simply add this as a reuirement? --> test
+from ldm.util import default
+import clip
+
+
+class AbstractEncoder(nn.Module):
+    def __init__(self):
+        super().__init__()
+
+    def encode(self, *args, **kwargs):
+        raise NotImplementedError
+
+class IdentityEncoder(AbstractEncoder):
+
+    def encode(self, x):
+        return x
+
+class FaceClipEncoder(AbstractEncoder):
+    def __init__(self, augment=True, retreival_key=None):
+        super().__init__()
+        self.encoder = FrozenCLIPImageEmbedder()
+        self.augment = augment
+        self.retreival_key = retreival_key
+
+    def forward(self, img):
+        encodings = []
+        with torch.no_grad():
+            x_offset = 125
+            if self.retreival_key:
+                # Assumes retrieved image are packed into the second half of channels
+                face = img[:,3:,190:440,x_offset:(512-x_offset)]
+                other = img[:,:3,...].clone()
+            else:
+                face = img[:,:,190:440,x_offset:(512-x_offset)]
+                other = img.clone()
+
+            if self.augment:
+                face = K.RandomHorizontalFlip()(face)
+
+            other[:,:,190:440,x_offset:(512-x_offset)] *= 0
+            encodings = [
+                self.encoder.encode(face),
+                self.encoder.encode(other),
+            ]
+
+        return torch.cat(encodings, dim=1)
+
+    def encode(self, img):
+        if isinstance(img, list):
+            # Uncondition
+            return torch.zeros((1, 2, 768), device=self.encoder.model.visual.conv1.weight.device)
+
+        return self(img)
+
+class FaceIdClipEncoder(AbstractEncoder):
+    def __init__(self):
+        super().__init__()
+        self.encoder = FrozenCLIPImageEmbedder()
+        for p in self.encoder.parameters():
+            p.requires_grad = False
+        self.id = FrozenFaceEncoder("/home/jpinkney/code/stable-diffusion/model_ir_se50.pth", augment=True)
+
+    def forward(self, img):
+        encodings = []
+        with torch.no_grad():
+            face = kornia.geometry.resize(img, (256, 256),
+                            interpolation='bilinear', align_corners=True)
+
+            other = img.clone()
+            other[:,:,184:452,122:396] *= 0
+            encodings = [
+                self.id.encode(face),
+                self.encoder.encode(other),
+            ]
+
+        return torch.cat(encodings, dim=1)
+
+    def encode(self, img):
+        if isinstance(img, list):
+            # Uncondition
+            return torch.zeros((1, 2, 768), device=self.encoder.model.visual.conv1.weight.device)
+
+        return self(img)
+
+class ClassEmbedder(nn.Module):
+    def __init__(self, embed_dim, n_classes=1000, key='class'):
+        super().__init__()
+        self.key = key
+        self.embedding = nn.Embedding(n_classes, embed_dim)
+
+    def forward(self, batch, key=None):
+        if key is None:
+            key = self.key
+        # this is for use in crossattn
+        c = batch[key][:, None]
+        c = self.embedding(c)
+        return c
+
+
+class TransformerEmbedder(AbstractEncoder):
+    """Some transformer encoder layers"""
+    def __init__(self, n_embed, n_layer, vocab_size, max_seq_len=77, device="cuda"):
+        super().__init__()
+        self.device = device
+        self.transformer = TransformerWrapper(num_tokens=vocab_size, max_seq_len=max_seq_len,
+                                              attn_layers=Encoder(dim=n_embed, depth=n_layer))
+
+    def forward(self, tokens):
+        tokens = tokens.to(self.device)  # meh
+        z = self.transformer(tokens, return_embeddings=True)
+        return z
+
+    def encode(self, x):
+        return self(x)
+
+
+class BERTTokenizer(AbstractEncoder):
+    """ Uses a pretrained BERT tokenizer by huggingface. Vocab size: 30522 (?)"""
+    def __init__(self, device="cuda", vq_interface=True, max_length=77):
+        super().__init__()
+        from transformers import BertTokenizerFast  # TODO: add to reuquirements
+        self.tokenizer = BertTokenizerFast.from_pretrained("bert-base-uncased")
+        self.device = device
+        self.vq_interface = vq_interface
+        self.max_length = max_length
+
+    def forward(self, text):
+        batch_encoding = self.tokenizer(text, truncation=True, max_length=self.max_length, return_length=True,
+                                        return_overflowing_tokens=False, padding="max_length", return_tensors="pt")
+        tokens = batch_encoding["input_ids"].to(self.device)
+        return tokens
+
+    @torch.no_grad()
+    def encode(self, text):
+        tokens = self(text)
+        if not self.vq_interface:
+            return tokens
+        return None, None, [None, None, tokens]
+
+    def decode(self, text):
+        return text
+
+
+class BERTEmbedder(AbstractEncoder):
+    """Uses the BERT tokenizr model and add some transformer encoder layers"""
+    def __init__(self, n_embed, n_layer, vocab_size=30522, max_seq_len=77,
+                 device="cuda",use_tokenizer=True, embedding_dropout=0.0):
+        super().__init__()
+        self.use_tknz_fn = use_tokenizer
+        if self.use_tknz_fn:
+            self.tknz_fn = BERTTokenizer(vq_interface=False, max_length=max_seq_len)
+        self.device = device
+        self.transformer = TransformerWrapper(num_tokens=vocab_size, max_seq_len=max_seq_len,
+                                              attn_layers=Encoder(dim=n_embed, depth=n_layer),
+                                              emb_dropout=embedding_dropout)
+
+    def forward(self, text):
+        if self.use_tknz_fn:
+            tokens = self.tknz_fn(text)#.to(self.device)
+        else:
+            tokens = text
+        z = self.transformer(tokens, return_embeddings=True)
+        return z
+
+    def encode(self, text):
+        # output of length 77
+        return self(text)
+
+
+from transformers import T5Tokenizer, T5EncoderModel, CLIPTokenizer, CLIPTextModel
+
+def disabled_train(self, mode=True):
+    """Overwrite model.train with this function to make sure train/eval mode
+    does not change anymore."""
+    return self
+
+
+class FrozenT5Embedder(AbstractEncoder):
+    """Uses the T5 transformer encoder for text"""
+    def __init__(self, version="google/t5-v1_1-large", device="cuda", max_length=77):  # others are google/t5-v1_1-xl and google/t5-v1_1-xxl
+        super().__init__()
+        self.tokenizer = T5Tokenizer.from_pretrained(version)
+        self.transformer = T5EncoderModel.from_pretrained(version)
+        self.device = device
+        self.max_length = max_length   # TODO: typical value?
+        self.freeze()
+
+    def freeze(self):
+        self.transformer = self.transformer.eval()
+        #self.train = disabled_train
+        for param in self.parameters():
+            param.requires_grad = False
+
+    def forward(self, text):
+        batch_encoding = self.tokenizer(text, truncation=True, max_length=self.max_length, return_length=True,
+                                        return_overflowing_tokens=False, padding="max_length", return_tensors="pt")
+        tokens = batch_encoding["input_ids"].to(self.device)
+        outputs = self.transformer(input_ids=tokens)
+
+        z = outputs.last_hidden_state
+        return z
+
+    def encode(self, text):
+        return self(text)
+
+from ldm.thirdp.psp.id_loss import IDFeatures
+import kornia.augmentation as K
+
+class FrozenFaceEncoder(AbstractEncoder):
+    def __init__(self, model_path, augment=False):
+        super().__init__()
+        self.loss_fn = IDFeatures(model_path)
+        # face encoder is frozen
+        for p in self.loss_fn.parameters():
+            p.requires_grad = False
+        # Mapper is trainable
+        self.mapper = torch.nn.Linear(512, 768)
+        p = 0.25
+        if augment:
+            self.augment = K.AugmentationSequential(
+                K.RandomHorizontalFlip(p=0.5),
+                K.RandomEqualize(p=p),
+                # K.RandomPlanckianJitter(p=p),
+                # K.RandomPlasmaBrightness(p=p),
+                # K.RandomPlasmaContrast(p=p),
+                # K.ColorJiggle(0.02, 0.2, 0.2, p=p),
+            )
+        else:
+            self.augment = False
+
+    def forward(self, img):
+        if isinstance(img, list):
+            # Uncondition
+            return torch.zeros((1, 1, 768), device=self.mapper.weight.device)
+
+        if self.augment is not None:
+            # Transforms require 0-1
+            img = self.augment((img + 1)/2)
+            img = 2*img - 1
+
+        feat = self.loss_fn(img, crop=True)
+        feat = self.mapper(feat.unsqueeze(1))
+        return feat
+
+    def encode(self, img):
+        return self(img)
+
+class FrozenCLIPEmbedder(AbstractEncoder):
+    """Uses the CLIP transformer encoder for text (from huggingface)"""
+    def __init__(self, version="openai/clip-vit-large-patch14", device="cuda", max_length=77):  # clip-vit-base-patch32
+        super().__init__()
+        self.tokenizer = CLIPTokenizer.from_pretrained(version)
+        self.transformer = CLIPTextModel.from_pretrained(version)
+        self.device = device
+        self.max_length = max_length   # TODO: typical value?
+        self.freeze()
+
+    def freeze(self):
+        self.transformer = self.transformer.eval()
+        #self.train = disabled_train
+        for param in self.parameters():
+            param.requires_grad = False
+
+    def forward(self, text):
+        batch_encoding = self.tokenizer(text, truncation=True, max_length=self.max_length, return_length=True,
+                                        return_overflowing_tokens=False, padding="max_length", return_tensors="pt")
+        tokens = batch_encoding["input_ids"].to(self.device)
+        outputs = self.transformer(input_ids=tokens)
+
+        z = outputs.last_hidden_state
+        return z
+
+    def encode(self, text):
+        return self(text)
+
+import torch.nn.functional as F
+from transformers import CLIPVisionModel
+class ClipImageProjector(AbstractEncoder):
+    """
+        Uses the CLIP image encoder.
+        """
+    def __init__(self, version="openai/clip-vit-large-patch14", max_length=77):  # clip-vit-base-patch32
+        super().__init__()
+        self.model = CLIPVisionModel.from_pretrained(version)
+        self.model.train()
+        self.max_length = max_length   # TODO: typical value?
+        self.antialias = True
+        self.mapper = torch.nn.Linear(1024, 768)
+        self.register_buffer('mean', torch.Tensor([0.48145466, 0.4578275, 0.40821073]), persistent=False)
+        self.register_buffer('std', torch.Tensor([0.26862954, 0.26130258, 0.27577711]), persistent=False)
+        null_cond = self.get_null_cond(version, max_length)
+        self.register_buffer('null_cond', null_cond)
+
+    @torch.no_grad()
+    def get_null_cond(self, version, max_length):
+        device = self.mean.device
+        embedder = FrozenCLIPEmbedder(version=version, device=device, max_length=max_length)
+        null_cond = embedder([""])
+        return null_cond
+
+    def preprocess(self, x):
+        # Expects inputs in the range -1, 1
+        x = kornia.geometry.resize(x, (224, 224),
+                                   interpolation='bicubic',align_corners=True,
+                                   antialias=self.antialias)
+        x = (x + 1.) / 2.
+        # renormalize according to clip
+        x = kornia.enhance.normalize(x, self.mean, self.std)
+        return x
+
+    def forward(self, x):
+        if isinstance(x, list):
+            return self.null_cond
+        # x is assumed to be in range [-1,1]
+        x = self.preprocess(x)
+        outputs = self.model(pixel_values=x)
+        last_hidden_state = outputs.last_hidden_state
+        last_hidden_state = self.mapper(last_hidden_state)
+        return F.pad(last_hidden_state, [0,0, 0,self.max_length-last_hidden_state.shape[1], 0,0])
+
+    def encode(self, im):
+        return self(im)
+
+class ProjectedFrozenCLIPEmbedder(AbstractEncoder):
+    def __init__(self, version="openai/clip-vit-large-patch14", device="cuda", max_length=77):  # clip-vit-base-patch32
+        super().__init__()
+        self.embedder = FrozenCLIPEmbedder(version=version, device=device, max_length=max_length)
+        self.projection = torch.nn.Linear(768, 768)
+
+    def forward(self, text):
+        z = self.embedder(text)
+        return self.projection(z)
+
+    def encode(self, text):
+        return self(text)
+
+class FrozenCLIPImageEmbedder(AbstractEncoder):
+    """
+        Uses the CLIP image encoder.
+        Not actually frozen... If you want that set cond_stage_trainable=False in cfg
+        """
+    def __init__(
+            self,
+            model='ViT-L/14',
+            jit=False,
+            device='cpu',
+            antialias=False,
+        ):
+        super().__init__()
+        self.model, _ = clip.load(name=model, device=device, jit=jit)
+        # We don't use the text part so delete it
+        del self.model.transformer
+        self.antialias = antialias
+        self.register_buffer('mean', torch.Tensor([0.48145466, 0.4578275, 0.40821073]), persistent=False)
+        self.register_buffer('std', torch.Tensor([0.26862954, 0.26130258, 0.27577711]), persistent=False)
+
+    def preprocess(self, x):
+        # Expects inputs in the range -1, 1
+        x = kornia.geometry.resize(x, (224, 224),
+                                   interpolation='bicubic',align_corners=True,
+                                   antialias=self.antialias)
+        x = (x + 1.) / 2.
+        # renormalize according to clip
+        x = kornia.enhance.normalize(x, self.mean, self.std)
+        return x
+
+    def forward(self, x):
+        # x is assumed to be in range [-1,1]
+        if isinstance(x, list):
+            # [""] denotes condition dropout for ucg
+            device = self.model.visual.conv1.weight.device
+            return torch.zeros(1, 768, device=device)
+        return self.model.encode_image(self.preprocess(x)).float()
+
+    def encode(self, im):
+        return self(im).unsqueeze(1)
+
+from torchvision import transforms
+import random
+
+class FrozenCLIPImageMutliEmbedder(AbstractEncoder):
+    """
+        Uses the CLIP image encoder.
+        Not actually frozen... If you want that set cond_stage_trainable=False in cfg
+        """
+    def __init__(
+            self,
+            model='ViT-L/14',
+            jit=False,
+            device='cpu',
+            antialias=True,
+            max_crops=5,
+        ):
+        super().__init__()
+        self.model, _ = clip.load(name=model, device=device, jit=jit)
+        # We don't use the text part so delete it
+        del self.model.transformer
+        self.antialias = antialias
+        self.register_buffer('mean', torch.Tensor([0.48145466, 0.4578275, 0.40821073]), persistent=False)
+        self.register_buffer('std', torch.Tensor([0.26862954, 0.26130258, 0.27577711]), persistent=False)
+        self.max_crops = max_crops
+
+    def preprocess(self, x):
+
+        # Expects inputs in the range -1, 1
+        randcrop = transforms.RandomResizedCrop(224, scale=(0.085, 1.0), ratio=(1,1))
+        max_crops = self.max_crops
+        patches = []
+        crops = [randcrop(x) for _ in range(max_crops)]
+        patches.extend(crops)
+        x = torch.cat(patches, dim=0)
+        x = (x + 1.) / 2.
+        # renormalize according to clip
+        x = kornia.enhance.normalize(x, self.mean, self.std)
+        return x
+
+    def forward(self, x):
+        # x is assumed to be in range [-1,1]
+        if isinstance(x, list):
+            # [""] denotes condition dropout for ucg
+            device = self.model.visual.conv1.weight.device
+            return torch.zeros(1, self.max_crops, 768, device=device)
+        batch_tokens = []
+        for im in x:
+            patches = self.preprocess(im.unsqueeze(0))
+            tokens = self.model.encode_image(patches).float()
+            for t in tokens:
+                if random.random() < 0.1:
+                    t *= 0
+            batch_tokens.append(tokens.unsqueeze(0))
+
+        return torch.cat(batch_tokens, dim=0)
+
+    def encode(self, im):
+        return self(im)
+
+class SpatialRescaler(nn.Module):
+    def __init__(self,
+                 n_stages=1,
+                 method='bilinear',
+                 multiplier=0.5,
+                 in_channels=3,
+                 out_channels=None,
+                 bias=False):
+        super().__init__()
+        self.n_stages = n_stages
+        assert self.n_stages >= 0
+        assert method in ['nearest','linear','bilinear','trilinear','bicubic','area']
+        self.multiplier = multiplier
+        self.interpolator = partial(torch.nn.functional.interpolate, mode=method)
+        self.remap_output = out_channels is not None
+        if self.remap_output:
+            print(f'Spatial Rescaler mapping from {in_channels} to {out_channels} channels after resizing.')
+            self.channel_mapper = nn.Conv2d(in_channels,out_channels,1,bias=bias)
+
+    def forward(self,x):
+        for stage in range(self.n_stages):
+            x = self.interpolator(x, scale_factor=self.multiplier)
+
+
+        if self.remap_output:
+            x = self.channel_mapper(x)
+        return x
+
+    def encode(self, x):
+        return self(x)
+
+
+from ldm.util import instantiate_from_config
+from ldm.modules.diffusionmodules.util import make_beta_schedule, extract_into_tensor, noise_like
+
+
+class LowScaleEncoder(nn.Module):
+    def __init__(self, model_config, linear_start, linear_end, timesteps=1000, max_noise_level=250, output_size=64,
+                 scale_factor=1.0):
+        super().__init__()
+        self.max_noise_level = max_noise_level
+        self.model = instantiate_from_config(model_config)
+        self.augmentation_schedule = self.register_schedule(timesteps=timesteps, linear_start=linear_start,
+                                                            linear_end=linear_end)
+        self.out_size = output_size
+        self.scale_factor = scale_factor
+
+    def register_schedule(self, beta_schedule="linear", timesteps=1000,
+                          linear_start=1e-4, linear_end=2e-2, cosine_s=8e-3):
+        betas = make_beta_schedule(beta_schedule, timesteps, linear_start=linear_start, linear_end=linear_end,
+                                   cosine_s=cosine_s)
+        alphas = 1. - betas
+        alphas_cumprod = np.cumprod(alphas, axis=0)
+        alphas_cumprod_prev = np.append(1., alphas_cumprod[:-1])
+
+        timesteps, = betas.shape
+        self.num_timesteps = int(timesteps)
+        self.linear_start = linear_start
+        self.linear_end = linear_end
+        assert alphas_cumprod.shape[0] == self.num_timesteps, 'alphas have to be defined for each timestep'
+
+        to_torch = partial(torch.tensor, dtype=torch.float32)
+
+        self.register_buffer('betas', to_torch(betas))
+        self.register_buffer('alphas_cumprod', to_torch(alphas_cumprod))
+        self.register_buffer('alphas_cumprod_prev', to_torch(alphas_cumprod_prev))
+
+        # calculations for diffusion q(x_t | x_{t-1}) and others
+        self.register_buffer('sqrt_alphas_cumprod', to_torch(np.sqrt(alphas_cumprod)))
+        self.register_buffer('sqrt_one_minus_alphas_cumprod', to_torch(np.sqrt(1. - alphas_cumprod)))
+        self.register_buffer('log_one_minus_alphas_cumprod', to_torch(np.log(1. - alphas_cumprod)))
+        self.register_buffer('sqrt_recip_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod)))
+        self.register_buffer('sqrt_recipm1_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod - 1)))
+
+    def q_sample(self, x_start, t, noise=None):
+        noise = default(noise, lambda: torch.randn_like(x_start))
+        return (extract_into_tensor(self.sqrt_alphas_cumprod, t, x_start.shape) * x_start +
+                extract_into_tensor(self.sqrt_one_minus_alphas_cumprod, t, x_start.shape) * noise)
+
+    def forward(self, x):
+        z = self.model.encode(x).sample()
+        z = z * self.scale_factor
+        noise_level = torch.randint(0, self.max_noise_level, (x.shape[0],), device=x.device).long()
+        z = self.q_sample(z, noise_level)
+        if self.out_size is not None:
+            z = torch.nn.functional.interpolate(z, size=self.out_size, mode="nearest")  # TODO: experiment with mode
+        # z = z.repeat_interleave(2, -2).repeat_interleave(2, -1)
+        return z, noise_level
+
+    def decode(self, z):
+        z = z / self.scale_factor
+        return self.model.decode(z)
+
+
+if __name__ == "__main__":
+    from ldm.util import count_params
+    sentences = ["a hedgehog drinking a whiskey", "der mond ist aufgegangen", "Ein Satz mit vielen Sonderzeichen: äöü ß ?! : 'xx-y/@s'"]
+    model = FrozenT5Embedder(version="google/t5-v1_1-xl").cuda()
+    count_params(model, True)
+    z = model(sentences)
+    print(z.shape)
+
+    model = FrozenCLIPEmbedder().cuda()
+    count_params(model, True)
+    z = model(sentences)
+    print(z.shape)
+
+    print("done.")

ldm/modules/evaluate/adm_evaluator.py

@@ -0,0 +1,676 @@
+import argparse
+import io
+import os
+import random
+import warnings
+import zipfile
+from abc import ABC, abstractmethod
+from contextlib import contextmanager
+from functools import partial
+from multiprocessing import cpu_count
+from multiprocessing.pool import ThreadPool
+from typing import Iterable, Optional, Tuple
+import yaml
+
+import numpy as np
+import requests
+import tensorflow.compat.v1 as tf
+from scipy import linalg
+from tqdm.auto import tqdm
+
+INCEPTION_V3_URL = "https://openaipublic.blob.core.windows.net/diffusion/jul-2021/ref_batches/classify_image_graph_def.pb"
+INCEPTION_V3_PATH = "classify_image_graph_def.pb"
+
+FID_POOL_NAME = "pool_3:0"
+FID_SPATIAL_NAME = "mixed_6/conv:0"
+
+REQUIREMENTS = f"This script has the following requirements: \n" \
+               'tensorflow-gpu>=2.0' + "\n" + 'scipy' + "\n" + "requests" + "\n" + "tqdm"
+
+
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--ref_batch", help="path to reference batch npz file")
+    parser.add_argument("--sample_batch", help="path to sample batch npz file")
+    args = parser.parse_args()
+
+    config = tf.ConfigProto(
+        allow_soft_placement=True  # allows DecodeJpeg to run on CPU in Inception graph
+    )
+    config.gpu_options.allow_growth = True
+    evaluator = Evaluator(tf.Session(config=config))
+
+    print("warming up TensorFlow...")
+    # This will cause TF to print a bunch of verbose stuff now rather
+    # than after the next print(), to help prevent confusion.
+    evaluator.warmup()
+
+    print("computing reference batch activations...")
+    ref_acts = evaluator.read_activations(args.ref_batch)
+    print("computing/reading reference batch statistics...")
+    ref_stats, ref_stats_spatial = evaluator.read_statistics(args.ref_batch, ref_acts)
+
+    print("computing sample batch activations...")
+    sample_acts = evaluator.read_activations(args.sample_batch)
+    print("computing/reading sample batch statistics...")
+    sample_stats, sample_stats_spatial = evaluator.read_statistics(args.sample_batch, sample_acts)
+
+    print("Computing evaluations...")
+    is_ = evaluator.compute_inception_score(sample_acts[0])
+    print("Inception Score:", is_)
+    fid = sample_stats.frechet_distance(ref_stats)
+    print("FID:", fid)
+    sfid = sample_stats_spatial.frechet_distance(ref_stats_spatial)
+    print("sFID:", sfid)
+    prec, recall = evaluator.compute_prec_recall(ref_acts[0], sample_acts[0])
+    print("Precision:", prec)
+    print("Recall:", recall)
+
+    savepath = '/'.join(args.sample_batch.split('/')[:-1])
+    results_file = os.path.join(savepath,'evaluation_metrics.yaml')
+    print(f'Saving evaluation results to "{results_file}"')
+
+    results = {
+        'IS': is_,
+        'FID': fid,
+        'sFID': sfid,
+        'Precision:':prec,
+        'Recall': recall
+    }
+
+    with open(results_file, 'w') as f:
+        yaml.dump(results, f, default_flow_style=False)
+
+class InvalidFIDException(Exception):
+    pass
+
+
+class FIDStatistics:
+    def __init__(self, mu: np.ndarray, sigma: np.ndarray):
+        self.mu = mu
+        self.sigma = sigma
+
+    def frechet_distance(self, other, eps=1e-6):
+        """
+        Compute the Frechet distance between two sets of statistics.
+        """
+        # https://github.com/bioinf-jku/TTUR/blob/73ab375cdf952a12686d9aa7978567771084da42/fid.py#L132
+        mu1, sigma1 = self.mu, self.sigma
+        mu2, sigma2 = other.mu, other.sigma
+
+        mu1 = np.atleast_1d(mu1)
+        mu2 = np.atleast_1d(mu2)
+
+        sigma1 = np.atleast_2d(sigma1)
+        sigma2 = np.atleast_2d(sigma2)
+
+        assert (
+            mu1.shape == mu2.shape
+        ), f"Training and test mean vectors have different lengths: {mu1.shape}, {mu2.shape}"
+        assert (
+            sigma1.shape == sigma2.shape
+        ), f"Training and test covariances have different dimensions: {sigma1.shape}, {sigma2.shape}"
+
+        diff = mu1 - mu2
+
+        # product might be almost singular
+        covmean, _ = linalg.sqrtm(sigma1.dot(sigma2), disp=False)
+        if not np.isfinite(covmean).all():
+            msg = (
+                "fid calculation produces singular product; adding %s to diagonal of cov estimates"
+                % eps
+            )
+            warnings.warn(msg)
+            offset = np.eye(sigma1.shape[0]) * eps
+            covmean = linalg.sqrtm((sigma1 + offset).dot(sigma2 + offset))
+
+        # numerical error might give slight imaginary component
+        if np.iscomplexobj(covmean):
+            if not np.allclose(np.diagonal(covmean).imag, 0, atol=1e-3):
+                m = np.max(np.abs(covmean.imag))
+                raise ValueError("Imaginary component {}".format(m))
+            covmean = covmean.real
+
+        tr_covmean = np.trace(covmean)
+
+        return diff.dot(diff) + np.trace(sigma1) + np.trace(sigma2) - 2 * tr_covmean
+
+
+class Evaluator:
+    def __init__(
+        self,
+        session,
+        batch_size=64,
+        softmax_batch_size=512,
+    ):
+        self.sess = session
+        self.batch_size = batch_size
+        self.softmax_batch_size = softmax_batch_size
+        self.manifold_estimator = ManifoldEstimator(session)
+        with self.sess.graph.as_default():
+            self.image_input = tf.placeholder(tf.float32, shape=[None, None, None, 3])
+            self.softmax_input = tf.placeholder(tf.float32, shape=[None, 2048])
+            self.pool_features, self.spatial_features = _create_feature_graph(self.image_input)
+            self.softmax = _create_softmax_graph(self.softmax_input)
+
+    def warmup(self):
+        self.compute_activations(np.zeros([1, 8, 64, 64, 3]))
+
+    def read_activations(self, npz_path: str) -> Tuple[np.ndarray, np.ndarray]:
+        with open_npz_array(npz_path, "arr_0") as reader:
+            return self.compute_activations(reader.read_batches(self.batch_size))
+
+    def compute_activations(self, batches: Iterable[np.ndarray],silent=False) -> Tuple[np.ndarray, np.ndarray]:
+        """
+        Compute image features for downstream evals.
+
+        :param batches: a iterator over NHWC numpy arrays in [0, 255].
+        :return: a tuple of numpy arrays of shape [N x X], where X is a feature
+                 dimension. The tuple is (pool_3, spatial).
+        """
+        preds = []
+        spatial_preds = []
+        it = batches if silent else tqdm(batches)
+        for batch in it:
+            batch = batch.astype(np.float32)
+            pred, spatial_pred = self.sess.run(
+                [self.pool_features, self.spatial_features], {self.image_input: batch}
+            )
+            preds.append(pred.reshape([pred.shape[0], -1]))
+            spatial_preds.append(spatial_pred.reshape([spatial_pred.shape[0], -1]))
+        return (
+            np.concatenate(preds, axis=0),
+            np.concatenate(spatial_preds, axis=0),
+        )
+
+    def read_statistics(
+        self, npz_path: str, activations: Tuple[np.ndarray, np.ndarray]
+    ) -> Tuple[FIDStatistics, FIDStatistics]:
+        obj = np.load(npz_path)
+        if "mu" in list(obj.keys()):
+            return FIDStatistics(obj["mu"], obj["sigma"]), FIDStatistics(
+                obj["mu_s"], obj["sigma_s"]
+            )
+        return tuple(self.compute_statistics(x) for x in activations)
+
+    def compute_statistics(self, activations: np.ndarray) -> FIDStatistics:
+        mu = np.mean(activations, axis=0)
+        sigma = np.cov(activations, rowvar=False)
+        return FIDStatistics(mu, sigma)
+
+    def compute_inception_score(self, activations: np.ndarray, split_size: int = 5000) -> float:
+        softmax_out = []
+        for i in range(0, len(activations), self.softmax_batch_size):
+            acts = activations[i : i + self.softmax_batch_size]
+            softmax_out.append(self.sess.run(self.softmax, feed_dict={self.softmax_input: acts}))
+        preds = np.concatenate(softmax_out, axis=0)
+        # https://github.com/openai/improved-gan/blob/4f5d1ec5c16a7eceb206f42bfc652693601e1d5c/inception_score/model.py#L46
+        scores = []
+        for i in range(0, len(preds), split_size):
+            part = preds[i : i + split_size]
+            kl = part * (np.log(part) - np.log(np.expand_dims(np.mean(part, 0), 0)))
+            kl = np.mean(np.sum(kl, 1))
+            scores.append(np.exp(kl))
+        return float(np.mean(scores))
+
+    def compute_prec_recall(
+        self, activations_ref: np.ndarray, activations_sample: np.ndarray
+    ) -> Tuple[float, float]:
+        radii_1 = self.manifold_estimator.manifold_radii(activations_ref)
+        radii_2 = self.manifold_estimator.manifold_radii(activations_sample)
+        pr = self.manifold_estimator.evaluate_pr(
+            activations_ref, radii_1, activations_sample, radii_2
+        )
+        return (float(pr[0][0]), float(pr[1][0]))
+
+
+class ManifoldEstimator:
+    """
+    A helper for comparing manifolds of feature vectors.
+
+    Adapted from https://github.com/kynkaat/improved-precision-and-recall-metric/blob/f60f25e5ad933a79135c783fcda53de30f42c9b9/precision_recall.py#L57
+    """
+
+    def __init__(
+        self,
+        session,
+        row_batch_size=10000,
+        col_batch_size=10000,
+        nhood_sizes=(3,),
+        clamp_to_percentile=None,
+        eps=1e-5,
+    ):
+        """
+        Estimate the manifold of given feature vectors.
+
+        :param session: the TensorFlow session.
+        :param row_batch_size: row batch size to compute pairwise distances
+                               (parameter to trade-off between memory usage and performance).
+        :param col_batch_size: column batch size to compute pairwise distances.
+        :param nhood_sizes: number of neighbors used to estimate the manifold.
+        :param clamp_to_percentile: prune hyperspheres that have radius larger than
+                                    the given percentile.
+        :param eps: small number for numerical stability.
+        """
+        self.distance_block = DistanceBlock(session)
+        self.row_batch_size = row_batch_size
+        self.col_batch_size = col_batch_size
+        self.nhood_sizes = nhood_sizes
+        self.num_nhoods = len(nhood_sizes)
+        self.clamp_to_percentile = clamp_to_percentile
+        self.eps = eps
+
+    def warmup(self):
+        feats, radii = (
+            np.zeros([1, 2048], dtype=np.float32),
+            np.zeros([1, 1], dtype=np.float32),
+        )
+        self.evaluate_pr(feats, radii, feats, radii)
+
+    def manifold_radii(self, features: np.ndarray) -> np.ndarray:
+        num_images = len(features)
+
+        # Estimate manifold of features by calculating distances to k-NN of each sample.
+        radii = np.zeros([num_images, self.num_nhoods], dtype=np.float32)
+        distance_batch = np.zeros([self.row_batch_size, num_images], dtype=np.float32)
+        seq = np.arange(max(self.nhood_sizes) + 1, dtype=np.int32)
+
+        for begin1 in range(0, num_images, self.row_batch_size):
+            end1 = min(begin1 + self.row_batch_size, num_images)
+            row_batch = features[begin1:end1]
+
+            for begin2 in range(0, num_images, self.col_batch_size):
+                end2 = min(begin2 + self.col_batch_size, num_images)
+                col_batch = features[begin2:end2]
+
+                # Compute distances between batches.
+                distance_batch[
+                    0 : end1 - begin1, begin2:end2
+                ] = self.distance_block.pairwise_distances(row_batch, col_batch)
+
+            # Find the k-nearest neighbor from the current batch.
+            radii[begin1:end1, :] = np.concatenate(
+                [
+                    x[:, self.nhood_sizes]
+                    for x in _numpy_partition(distance_batch[0 : end1 - begin1, :], seq, axis=1)
+                ],
+                axis=0,
+            )
+
+        if self.clamp_to_percentile is not None:
+            max_distances = np.percentile(radii, self.clamp_to_percentile, axis=0)
+            radii[radii > max_distances] = 0
+        return radii
+
+    def evaluate(self, features: np.ndarray, radii: np.ndarray, eval_features: np.ndarray):
+        """
+        Evaluate if new feature vectors are at the manifold.
+        """
+        num_eval_images = eval_features.shape[0]
+        num_ref_images = radii.shape[0]
+        distance_batch = np.zeros([self.row_batch_size, num_ref_images], dtype=np.float32)
+        batch_predictions = np.zeros([num_eval_images, self.num_nhoods], dtype=np.int32)
+        max_realism_score = np.zeros([num_eval_images], dtype=np.float32)
+        nearest_indices = np.zeros([num_eval_images], dtype=np.int32)
+
+        for begin1 in range(0, num_eval_images, self.row_batch_size):
+            end1 = min(begin1 + self.row_batch_size, num_eval_images)
+            feature_batch = eval_features[begin1:end1]
+
+            for begin2 in range(0, num_ref_images, self.col_batch_size):
+                end2 = min(begin2 + self.col_batch_size, num_ref_images)
+                ref_batch = features[begin2:end2]
+
+                distance_batch[
+                    0 : end1 - begin1, begin2:end2
+                ] = self.distance_block.pairwise_distances(feature_batch, ref_batch)
+
+            # From the minibatch of new feature vectors, determine if they are in the estimated manifold.
+            # If a feature vector is inside a hypersphere of some reference sample, then
+            # the new sample lies at the estimated manifold.
+            # The radii of the hyperspheres are determined from distances of neighborhood size k.
+            samples_in_manifold = distance_batch[0 : end1 - begin1, :, None] <= radii
+            batch_predictions[begin1:end1] = np.any(samples_in_manifold, axis=1).astype(np.int32)
+
+            max_realism_score[begin1:end1] = np.max(
+                radii[:, 0] / (distance_batch[0 : end1 - begin1, :] + self.eps), axis=1
+            )
+            nearest_indices[begin1:end1] = np.argmin(distance_batch[0 : end1 - begin1, :], axis=1)
+
+        return {
+            "fraction": float(np.mean(batch_predictions)),
+            "batch_predictions": batch_predictions,
+            "max_realisim_score": max_realism_score,
+            "nearest_indices": nearest_indices,
+        }
+
+    def evaluate_pr(
+        self,
+        features_1: np.ndarray,
+        radii_1: np.ndarray,
+        features_2: np.ndarray,
+        radii_2: np.ndarray,
+    ) -> Tuple[np.ndarray, np.ndarray]:
+        """
+        Evaluate precision and recall efficiently.
+
+        :param features_1: [N1 x D] feature vectors for reference batch.
+        :param radii_1: [N1 x K1] radii for reference vectors.
+        :param features_2: [N2 x D] feature vectors for the other batch.
+        :param radii_2: [N x K2] radii for other vectors.
+        :return: a tuple of arrays for (precision, recall):
+                 - precision: an np.ndarray of length K1
+                 - recall: an np.ndarray of length K2
+        """
+        features_1_status = np.zeros([len(features_1), radii_2.shape[1]], dtype=np.bool)
+        features_2_status = np.zeros([len(features_2), radii_1.shape[1]], dtype=np.bool)
+        for begin_1 in range(0, len(features_1), self.row_batch_size):
+            end_1 = begin_1 + self.row_batch_size
+            batch_1 = features_1[begin_1:end_1]
+            for begin_2 in range(0, len(features_2), self.col_batch_size):
+                end_2 = begin_2 + self.col_batch_size
+                batch_2 = features_2[begin_2:end_2]
+                batch_1_in, batch_2_in = self.distance_block.less_thans(
+                    batch_1, radii_1[begin_1:end_1], batch_2, radii_2[begin_2:end_2]
+                )
+                features_1_status[begin_1:end_1] |= batch_1_in
+                features_2_status[begin_2:end_2] |= batch_2_in
+        return (
+            np.mean(features_2_status.astype(np.float64), axis=0),
+            np.mean(features_1_status.astype(np.float64), axis=0),
+        )
+
+
+class DistanceBlock:
+    """
+    Calculate pairwise distances between vectors.
+
+    Adapted from https://github.com/kynkaat/improved-precision-and-recall-metric/blob/f60f25e5ad933a79135c783fcda53de30f42c9b9/precision_recall.py#L34
+    """
+
+    def __init__(self, session):
+        self.session = session
+
+        # Initialize TF graph to calculate pairwise distances.
+        with session.graph.as_default():
+            self._features_batch1 = tf.placeholder(tf.float32, shape=[None, None])
+            self._features_batch2 = tf.placeholder(tf.float32, shape=[None, None])
+            distance_block_16 = _batch_pairwise_distances(
+                tf.cast(self._features_batch1, tf.float16),
+                tf.cast(self._features_batch2, tf.float16),
+            )
+            self.distance_block = tf.cond(
+                tf.reduce_all(tf.math.is_finite(distance_block_16)),
+                lambda: tf.cast(distance_block_16, tf.float32),
+                lambda: _batch_pairwise_distances(self._features_batch1, self._features_batch2),
+            )
+
+            # Extra logic for less thans.
+            self._radii1 = tf.placeholder(tf.float32, shape=[None, None])
+            self._radii2 = tf.placeholder(tf.float32, shape=[None, None])
+            dist32 = tf.cast(self.distance_block, tf.float32)[..., None]
+            self._batch_1_in = tf.math.reduce_any(dist32 <= self._radii2, axis=1)
+            self._batch_2_in = tf.math.reduce_any(dist32 <= self._radii1[:, None], axis=0)
+
+    def pairwise_distances(self, U, V):
+        """
+        Evaluate pairwise distances between two batches of feature vectors.
+        """
+        return self.session.run(
+            self.distance_block,
+            feed_dict={self._features_batch1: U, self._features_batch2: V},
+        )
+
+    def less_thans(self, batch_1, radii_1, batch_2, radii_2):
+        return self.session.run(
+            [self._batch_1_in, self._batch_2_in],
+            feed_dict={
+                self._features_batch1: batch_1,
+                self._features_batch2: batch_2,
+                self._radii1: radii_1,
+                self._radii2: radii_2,
+            },
+        )
+
+
+def _batch_pairwise_distances(U, V):
+    """
+    Compute pairwise distances between two batches of feature vectors.
+    """
+    with tf.variable_scope("pairwise_dist_block"):
+        # Squared norms of each row in U and V.
+        norm_u = tf.reduce_sum(tf.square(U), 1)
+        norm_v = tf.reduce_sum(tf.square(V), 1)
+
+        # norm_u as a column and norm_v as a row vectors.
+        norm_u = tf.reshape(norm_u, [-1, 1])
+        norm_v = tf.reshape(norm_v, [1, -1])
+
+        # Pairwise squared Euclidean distances.
+        D = tf.maximum(norm_u - 2 * tf.matmul(U, V, False, True) + norm_v, 0.0)
+
+    return D
+
+
+class NpzArrayReader(ABC):
+    @abstractmethod
+    def read_batch(self, batch_size: int) -> Optional[np.ndarray]:
+        pass
+
+    @abstractmethod
+    def remaining(self) -> int:
+        pass
+
+    def read_batches(self, batch_size: int) -> Iterable[np.ndarray]:
+        def gen_fn():
+            while True:
+                batch = self.read_batch(batch_size)
+                if batch is None:
+                    break
+                yield batch
+
+        rem = self.remaining()
+        num_batches = rem // batch_size + int(rem % batch_size != 0)
+        return BatchIterator(gen_fn, num_batches)
+
+
+class BatchIterator:
+    def __init__(self, gen_fn, length):
+        self.gen_fn = gen_fn
+        self.length = length
+
+    def __len__(self):
+        return self.length
+
+    def __iter__(self):
+        return self.gen_fn()
+
+
+class StreamingNpzArrayReader(NpzArrayReader):
+    def __init__(self, arr_f, shape, dtype):
+        self.arr_f = arr_f
+        self.shape = shape
+        self.dtype = dtype
+        self.idx = 0
+
+    def read_batch(self, batch_size: int) -> Optional[np.ndarray]:
+        if self.idx >= self.shape[0]:
+            return None
+
+        bs = min(batch_size, self.shape[0] - self.idx)
+        self.idx += bs
+
+        if self.dtype.itemsize == 0:
+            return np.ndarray([bs, *self.shape[1:]], dtype=self.dtype)
+
+        read_count = bs * np.prod(self.shape[1:])
+        read_size = int(read_count * self.dtype.itemsize)
+        data = _read_bytes(self.arr_f, read_size, "array data")
+        return np.frombuffer(data, dtype=self.dtype).reshape([bs, *self.shape[1:]])
+
+    def remaining(self) -> int:
+        return max(0, self.shape[0] - self.idx)
+
+
+class MemoryNpzArrayReader(NpzArrayReader):
+    def __init__(self, arr):
+        self.arr = arr
+        self.idx = 0
+
+    @classmethod
+    def load(cls, path: str, arr_name: str):
+        with open(path, "rb") as f:
+            arr = np.load(f)[arr_name]
+        return cls(arr)
+
+    def read_batch(self, batch_size: int) -> Optional[np.ndarray]:
+        if self.idx >= self.arr.shape[0]:
+            return None
+
+        res = self.arr[self.idx : self.idx + batch_size]
+        self.idx += batch_size
+        return res
+
+    def remaining(self) -> int:
+        return max(0, self.arr.shape[0] - self.idx)
+
+
+@contextmanager
+def open_npz_array(path: str, arr_name: str) -> NpzArrayReader:
+    with _open_npy_file(path, arr_name) as arr_f:
+        version = np.lib.format.read_magic(arr_f)
+        if version == (1, 0):
+            header = np.lib.format.read_array_header_1_0(arr_f)
+        elif version == (2, 0):
+            header = np.lib.format.read_array_header_2_0(arr_f)
+        else:
+            yield MemoryNpzArrayReader.load(path, arr_name)
+            return
+        shape, fortran, dtype = header
+        if fortran or dtype.hasobject:
+            yield MemoryNpzArrayReader.load(path, arr_name)
+        else:
+            yield StreamingNpzArrayReader(arr_f, shape, dtype)
+
+
+def _read_bytes(fp, size, error_template="ran out of data"):
+    """
+    Copied from: https://github.com/numpy/numpy/blob/fb215c76967739268de71aa4bda55dd1b062bc2e/numpy/lib/format.py#L788-L886
+
+    Read from file-like object until size bytes are read.
+    Raises ValueError if not EOF is encountered before size bytes are read.
+    Non-blocking objects only supported if they derive from io objects.
+    Required as e.g. ZipExtFile in python 2.6 can return less data than
+    requested.
+    """
+    data = bytes()
+    while True:
+        # io files (default in python3) return None or raise on
+        # would-block, python2 file will truncate, probably nothing can be
+        # done about that.  note that regular files can't be non-blocking
+        try:
+            r = fp.read(size - len(data))
+            data += r
+            if len(r) == 0 or len(data) == size:
+                break
+        except io.BlockingIOError:
+            pass
+    if len(data) != size:
+        msg = "EOF: reading %s, expected %d bytes got %d"
+        raise ValueError(msg % (error_template, size, len(data)))
+    else:
+        return data
+
+
+@contextmanager
+def _open_npy_file(path: str, arr_name: str):
+    with open(path, "rb") as f:
+        with zipfile.ZipFile(f, "r") as zip_f:
+            if f"{arr_name}.npy" not in zip_f.namelist():
+                raise ValueError(f"missing {arr_name} in npz file")
+            with zip_f.open(f"{arr_name}.npy", "r") as arr_f:
+                yield arr_f
+
+
+def _download_inception_model():
+    if os.path.exists(INCEPTION_V3_PATH):
+        return
+    print("downloading InceptionV3 model...")
+    with requests.get(INCEPTION_V3_URL, stream=True) as r:
+        r.raise_for_status()
+        tmp_path = INCEPTION_V3_PATH + ".tmp"
+        with open(tmp_path, "wb") as f:
+            for chunk in tqdm(r.iter_content(chunk_size=8192)):
+                f.write(chunk)
+        os.rename(tmp_path, INCEPTION_V3_PATH)
+
+
+def _create_feature_graph(input_batch):
+    _download_inception_model()
+    prefix = f"{random.randrange(2**32)}_{random.randrange(2**32)}"
+    with open(INCEPTION_V3_PATH, "rb") as f:
+        graph_def = tf.GraphDef()
+        graph_def.ParseFromString(f.read())
+    pool3, spatial = tf.import_graph_def(
+        graph_def,
+        input_map={f"ExpandDims:0": input_batch},
+        return_elements=[FID_POOL_NAME, FID_SPATIAL_NAME],
+        name=prefix,
+    )
+    _update_shapes(pool3)
+    spatial = spatial[..., :7]
+    return pool3, spatial
+
+
+def _create_softmax_graph(input_batch):
+    _download_inception_model()
+    prefix = f"{random.randrange(2**32)}_{random.randrange(2**32)}"
+    with open(INCEPTION_V3_PATH, "rb") as f:
+        graph_def = tf.GraphDef()
+        graph_def.ParseFromString(f.read())
+    (matmul,) = tf.import_graph_def(
+        graph_def, return_elements=[f"softmax/logits/MatMul"], name=prefix
+    )
+    w = matmul.inputs[1]
+    logits = tf.matmul(input_batch, w)
+    return tf.nn.softmax(logits)
+
+
+def _update_shapes(pool3):
+    # https://github.com/bioinf-jku/TTUR/blob/73ab375cdf952a12686d9aa7978567771084da42/fid.py#L50-L63
+    ops = pool3.graph.get_operations()
+    for op in ops:
+        for o in op.outputs:
+            shape = o.get_shape()
+            if shape._dims is not None:  # pylint: disable=protected-access
+                # shape = [s.value for s in shape] TF 1.x
+                shape = [s for s in shape]  # TF 2.x
+                new_shape = []
+                for j, s in enumerate(shape):
+                    if s == 1 and j == 0:
+                        new_shape.append(None)
+                    else:
+                        new_shape.append(s)
+                o.__dict__["_shape_val"] = tf.TensorShape(new_shape)
+    return pool3
+
+
+def _numpy_partition(arr, kth, **kwargs):
+    num_workers = min(cpu_count(), len(arr))
+    chunk_size = len(arr) // num_workers
+    extra = len(arr) % num_workers
+
+    start_idx = 0
+    batches = []
+    for i in range(num_workers):
+        size = chunk_size + (1 if i < extra else 0)
+        batches.append(arr[start_idx : start_idx + size])
+        start_idx += size
+
+    with ThreadPool(num_workers) as pool:
+        return list(pool.map(partial(np.partition, kth=kth, **kwargs), batches))
+
+
+if __name__ == "__main__":
+    print(REQUIREMENTS)
+    main()

ldm/modules/evaluate/evaluate_perceptualsim.py

@@ -0,0 +1,630 @@
+import argparse
+import glob
+import os
+from tqdm import tqdm
+from collections import namedtuple
+
+import numpy as np
+import torch
+import torchvision.transforms as transforms
+from torchvision import models
+from PIL import Image
+
+from ldm.modules.evaluate.ssim import ssim
+
+
+transform = transforms.Compose([transforms.ToTensor()])
+
+def normalize_tensor(in_feat, eps=1e-10):
+    norm_factor = torch.sqrt(torch.sum(in_feat ** 2, dim=1)).view(
+        in_feat.size()[0], 1, in_feat.size()[2], in_feat.size()[3]
+    )
+    return in_feat / (norm_factor.expand_as(in_feat) + eps)
+
+
+def cos_sim(in0, in1):
+    in0_norm = normalize_tensor(in0)
+    in1_norm = normalize_tensor(in1)
+    N = in0.size()[0]
+    X = in0.size()[2]
+    Y = in0.size()[3]
+
+    return torch.mean(
+        torch.mean(
+            torch.sum(in0_norm * in1_norm, dim=1).view(N, 1, X, Y), dim=2
+        ).view(N, 1, 1, Y),
+        dim=3,
+    ).view(N)
+
+
+class squeezenet(torch.nn.Module):
+    def __init__(self, requires_grad=False, pretrained=True):
+        super(squeezenet, self).__init__()
+        pretrained_features = models.squeezenet1_1(
+            pretrained=pretrained
+        ).features
+        self.slice1 = torch.nn.Sequential()
+        self.slice2 = torch.nn.Sequential()
+        self.slice3 = torch.nn.Sequential()
+        self.slice4 = torch.nn.Sequential()
+        self.slice5 = torch.nn.Sequential()
+        self.slice6 = torch.nn.Sequential()
+        self.slice7 = torch.nn.Sequential()
+        self.N_slices = 7
+        for x in range(2):
+            self.slice1.add_module(str(x), pretrained_features[x])
+        for x in range(2, 5):
+            self.slice2.add_module(str(x), pretrained_features[x])
+        for x in range(5, 8):
+            self.slice3.add_module(str(x), pretrained_features[x])
+        for x in range(8, 10):
+            self.slice4.add_module(str(x), pretrained_features[x])
+        for x in range(10, 11):
+            self.slice5.add_module(str(x), pretrained_features[x])
+        for x in range(11, 12):
+            self.slice6.add_module(str(x), pretrained_features[x])
+        for x in range(12, 13):
+            self.slice7.add_module(str(x), pretrained_features[x])
+        if not requires_grad:
+            for param in self.parameters():
+                param.requires_grad = False
+
+    def forward(self, X):
+        h = self.slice1(X)
+        h_relu1 = h
+        h = self.slice2(h)
+        h_relu2 = h
+        h = self.slice3(h)
+        h_relu3 = h
+        h = self.slice4(h)
+        h_relu4 = h
+        h = self.slice5(h)
+        h_relu5 = h
+        h = self.slice6(h)
+        h_relu6 = h
+        h = self.slice7(h)
+        h_relu7 = h
+        vgg_outputs = namedtuple(
+            "SqueezeOutputs",
+            ["relu1", "relu2", "relu3", "relu4", "relu5", "relu6", "relu7"],
+        )
+        out = vgg_outputs(
+            h_relu1, h_relu2, h_relu3, h_relu4, h_relu5, h_relu6, h_relu7
+        )
+
+        return out
+
+
+class alexnet(torch.nn.Module):
+    def __init__(self, requires_grad=False, pretrained=True):
+        super(alexnet, self).__init__()
+        alexnet_pretrained_features = models.alexnet(
+            pretrained=pretrained
+        ).features
+        self.slice1 = torch.nn.Sequential()
+        self.slice2 = torch.nn.Sequential()
+        self.slice3 = torch.nn.Sequential()
+        self.slice4 = torch.nn.Sequential()
+        self.slice5 = torch.nn.Sequential()
+        self.N_slices = 5
+        for x in range(2):
+            self.slice1.add_module(str(x), alexnet_pretrained_features[x])
+        for x in range(2, 5):
+            self.slice2.add_module(str(x), alexnet_pretrained_features[x])
+        for x in range(5, 8):
+            self.slice3.add_module(str(x), alexnet_pretrained_features[x])
+        for x in range(8, 10):
+            self.slice4.add_module(str(x), alexnet_pretrained_features[x])
+        for x in range(10, 12):
+            self.slice5.add_module(str(x), alexnet_pretrained_features[x])
+        if not requires_grad:
+            for param in self.parameters():
+                param.requires_grad = False
+
+    def forward(self, X):
+        h = self.slice1(X)
+        h_relu1 = h
+        h = self.slice2(h)
+        h_relu2 = h
+        h = self.slice3(h)
+        h_relu3 = h
+        h = self.slice4(h)
+        h_relu4 = h
+        h = self.slice5(h)
+        h_relu5 = h
+        alexnet_outputs = namedtuple(
+            "AlexnetOutputs", ["relu1", "relu2", "relu3", "relu4", "relu5"]
+        )
+        out = alexnet_outputs(h_relu1, h_relu2, h_relu3, h_relu4, h_relu5)
+
+        return out
+
+
+class vgg16(torch.nn.Module):
+    def __init__(self, requires_grad=False, pretrained=True):
+        super(vgg16, self).__init__()
+        vgg_pretrained_features = models.vgg16(pretrained=pretrained).features
+        self.slice1 = torch.nn.Sequential()
+        self.slice2 = torch.nn.Sequential()
+        self.slice3 = torch.nn.Sequential()
+        self.slice4 = torch.nn.Sequential()
+        self.slice5 = torch.nn.Sequential()
+        self.N_slices = 5
+        for x in range(4):
+            self.slice1.add_module(str(x), vgg_pretrained_features[x])
+        for x in range(4, 9):
+            self.slice2.add_module(str(x), vgg_pretrained_features[x])
+        for x in range(9, 16):
+            self.slice3.add_module(str(x), vgg_pretrained_features[x])
+        for x in range(16, 23):
+            self.slice4.add_module(str(x), vgg_pretrained_features[x])
+        for x in range(23, 30):
+            self.slice5.add_module(str(x), vgg_pretrained_features[x])
+        if not requires_grad:
+            for param in self.parameters():
+                param.requires_grad = False
+
+    def forward(self, X):
+        h = self.slice1(X)
+        h_relu1_2 = h
+        h = self.slice2(h)
+        h_relu2_2 = h
+        h = self.slice3(h)
+        h_relu3_3 = h
+        h = self.slice4(h)
+        h_relu4_3 = h
+        h = self.slice5(h)
+        h_relu5_3 = h
+        vgg_outputs = namedtuple(
+            "VggOutputs",
+            ["relu1_2", "relu2_2", "relu3_3", "relu4_3", "relu5_3"],
+        )
+        out = vgg_outputs(h_relu1_2, h_relu2_2, h_relu3_3, h_relu4_3, h_relu5_3)
+
+        return out
+
+
+class resnet(torch.nn.Module):
+    def __init__(self, requires_grad=False, pretrained=True, num=18):
+        super(resnet, self).__init__()
+        if num == 18:
+            self.net = models.resnet18(pretrained=pretrained)
+        elif num == 34:
+            self.net = models.resnet34(pretrained=pretrained)
+        elif num == 50:
+            self.net = models.resnet50(pretrained=pretrained)
+        elif num == 101:
+            self.net = models.resnet101(pretrained=pretrained)
+        elif num == 152:
+            self.net = models.resnet152(pretrained=pretrained)
+        self.N_slices = 5
+
+        self.conv1 = self.net.conv1
+        self.bn1 = self.net.bn1
+        self.relu = self.net.relu
+        self.maxpool = self.net.maxpool
+        self.layer1 = self.net.layer1
+        self.layer2 = self.net.layer2
+        self.layer3 = self.net.layer3
+        self.layer4 = self.net.layer4
+
+    def forward(self, X):
+        h = self.conv1(X)
+        h = self.bn1(h)
+        h = self.relu(h)
+        h_relu1 = h
+        h = self.maxpool(h)
+        h = self.layer1(h)
+        h_conv2 = h
+        h = self.layer2(h)
+        h_conv3 = h
+        h = self.layer3(h)
+        h_conv4 = h
+        h = self.layer4(h)
+        h_conv5 = h
+
+        outputs = namedtuple(
+            "Outputs", ["relu1", "conv2", "conv3", "conv4", "conv5"]
+        )
+        out = outputs(h_relu1, h_conv2, h_conv3, h_conv4, h_conv5)
+
+        return out
+
+# Off-the-shelf deep network
+class PNet(torch.nn.Module):
+    """Pre-trained network with all channels equally weighted by default"""
+
+    def __init__(self, pnet_type="vgg", pnet_rand=False, use_gpu=True):
+        super(PNet, self).__init__()
+
+        self.use_gpu = use_gpu
+
+        self.pnet_type = pnet_type
+        self.pnet_rand = pnet_rand
+
+        self.shift = torch.Tensor([-0.030, -0.088, -0.188]).view(1, 3, 1, 1)
+        self.scale = torch.Tensor([0.458, 0.448, 0.450]).view(1, 3, 1, 1)
+
+        if self.pnet_type in ["vgg", "vgg16"]:
+            self.net = vgg16(pretrained=not self.pnet_rand, requires_grad=False)
+        elif self.pnet_type == "alex":
+            self.net = alexnet(
+                pretrained=not self.pnet_rand, requires_grad=False
+            )
+        elif self.pnet_type[:-2] == "resnet":
+            self.net = resnet(
+                pretrained=not self.pnet_rand,
+                requires_grad=False,
+                num=int(self.pnet_type[-2:]),
+            )
+        elif self.pnet_type == "squeeze":
+            self.net = squeezenet(
+                pretrained=not self.pnet_rand, requires_grad=False
+            )
+
+        self.L = self.net.N_slices
+
+        if use_gpu:
+            self.net.cuda()
+            self.shift = self.shift.cuda()
+            self.scale = self.scale.cuda()
+
+    def forward(self, in0, in1, retPerLayer=False):
+        in0_sc = (in0 - self.shift.expand_as(in0)) / self.scale.expand_as(in0)
+        in1_sc = (in1 - self.shift.expand_as(in0)) / self.scale.expand_as(in0)
+
+        outs0 = self.net.forward(in0_sc)
+        outs1 = self.net.forward(in1_sc)
+
+        if retPerLayer:
+            all_scores = []
+        for (kk, out0) in enumerate(outs0):
+            cur_score = 1.0 - cos_sim(outs0[kk], outs1[kk])
+            if kk == 0:
+                val = 1.0 * cur_score
+            else:
+                val = val + cur_score
+            if retPerLayer:
+                all_scores += [cur_score]
+
+        if retPerLayer:
+            return (val, all_scores)
+        else:
+            return val
+
+
+
+
+# The SSIM metric
+def ssim_metric(img1, img2, mask=None):
+    return ssim(img1, img2, mask=mask, size_average=False)
+
+
+# The PSNR metric
+def psnr(img1, img2, mask=None,reshape=False):
+    b = img1.size(0)
+    if not (mask is None):
+        b = img1.size(0)
+        mse_err = (img1 - img2).pow(2) * mask
+        if reshape:
+            mse_err = mse_err.reshape(b, -1).sum(dim=1) / (
+                    3 * mask.reshape(b, -1).sum(dim=1).clamp(min=1)
+            )
+        else:
+            mse_err = mse_err.view(b, -1).sum(dim=1) / (
+                3 * mask.view(b, -1).sum(dim=1).clamp(min=1)
+            )
+    else:
+        if reshape:
+            mse_err = (img1 - img2).pow(2).reshape(b, -1).mean(dim=1)
+        else:
+            mse_err = (img1 - img2).pow(2).view(b, -1).mean(dim=1)
+
+    psnr = 10 * (1 / mse_err).log10()
+    return psnr
+
+
+# The perceptual similarity metric
+def perceptual_sim(img1, img2, vgg16):
+    # First extract features
+    dist = vgg16(img1 * 2 - 1, img2 * 2 - 1)
+
+    return dist
+
+def load_img(img_name, size=None):
+    try:
+        img = Image.open(img_name)
+
+        if type(size) == int:
+            img = img.resize((size, size))
+        elif size is not None:
+            img = img.resize((size[1], size[0]))
+
+        img = transform(img).cuda()
+        img = img.unsqueeze(0)
+    except Exception as e:
+        print("Failed at loading %s " % img_name)
+        print(e)
+        img = torch.zeros(1, 3, 256, 256).cuda()
+        raise
+    return img
+
+
+def compute_perceptual_similarity(folder, pred_img, tgt_img, take_every_other):
+
+    # Load VGG16 for feature similarity
+    vgg16 = PNet().to("cuda")
+    vgg16.eval()
+    vgg16.cuda()
+
+    values_percsim = []
+    values_ssim = []
+    values_psnr = []
+    folders = os.listdir(folder)
+    for i, f in tqdm(enumerate(sorted(folders))):
+        pred_imgs = glob.glob(folder + f + "/" + pred_img)
+        tgt_imgs = glob.glob(folder + f + "/" + tgt_img)
+        assert len(tgt_imgs) == 1
+
+        perc_sim = 10000
+        ssim_sim = -10
+        psnr_sim = -10
+        for p_img in pred_imgs:
+            t_img = load_img(tgt_imgs[0])
+            p_img = load_img(p_img, size=t_img.shape[2:])
+            t_perc_sim = perceptual_sim(p_img, t_img, vgg16).item()
+            perc_sim = min(perc_sim, t_perc_sim)
+
+            ssim_sim = max(ssim_sim, ssim_metric(p_img, t_img).item())
+            psnr_sim = max(psnr_sim, psnr(p_img, t_img).item())
+
+        values_percsim += [perc_sim]
+        values_ssim += [ssim_sim]
+        values_psnr += [psnr_sim]
+
+    if take_every_other:
+        n_valuespercsim = []
+        n_valuesssim = []
+        n_valuespsnr = []
+        for i in range(0, len(values_percsim) // 2):
+            n_valuespercsim += [
+                min(values_percsim[2 * i], values_percsim[2 * i + 1])
+            ]
+            n_valuespsnr += [max(values_psnr[2 * i], values_psnr[2 * i + 1])]
+            n_valuesssim += [max(values_ssim[2 * i], values_ssim[2 * i + 1])]
+
+        values_percsim = n_valuespercsim
+        values_ssim = n_valuesssim
+        values_psnr = n_valuespsnr
+
+    avg_percsim = np.mean(np.array(values_percsim))
+    std_percsim = np.std(np.array(values_percsim))
+
+    avg_psnr = np.mean(np.array(values_psnr))
+    std_psnr = np.std(np.array(values_psnr))
+
+    avg_ssim = np.mean(np.array(values_ssim))
+    std_ssim = np.std(np.array(values_ssim))
+
+    return {
+        "Perceptual similarity": (avg_percsim, std_percsim),
+        "PSNR": (avg_psnr, std_psnr),
+        "SSIM": (avg_ssim, std_ssim),
+    }
+
+
+def compute_perceptual_similarity_from_list(pred_imgs_list, tgt_imgs_list,
+                                            take_every_other,
+                                            simple_format=True):
+
+    # Load VGG16 for feature similarity
+    vgg16 = PNet().to("cuda")
+    vgg16.eval()
+    vgg16.cuda()
+
+    values_percsim = []
+    values_ssim = []
+    values_psnr = []
+    equal_count = 0
+    ambig_count = 0
+    for i, tgt_img in enumerate(tqdm(tgt_imgs_list)):
+        pred_imgs = pred_imgs_list[i]
+        tgt_imgs = [tgt_img]
+        assert len(tgt_imgs) == 1
+
+        if type(pred_imgs) != list:
+            pred_imgs = [pred_imgs]
+
+        perc_sim = 10000
+        ssim_sim = -10
+        psnr_sim = -10
+        assert len(pred_imgs)>0
+        for p_img in pred_imgs:
+            t_img = load_img(tgt_imgs[0])
+            p_img = load_img(p_img, size=t_img.shape[2:])
+            t_perc_sim = perceptual_sim(p_img, t_img, vgg16).item()
+            perc_sim = min(perc_sim, t_perc_sim)
+
+            ssim_sim = max(ssim_sim, ssim_metric(p_img, t_img).item())
+            psnr_sim = max(psnr_sim, psnr(p_img, t_img).item())
+
+        values_percsim += [perc_sim]
+        values_ssim += [ssim_sim]
+        if psnr_sim != np.float("inf"):
+            values_psnr += [psnr_sim]
+        else:
+            if torch.allclose(p_img, t_img):
+                equal_count += 1
+                print("{} equal src and wrp images.".format(equal_count))
+            else:
+                ambig_count += 1
+                print("{} ambiguous src and wrp images.".format(ambig_count))
+
+    if take_every_other:
+        n_valuespercsim = []
+        n_valuesssim = []
+        n_valuespsnr = []
+        for i in range(0, len(values_percsim) // 2):
+            n_valuespercsim += [
+                min(values_percsim[2 * i], values_percsim[2 * i + 1])
+            ]
+            n_valuespsnr += [max(values_psnr[2 * i], values_psnr[2 * i + 1])]
+            n_valuesssim += [max(values_ssim[2 * i], values_ssim[2 * i + 1])]
+
+        values_percsim = n_valuespercsim
+        values_ssim = n_valuesssim
+        values_psnr = n_valuespsnr
+
+    avg_percsim = np.mean(np.array(values_percsim))
+    std_percsim = np.std(np.array(values_percsim))
+
+    avg_psnr = np.mean(np.array(values_psnr))
+    std_psnr = np.std(np.array(values_psnr))
+
+    avg_ssim = np.mean(np.array(values_ssim))
+    std_ssim = np.std(np.array(values_ssim))
+
+    if simple_format:
+        # just to make yaml formatting readable
+        return {
+            "Perceptual similarity": [float(avg_percsim), float(std_percsim)],
+            "PSNR": [float(avg_psnr), float(std_psnr)],
+            "SSIM": [float(avg_ssim), float(std_ssim)],
+        }
+    else:
+        return {
+            "Perceptual similarity": (avg_percsim, std_percsim),
+            "PSNR": (avg_psnr, std_psnr),
+            "SSIM": (avg_ssim, std_ssim),
+        }
+
+
+def compute_perceptual_similarity_from_list_topk(pred_imgs_list, tgt_imgs_list,
+                                                 take_every_other, resize=False):
+
+    # Load VGG16 for feature similarity
+    vgg16 = PNet().to("cuda")
+    vgg16.eval()
+    vgg16.cuda()
+
+    values_percsim = []
+    values_ssim = []
+    values_psnr = []
+    individual_percsim = []
+    individual_ssim = []
+    individual_psnr = []
+    for i, tgt_img in enumerate(tqdm(tgt_imgs_list)):
+        pred_imgs = pred_imgs_list[i]
+        tgt_imgs = [tgt_img]
+        assert len(tgt_imgs) == 1
+
+        if type(pred_imgs) != list:
+            assert False
+            pred_imgs = [pred_imgs]
+
+        perc_sim = 10000
+        ssim_sim = -10
+        psnr_sim = -10
+        sample_percsim = list()
+        sample_ssim = list()
+        sample_psnr = list()
+        for p_img in pred_imgs:
+            if resize:
+                t_img = load_img(tgt_imgs[0], size=(256,256))
+            else:
+                t_img = load_img(tgt_imgs[0])
+            p_img = load_img(p_img, size=t_img.shape[2:])
+
+            t_perc_sim = perceptual_sim(p_img, t_img, vgg16).item()
+            sample_percsim.append(t_perc_sim)
+            perc_sim = min(perc_sim, t_perc_sim)
+
+            t_ssim = ssim_metric(p_img, t_img).item()
+            sample_ssim.append(t_ssim)
+            ssim_sim = max(ssim_sim, t_ssim)
+
+            t_psnr = psnr(p_img, t_img).item()
+            sample_psnr.append(t_psnr)
+            psnr_sim = max(psnr_sim, t_psnr)
+
+        values_percsim += [perc_sim]
+        values_ssim += [ssim_sim]
+        values_psnr += [psnr_sim]
+        individual_percsim.append(sample_percsim)
+        individual_ssim.append(sample_ssim)
+        individual_psnr.append(sample_psnr)
+
+    if take_every_other:
+        assert False, "Do this later, after specifying topk to get proper results"
+        n_valuespercsim = []
+        n_valuesssim = []
+        n_valuespsnr = []
+        for i in range(0, len(values_percsim) // 2):
+            n_valuespercsim += [
+                min(values_percsim[2 * i], values_percsim[2 * i + 1])
+            ]
+            n_valuespsnr += [max(values_psnr[2 * i], values_psnr[2 * i + 1])]
+            n_valuesssim += [max(values_ssim[2 * i], values_ssim[2 * i + 1])]
+
+        values_percsim = n_valuespercsim
+        values_ssim = n_valuesssim
+        values_psnr = n_valuespsnr
+
+    avg_percsim = np.mean(np.array(values_percsim))
+    std_percsim = np.std(np.array(values_percsim))
+
+    avg_psnr = np.mean(np.array(values_psnr))
+    std_psnr = np.std(np.array(values_psnr))
+
+    avg_ssim = np.mean(np.array(values_ssim))
+    std_ssim = np.std(np.array(values_ssim))
+
+    individual_percsim = np.array(individual_percsim)
+    individual_psnr = np.array(individual_psnr)
+    individual_ssim = np.array(individual_ssim)
+
+    return {
+        "avg_of_best": {
+            "Perceptual similarity": [float(avg_percsim), float(std_percsim)],
+            "PSNR": [float(avg_psnr), float(std_psnr)],
+            "SSIM": [float(avg_ssim), float(std_ssim)],
+        },
+        "individual": {
+            "PSIM": individual_percsim,
+            "PSNR": individual_psnr,
+            "SSIM": individual_ssim,
+        }
+    }
+
+
+if __name__ == "__main__":
+    args = argparse.ArgumentParser()
+    args.add_argument("--folder", type=str, default="")
+    args.add_argument("--pred_image", type=str, default="")
+    args.add_argument("--target_image", type=str, default="")
+    args.add_argument("--take_every_other", action="store_true", default=False)
+    args.add_argument("--output_file", type=str, default="")
+
+    opts = args.parse_args()
+
+    folder = opts.folder
+    pred_img = opts.pred_image
+    tgt_img = opts.target_image
+
+    results = compute_perceptual_similarity(
+        folder, pred_img, tgt_img, opts.take_every_other
+    )
+
+    f = open(opts.output_file, 'w')
+    for key in results:
+        print("%s for %s: \n" % (key, opts.folder))
+        print(
+            "\t {:0.4f} | {:0.4f} \n".format(results[key][0], results[key][1])
+        )
+
+        f.write("%s for %s: \n" % (key, opts.folder))
+        f.write(
+            "\t {:0.4f} | {:0.4f} \n".format(results[key][0], results[key][1])
+        )
+
+    f.close()

ldm/modules/evaluate/frechet_video_distance.py

@@ -0,0 +1,147 @@
+# coding=utf-8
+# Copyright 2022 The Google Research Authors.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# Lint as: python2, python3
+"""Minimal Reference implementation for the Frechet Video Distance (FVD).
+
+FVD is a metric for the quality of video generation models. It is inspired by
+the FID (Frechet Inception Distance) used for images, but uses a different
+embedding to be better suitable for videos.
+"""
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+
+import six
+import tensorflow.compat.v1 as tf
+import tensorflow_gan as tfgan
+import tensorflow_hub as hub
+
+
+def preprocess(videos, target_resolution):
+  """Runs some preprocessing on the videos for I3D model.
+
+  Args:
+    videos: <T>[batch_size, num_frames, height, width, depth] The videos to be
+      preprocessed. We don't care about the specific dtype of the videos, it can
+      be anything that tf.image.resize_bilinear accepts. Values are expected to
+      be in the range 0-255.
+    target_resolution: (width, height): target video resolution
+
+  Returns:
+    videos: <float32>[batch_size, num_frames, height, width, depth]
+  """
+  videos_shape = list(videos.shape)
+  all_frames = tf.reshape(videos, [-1] + videos_shape[-3:])
+  resized_videos = tf.image.resize_bilinear(all_frames, size=target_resolution)
+  target_shape = [videos_shape[0], -1] + list(target_resolution) + [3]
+  output_videos = tf.reshape(resized_videos, target_shape)
+  scaled_videos = 2. * tf.cast(output_videos, tf.float32) / 255. - 1
+  return scaled_videos
+
+
+def _is_in_graph(tensor_name):
+  """Checks whether a given tensor does exists in the graph."""
+  try:
+    tf.get_default_graph().get_tensor_by_name(tensor_name)
+  except KeyError:
+    return False
+  return True
+
+
+def create_id3_embedding(videos,warmup=False,batch_size=16):
+  """Embeds the given videos using the Inflated 3D Convolution ne   twork.
+
+  Downloads the graph of the I3D from tf.hub and adds it to the graph on the
+  first call.
+
+  Args:
+    videos: <float32>[batch_size, num_frames, height=224, width=224, depth=3].
+      Expected range is [-1, 1].
+
+  Returns:
+    embedding: <float32>[batch_size, embedding_size]. embedding_size depends
+               on the model used.
+
+  Raises:
+    ValueError: when a provided embedding_layer is not supported.
+  """
+
+  # batch_size = 16
+  module_spec = "https://tfhub.dev/deepmind/i3d-kinetics-400/1"
+
+
+  # Making sure that we import the graph separately for
+  # each different input video tensor.
+  module_name = "fvd_kinetics-400_id3_module_" + six.ensure_str(
+      videos.name).replace(":", "_")
+
+
+
+  assert_ops = [
+      tf.Assert(
+          tf.reduce_max(videos) <= 1.001,
+          ["max value in frame is > 1", videos]),
+      tf.Assert(
+          tf.reduce_min(videos) >= -1.001,
+          ["min value in frame is < -1", videos]),
+      tf.assert_equal(
+          tf.shape(videos)[0],
+          batch_size, ["invalid frame batch size: ",
+                       tf.shape(videos)],
+          summarize=6),
+  ]
+  with tf.control_dependencies(assert_ops):
+    videos = tf.identity(videos)
+
+  module_scope = "%s_apply_default/" % module_name
+
+  # To check whether the module has already been loaded into the graph, we look
+  # for a given tensor name. If this tensor name exists, we assume the function
+  # has been called before and the graph was imported. Otherwise we import it.
+  # Note: in theory, the tensor could exist, but have wrong shapes.
+  # This will happen if create_id3_embedding is called with a frames_placehoder
+  # of wrong size/batch size, because even though that will throw a tf.Assert
+  # on graph-execution time, it will insert the tensor (with wrong shape) into
+  # the graph. This is why we need the following assert.
+  if warmup:
+      video_batch_size = int(videos.shape[0])
+      assert video_batch_size in [batch_size, -1, None], f"Invalid batch size {video_batch_size}"
+  tensor_name = module_scope + "RGB/inception_i3d/Mean:0"
+  if not _is_in_graph(tensor_name):
+    i3d_model = hub.Module(module_spec, name=module_name)
+    i3d_model(videos)
+
+  # gets the kinetics-i3d-400-logits layer
+  tensor_name = module_scope + "RGB/inception_i3d/Mean:0"
+  tensor = tf.get_default_graph().get_tensor_by_name(tensor_name)
+  return tensor
+
+
+def calculate_fvd(real_activations,
+                  generated_activations):
+  """Returns a list of ops that compute metrics as funcs of activations.
+
+  Args:
+    real_activations: <float32>[num_samples, embedding_size]
+    generated_activations: <float32>[num_samples, embedding_size]
+
+  Returns:
+    A scalar that contains the requested FVD.
+  """
+  return tfgan.eval.frechet_classifier_distance_from_activations(
+      real_activations, generated_activations)

ldm/modules/evaluate/ssim.py

@@ -0,0 +1,124 @@
+# MIT Licence
+
+# Methods to predict the SSIM, taken from
+# https://github.com/Po-Hsun-Su/pytorch-ssim/blob/master/pytorch_ssim/__init__.py
+
+from math import exp
+
+import torch
+import torch.nn.functional as F
+from torch.autograd import Variable
+
+def gaussian(window_size, sigma):
+    gauss = torch.Tensor(
+        [
+            exp(-((x - window_size // 2) ** 2) / float(2 * sigma ** 2))
+            for x in range(window_size)
+        ]
+    )
+    return gauss / gauss.sum()
+
+
+def create_window(window_size, channel):
+    _1D_window = gaussian(window_size, 1.5).unsqueeze(1)
+    _2D_window = _1D_window.mm(_1D_window.t()).float().unsqueeze(0).unsqueeze(0)
+    window = Variable(
+        _2D_window.expand(channel, 1, window_size, window_size).contiguous()
+    )
+    return window
+
+
+def _ssim(
+    img1, img2, window, window_size, channel, mask=None, size_average=True
+):
+    mu1 = F.conv2d(img1, window, padding=window_size // 2, groups=channel)
+    mu2 = F.conv2d(img2, window, padding=window_size // 2, groups=channel)
+
+    mu1_sq = mu1.pow(2)
+    mu2_sq = mu2.pow(2)
+    mu1_mu2 = mu1 * mu2
+
+    sigma1_sq = (
+        F.conv2d(img1 * img1, window, padding=window_size // 2, groups=channel)
+        - mu1_sq
+    )
+    sigma2_sq = (
+        F.conv2d(img2 * img2, window, padding=window_size // 2, groups=channel)
+        - mu2_sq
+    )
+    sigma12 = (
+        F.conv2d(img1 * img2, window, padding=window_size // 2, groups=channel)
+        - mu1_mu2
+    )
+
+    C1 = (0.01) ** 2
+    C2 = (0.03) ** 2
+
+    ssim_map = ((2 * mu1_mu2 + C1) * (2 * sigma12 + C2)) / (
+        (mu1_sq + mu2_sq + C1) * (sigma1_sq + sigma2_sq + C2)
+    )
+
+    if not (mask is None):
+        b = mask.size(0)
+        ssim_map = ssim_map.mean(dim=1, keepdim=True) * mask
+        ssim_map = ssim_map.view(b, -1).sum(dim=1) / mask.view(b, -1).sum(
+            dim=1
+        ).clamp(min=1)
+        return ssim_map
+
+    import pdb
+
+    pdb.set_trace
+
+    if size_average:
+        return ssim_map.mean()
+    else:
+        return ssim_map.mean(1).mean(1).mean(1)
+
+
+class SSIM(torch.nn.Module):
+    def __init__(self, window_size=11, size_average=True):
+        super(SSIM, self).__init__()
+        self.window_size = window_size
+        self.size_average = size_average
+        self.channel = 1
+        self.window = create_window(window_size, self.channel)
+
+    def forward(self, img1, img2, mask=None):
+        (_, channel, _, _) = img1.size()
+
+        if (
+            channel == self.channel
+            and self.window.data.type() == img1.data.type()
+        ):
+            window = self.window
+        else:
+            window = create_window(self.window_size, channel)
+
+            if img1.is_cuda:
+                window = window.cuda(img1.get_device())
+            window = window.type_as(img1)
+
+            self.window = window
+            self.channel = channel
+
+        return _ssim(
+            img1,
+            img2,
+            window,
+            self.window_size,
+            channel,
+            mask,
+            self.size_average,
+        )
+
+
+def ssim(img1, img2, window_size=11, mask=None, size_average=True):
+    (_, channel, _, _) = img1.size()
+    window = create_window(window_size, channel)
+
+    if img1.is_cuda:
+        window = window.cuda(img1.get_device())
+    window = window.type_as(img1)
+
+    return _ssim(img1, img2, window, window_size, channel, mask, size_average)

ldm/modules/evaluate/torch_frechet_video_distance.py

@@ -0,0 +1,294 @@
+# based on https://github.com/universome/fvd-comparison/blob/master/compare_models.py; huge thanks!
+import os
+import numpy as np
+import io
+import re
+import requests
+import html
+import hashlib
+import urllib
+import urllib.request
+import scipy.linalg
+import multiprocessing as mp
+import glob
+
+
+from tqdm import tqdm
+from typing import Any, List, Tuple, Union, Dict, Callable
+
+from torchvision.io import read_video
+import torch; torch.set_grad_enabled(False)
+from einops import rearrange
+
+from nitro.util import isvideo
+
+def compute_frechet_distance(mu_sample,sigma_sample,mu_ref,sigma_ref) -> float:
+    print('Calculate frechet distance...')
+    m = np.square(mu_sample - mu_ref).sum()
+    s, _ = scipy.linalg.sqrtm(np.dot(sigma_sample, sigma_ref), disp=False) # pylint: disable=no-member
+    fid = np.real(m + np.trace(sigma_sample + sigma_ref - s * 2))
+
+    return float(fid)
+
+
+def compute_stats(feats: np.ndarray) -> Tuple[np.ndarray, np.ndarray]:
+    mu = feats.mean(axis=0) # [d]
+    sigma = np.cov(feats, rowvar=False) # [d, d]
+
+    return mu, sigma
+
+
+def open_url(url: str, num_attempts: int = 10, verbose: bool = True, return_filename: bool = False) -> Any:
+    """Download the given URL and return a binary-mode file object to access the data."""
+    assert num_attempts >= 1
+
+    # Doesn't look like an URL scheme so interpret it as a local filename.
+    if not re.match('^[a-z]+://', url):
+        return url if return_filename else open(url, "rb")
+
+    # Handle file URLs.  This code handles unusual file:// patterns that
+    # arise on Windows:
+    #
+    # file:///c:/foo.txt
+    #
+    # which would translate to a local '/c:/foo.txt' filename that's
+    # invalid.  Drop the forward slash for such pathnames.
+    #
+    # If you touch this code path, you should test it on both Linux and
+    # Windows.
+    #
+    # Some internet resources suggest using urllib.request.url2pathname() but
+    # but that converts forward slashes to backslashes and this causes
+    # its own set of problems.
+    if url.startswith('file://'):
+        filename = urllib.parse.urlparse(url).path
+        if re.match(r'^/[a-zA-Z]:', filename):
+            filename = filename[1:]
+        return filename if return_filename else open(filename, "rb")
+
+    url_md5 = hashlib.md5(url.encode("utf-8")).hexdigest()
+
+    # Download.
+    url_name = None
+    url_data = None
+    with requests.Session() as session:
+        if verbose:
+            print("Downloading %s ..." % url, end="", flush=True)
+        for attempts_left in reversed(range(num_attempts)):
+            try:
+                with session.get(url) as res:
+                    res.raise_for_status()
+                    if len(res.content) == 0:
+                        raise IOError("No data received")
+
+                    if len(res.content) < 8192:
+                        content_str = res.content.decode("utf-8")
+                        if "download_warning" in res.headers.get("Set-Cookie", ""):
+                            links = [html.unescape(link) for link in content_str.split('"') if "export=download" in link]
+                            if len(links) == 1:
+                                url = requests.compat.urljoin(url, links[0])
+                                raise IOError("Google Drive virus checker nag")
+                        if "Google Drive - Quota exceeded" in content_str:
+                            raise IOError("Google Drive download quota exceeded -- please try again later")
+
+                    match = re.search(r'filename="([^"]*)"', res.headers.get("Content-Disposition", ""))
+                    url_name = match[1] if match else url
+                    url_data = res.content
+                    if verbose:
+                        print(" done")
+                    break
+            except KeyboardInterrupt:
+                raise
+            except:
+                if not attempts_left:
+                    if verbose:
+                        print(" failed")
+                    raise
+                if verbose:
+                    print(".", end="", flush=True)
+
+    # Return data as file object.
+    assert not return_filename
+    return io.BytesIO(url_data)
+
+def load_video(ip):
+    vid, *_ = read_video(ip)
+    vid = rearrange(vid, 't h w c -> t c h w').to(torch.uint8)
+    return vid
+
+def get_data_from_str(input_str,nprc = None):
+    assert os.path.isdir(input_str), f'Specified input folder "{input_str}" is not a directory'
+    vid_filelist = glob.glob(os.path.join(input_str,'*.mp4'))
+    print(f'Found {len(vid_filelist)} videos in dir {input_str}')
+
+    if nprc is None:
+        try:
+            nprc = mp.cpu_count()
+        except NotImplementedError:
+            print('WARNING: cpu_count() not avlailable, using only 1 cpu for video loading')
+            nprc = 1
+
+    pool = mp.Pool(processes=nprc)
+
+    vids = []
+    for v in tqdm(pool.imap_unordered(load_video,vid_filelist),total=len(vid_filelist),desc='Loading videos...'):
+        vids.append(v)
+
+
+    vids = torch.stack(vids,dim=0).float()
+
+    return vids
+
+def get_stats(stats):
+    assert os.path.isfile(stats) and stats.endswith('.npz'), f'no stats found under {stats}'
+
+    print(f'Using precomputed statistics under {stats}')
+    stats = np.load(stats)
+    stats = {key: stats[key] for key in stats.files}
+
+    return stats
+
+
+
+
+@torch.no_grad()
+def compute_fvd(ref_input, sample_input, bs=32,
+                ref_stats=None,
+                sample_stats=None,
+                nprc_load=None):
+
+
+
+    calc_stats = ref_stats is None or sample_stats is None
+
+    if calc_stats:
+
+        only_ref = sample_stats is not None
+        only_sample = ref_stats is not None
+
+
+        if isinstance(ref_input,str) and not only_sample:
+            ref_input = get_data_from_str(ref_input,nprc_load)
+
+        if isinstance(sample_input, str) and not only_ref:
+            sample_input = get_data_from_str(sample_input, nprc_load)
+
+        stats = compute_statistics(sample_input,ref_input,
+                                        device='cuda' if torch.cuda.is_available() else 'cpu',
+                                        bs=bs,
+                                        only_ref=only_ref,
+                                        only_sample=only_sample)
+
+        if only_ref:
+            stats.update(get_stats(sample_stats))
+        elif only_sample:
+            stats.update(get_stats(ref_stats))
+
+
+
+    else:
+        stats = get_stats(sample_stats)
+        stats.update(get_stats(ref_stats))
+
+    fvd = compute_frechet_distance(**stats)
+
+    return {'FVD' : fvd,}
+
+
+@torch.no_grad()
+def compute_statistics(videos_fake, videos_real, device: str='cuda', bs=32, only_ref=False,only_sample=False) -> Dict:
+    detector_url = 'https://www.dropbox.com/s/ge9e5ujwgetktms/i3d_torchscript.pt?dl=1'
+    detector_kwargs = dict(rescale=True, resize=True, return_features=True) # Return raw features before the softmax layer.
+
+    with open_url(detector_url, verbose=False) as f:
+        detector = torch.jit.load(f).eval().to(device)
+
+
+
+    assert not (only_sample and only_ref), 'only_ref and only_sample arguments are mutually exclusive'
+
+    ref_embed, sample_embed = [], []
+
+    info = f'Computing I3D activations for FVD score with batch size {bs}'
+
+    if only_ref:
+
+        if not isvideo(videos_real):
+            # if not is video we assume to have numpy arrays pf shape (n_vids, t, h, w, c) in range [0,255]
+            videos_real = torch.from_numpy(videos_real).permute(0, 4, 1, 2, 3).float()
+            print(videos_real.shape)
+
+        if videos_real.shape[0] % bs == 0:
+            n_secs = videos_real.shape[0] // bs
+        else:
+            n_secs = videos_real.shape[0] // bs + 1
+
+        videos_real = torch.tensor_split(videos_real, n_secs, dim=0)
+
+        for ref_v in tqdm(videos_real, total=len(videos_real),desc=info):
+
+            feats_ref = detector(ref_v.to(device).contiguous(), **detector_kwargs).cpu().numpy()
+            ref_embed.append(feats_ref)
+
+    elif only_sample:
+
+        if not isvideo(videos_fake):
+            # if not is video we assume to have numpy arrays pf shape (n_vids, t, h, w, c) in range [0,255]
+            videos_fake = torch.from_numpy(videos_fake).permute(0, 4, 1, 2, 3).float()
+            print(videos_fake.shape)
+
+        if videos_fake.shape[0] % bs == 0:
+            n_secs = videos_fake.shape[0] // bs
+        else:
+            n_secs = videos_fake.shape[0] // bs + 1
+
+        videos_real = torch.tensor_split(videos_real, n_secs, dim=0)
+
+        for sample_v in tqdm(videos_fake, total=len(videos_real),desc=info):
+            feats_sample = detector(sample_v.to(device).contiguous(), **detector_kwargs).cpu().numpy()
+            sample_embed.append(feats_sample)
+
+
+    else:
+
+        if not isvideo(videos_real):
+            # if not is video we assume to have numpy arrays pf shape (n_vids, t, h, w, c) in range [0,255]
+            videos_real = torch.from_numpy(videos_real).permute(0, 4, 1, 2, 3).float()
+
+        if not isvideo(videos_fake):
+            videos_fake = torch.from_numpy(videos_fake).permute(0, 4, 1, 2, 3).float()
+
+        if videos_fake.shape[0] % bs == 0:
+            n_secs = videos_fake.shape[0] // bs
+        else:
+            n_secs = videos_fake.shape[0] // bs + 1
+
+        videos_real = torch.tensor_split(videos_real, n_secs, dim=0)
+        videos_fake = torch.tensor_split(videos_fake, n_secs, dim=0)
+
+        for ref_v, sample_v in tqdm(zip(videos_real,videos_fake),total=len(videos_fake),desc=info):
+            # print(ref_v.shape)
+            # ref_v = torch.nn.functional.interpolate(ref_v, size=(sample_v.shape[2], 256, 256), mode='trilinear', align_corners=False)
+            # sample_v = torch.nn.functional.interpolate(sample_v, size=(sample_v.shape[2], 256, 256), mode='trilinear', align_corners=False)
+
+
+            feats_sample = detector(sample_v.to(device).contiguous(), **detector_kwargs).cpu().numpy()
+            feats_ref = detector(ref_v.to(device).contiguous(), **detector_kwargs).cpu().numpy()
+            sample_embed.append(feats_sample)
+            ref_embed.append(feats_ref)
+
+    out = dict()
+    if len(sample_embed) > 0:
+        sample_embed = np.concatenate(sample_embed,axis=0)
+        mu_sample, sigma_sample = compute_stats(sample_embed)
+        out.update({'mu_sample': mu_sample,
+                    'sigma_sample': sigma_sample})
+
+    if len(ref_embed) > 0:
+        ref_embed = np.concatenate(ref_embed,axis=0)
+        mu_ref, sigma_ref = compute_stats(ref_embed)
+        out.update({'mu_ref': mu_ref,
+                    'sigma_ref': sigma_ref})
+
+
+    return out

ldm/modules/image_degradation/__init__.py

@@ -0,0 +1,2 @@
+from ldm.modules.image_degradation.bsrgan import degradation_bsrgan_variant as degradation_fn_bsr
+from ldm.modules.image_degradation.bsrgan_light import degradation_bsrgan_variant as degradation_fn_bsr_light

ldm/modules/image_degradation/bsrgan.py

@@ -0,0 +1,730 @@
+# -*- coding: utf-8 -*-
+"""
+# --------------------------------------------
+# Super-Resolution
+# --------------------------------------------
+#
+# Kai Zhang (cskaizhang@gmail.com)
+# https://github.com/cszn
+# From 2019/03--2021/08
+# --------------------------------------------
+"""
+
+import numpy as np
+import cv2
+import torch
+
+from functools import partial
+import random
+from scipy import ndimage
+import scipy
+import scipy.stats as ss
+from scipy.interpolate import interp2d
+from scipy.linalg import orth
+import albumentations
+
+import ldm.modules.image_degradation.utils_image as util
+
+
+def modcrop_np(img, sf):
+    '''
+    Args:
+        img: numpy image, WxH or WxHxC
+        sf: scale factor
+    Return:
+        cropped image
+    '''
+    w, h = img.shape[:2]
+    im = np.copy(img)
+    return im[:w - w % sf, :h - h % sf, ...]
+
+
+"""
+# --------------------------------------------
+# anisotropic Gaussian kernels
+# --------------------------------------------
+"""
+
+
+def analytic_kernel(k):
+    """Calculate the X4 kernel from the X2 kernel (for proof see appendix in paper)"""
+    k_size = k.shape[0]
+    # Calculate the big kernels size
+    big_k = np.zeros((3 * k_size - 2, 3 * k_size - 2))
+    # Loop over the small kernel to fill the big one
+    for r in range(k_size):
+        for c in range(k_size):
+            big_k[2 * r:2 * r + k_size, 2 * c:2 * c + k_size] += k[r, c] * k
+    # Crop the edges of the big kernel to ignore very small values and increase run time of SR
+    crop = k_size // 2
+    cropped_big_k = big_k[crop:-crop, crop:-crop]
+    # Normalize to 1
+    return cropped_big_k / cropped_big_k.sum()
+
+
+def anisotropic_Gaussian(ksize=15, theta=np.pi, l1=6, l2=6):
+    """ generate an anisotropic Gaussian kernel
+    Args:
+        ksize : e.g., 15, kernel size
+        theta : [0,  pi], rotation angle range
+        l1    : [0.1,50], scaling of eigenvalues
+        l2    : [0.1,l1], scaling of eigenvalues
+        If l1 = l2, will get an isotropic Gaussian kernel.
+    Returns:
+        k     : kernel
+    """
+
+    v = np.dot(np.array([[np.cos(theta), -np.sin(theta)], [np.sin(theta), np.cos(theta)]]), np.array([1., 0.]))
+    V = np.array([[v[0], v[1]], [v[1], -v[0]]])
+    D = np.array([[l1, 0], [0, l2]])
+    Sigma = np.dot(np.dot(V, D), np.linalg.inv(V))
+    k = gm_blur_kernel(mean=[0, 0], cov=Sigma, size=ksize)
+
+    return k
+
+
+def gm_blur_kernel(mean, cov, size=15):
+    center = size / 2.0 + 0.5
+    k = np.zeros([size, size])
+    for y in range(size):
+        for x in range(size):
+            cy = y - center + 1
+            cx = x - center + 1
+            k[y, x] = ss.multivariate_normal.pdf([cx, cy], mean=mean, cov=cov)
+
+    k = k / np.sum(k)
+    return k
+
+
+def shift_pixel(x, sf, upper_left=True):
+    """shift pixel for super-resolution with different scale factors
+    Args:
+        x: WxHxC or WxH
+        sf: scale factor
+        upper_left: shift direction
+    """
+    h, w = x.shape[:2]
+    shift = (sf - 1) * 0.5
+    xv, yv = np.arange(0, w, 1.0), np.arange(0, h, 1.0)
+    if upper_left:
+        x1 = xv + shift
+        y1 = yv + shift
+    else:
+        x1 = xv - shift
+        y1 = yv - shift
+
+    x1 = np.clip(x1, 0, w - 1)
+    y1 = np.clip(y1, 0, h - 1)
+
+    if x.ndim == 2:
+        x = interp2d(xv, yv, x)(x1, y1)
+    if x.ndim == 3:
+        for i in range(x.shape[-1]):
+            x[:, :, i] = interp2d(xv, yv, x[:, :, i])(x1, y1)
+
+    return x
+
+
+def blur(x, k):
+    '''
+    x: image, NxcxHxW
+    k: kernel, Nx1xhxw
+    '''
+    n, c = x.shape[:2]
+    p1, p2 = (k.shape[-2] - 1) // 2, (k.shape[-1] - 1) // 2
+    x = torch.nn.functional.pad(x, pad=(p1, p2, p1, p2), mode='replicate')
+    k = k.repeat(1, c, 1, 1)
+    k = k.view(-1, 1, k.shape[2], k.shape[3])
+    x = x.view(1, -1, x.shape[2], x.shape[3])
+    x = torch.nn.functional.conv2d(x, k, bias=None, stride=1, padding=0, groups=n * c)
+    x = x.view(n, c, x.shape[2], x.shape[3])
+
+    return x
+
+
+def gen_kernel(k_size=np.array([15, 15]), scale_factor=np.array([4, 4]), min_var=0.6, max_var=10., noise_level=0):
+    """"
+    # modified version of https://github.com/assafshocher/BlindSR_dataset_generator
+    # Kai Zhang
+    # min_var = 0.175 * sf  # variance of the gaussian kernel will be sampled between min_var and max_var
+    # max_var = 2.5 * sf
+    """
+    # Set random eigen-vals (lambdas) and angle (theta) for COV matrix
+    lambda_1 = min_var + np.random.rand() * (max_var - min_var)
+    lambda_2 = min_var + np.random.rand() * (max_var - min_var)
+    theta = np.random.rand() * np.pi  # random theta
+    noise = -noise_level + np.random.rand(*k_size) * noise_level * 2
+
+    # Set COV matrix using Lambdas and Theta
+    LAMBDA = np.diag([lambda_1, lambda_2])
+    Q = np.array([[np.cos(theta), -np.sin(theta)],
+                  [np.sin(theta), np.cos(theta)]])
+    SIGMA = Q @ LAMBDA @ Q.T
+    INV_SIGMA = np.linalg.inv(SIGMA)[None, None, :, :]
+
+    # Set expectation position (shifting kernel for aligned image)
+    MU = k_size // 2 - 0.5 * (scale_factor - 1)  # - 0.5 * (scale_factor - k_size % 2)
+    MU = MU[None, None, :, None]
+
+    # Create meshgrid for Gaussian
+    [X, Y] = np.meshgrid(range(k_size[0]), range(k_size[1]))
+    Z = np.stack([X, Y], 2)[:, :, :, None]
+
+    # Calcualte Gaussian for every pixel of the kernel
+    ZZ = Z - MU
+    ZZ_t = ZZ.transpose(0, 1, 3, 2)
+    raw_kernel = np.exp(-0.5 * np.squeeze(ZZ_t @ INV_SIGMA @ ZZ)) * (1 + noise)
+
+    # shift the kernel so it will be centered
+    # raw_kernel_centered = kernel_shift(raw_kernel, scale_factor)
+
+    # Normalize the kernel and return
+    # kernel = raw_kernel_centered / np.sum(raw_kernel_centered)
+    kernel = raw_kernel / np.sum(raw_kernel)
+    return kernel
+
+
+def fspecial_gaussian(hsize, sigma):
+    hsize = [hsize, hsize]
+    siz = [(hsize[0] - 1.0) / 2.0, (hsize[1] - 1.0) / 2.0]
+    std = sigma
+    [x, y] = np.meshgrid(np.arange(-siz[1], siz[1] + 1), np.arange(-siz[0], siz[0] + 1))
+    arg = -(x * x + y * y) / (2 * std * std)
+    h = np.exp(arg)
+    h[h < scipy.finfo(float).eps * h.max()] = 0
+    sumh = h.sum()
+    if sumh != 0:
+        h = h / sumh
+    return h
+
+
+def fspecial_laplacian(alpha):
+    alpha = max([0, min([alpha, 1])])
+    h1 = alpha / (alpha + 1)
+    h2 = (1 - alpha) / (alpha + 1)
+    h = [[h1, h2, h1], [h2, -4 / (alpha + 1), h2], [h1, h2, h1]]
+    h = np.array(h)
+    return h
+
+
+def fspecial(filter_type, *args, **kwargs):
+    '''
+    python code from:
+    https://github.com/ronaldosena/imagens-medicas-2/blob/40171a6c259edec7827a6693a93955de2bd39e76/Aulas/aula_2_-_uniform_filter/matlab_fspecial.py
+    '''
+    if filter_type == 'gaussian':
+        return fspecial_gaussian(*args, **kwargs)
+    if filter_type == 'laplacian':
+        return fspecial_laplacian(*args, **kwargs)
+
+
+"""
+# --------------------------------------------
+# degradation models
+# --------------------------------------------
+"""
+
+
+def bicubic_degradation(x, sf=3):
+    '''
+    Args:
+        x: HxWxC image, [0, 1]
+        sf: down-scale factor
+    Return:
+        bicubicly downsampled LR image
+    '''
+    x = util.imresize_np(x, scale=1 / sf)
+    return x
+
+
+def srmd_degradation(x, k, sf=3):
+    ''' blur + bicubic downsampling
+    Args:
+        x: HxWxC image, [0, 1]
+        k: hxw, double
+        sf: down-scale factor
+    Return:
+        downsampled LR image
+    Reference:
+        @inproceedings{zhang2018learning,
+          title={Learning a single convolutional super-resolution network for multiple degradations},
+          author={Zhang, Kai and Zuo, Wangmeng and Zhang, Lei},
+          booktitle={IEEE Conference on Computer Vision and Pattern Recognition},
+          pages={3262--3271},
+          year={2018}
+        }
+    '''
+    x = ndimage.filters.convolve(x, np.expand_dims(k, axis=2), mode='wrap')  # 'nearest' | 'mirror'
+    x = bicubic_degradation(x, sf=sf)
+    return x
+
+
+def dpsr_degradation(x, k, sf=3):
+    ''' bicubic downsampling + blur
+    Args:
+        x: HxWxC image, [0, 1]
+        k: hxw, double
+        sf: down-scale factor
+    Return:
+        downsampled LR image
+    Reference:
+        @inproceedings{zhang2019deep,
+          title={Deep Plug-and-Play Super-Resolution for Arbitrary Blur Kernels},
+          author={Zhang, Kai and Zuo, Wangmeng and Zhang, Lei},
+          booktitle={IEEE Conference on Computer Vision and Pattern Recognition},
+          pages={1671--1681},
+          year={2019}
+        }
+    '''
+    x = bicubic_degradation(x, sf=sf)
+    x = ndimage.filters.convolve(x, np.expand_dims(k, axis=2), mode='wrap')
+    return x
+
+
+def classical_degradation(x, k, sf=3):
+    ''' blur + downsampling
+    Args:
+        x: HxWxC image, [0, 1]/[0, 255]
+        k: hxw, double
+        sf: down-scale factor
+    Return:
+        downsampled LR image
+    '''
+    x = ndimage.filters.convolve(x, np.expand_dims(k, axis=2), mode='wrap')
+    # x = filters.correlate(x, np.expand_dims(np.flip(k), axis=2))
+    st = 0
+    return x[st::sf, st::sf, ...]
+
+
+def add_sharpening(img, weight=0.5, radius=50, threshold=10):
+    """USM sharpening. borrowed from real-ESRGAN
+    Input image: I; Blurry image: B.
+    1. K = I + weight * (I - B)
+    2. Mask = 1 if abs(I - B) > threshold, else: 0
+    3. Blur mask:
+    4. Out = Mask * K + (1 - Mask) * I
+    Args:
+        img (Numpy array): Input image, HWC, BGR; float32, [0, 1].
+        weight (float): Sharp weight. Default: 1.
+        radius (float): Kernel size of Gaussian blur. Default: 50.
+        threshold (int):
+    """
+    if radius % 2 == 0:
+        radius += 1
+    blur = cv2.GaussianBlur(img, (radius, radius), 0)
+    residual = img - blur
+    mask = np.abs(residual) * 255 > threshold
+    mask = mask.astype('float32')
+    soft_mask = cv2.GaussianBlur(mask, (radius, radius), 0)
+
+    K = img + weight * residual
+    K = np.clip(K, 0, 1)
+    return soft_mask * K + (1 - soft_mask) * img
+
+
+def add_blur(img, sf=4):
+    wd2 = 4.0 + sf
+    wd = 2.0 + 0.2 * sf
+    if random.random() < 0.5:
+        l1 = wd2 * random.random()
+        l2 = wd2 * random.random()
+        k = anisotropic_Gaussian(ksize=2 * random.randint(2, 11) + 3, theta=random.random() * np.pi, l1=l1, l2=l2)
+    else:
+        k = fspecial('gaussian', 2 * random.randint(2, 11) + 3, wd * random.random())
+    img = ndimage.filters.convolve(img, np.expand_dims(k, axis=2), mode='mirror')
+
+    return img
+
+
+def add_resize(img, sf=4):
+    rnum = np.random.rand()
+    if rnum > 0.8:  # up
+        sf1 = random.uniform(1, 2)
+    elif rnum < 0.7:  # down
+        sf1 = random.uniform(0.5 / sf, 1)
+    else:
+        sf1 = 1.0
+    img = cv2.resize(img, (int(sf1 * img.shape[1]), int(sf1 * img.shape[0])), interpolation=random.choice([1, 2, 3]))
+    img = np.clip(img, 0.0, 1.0)
+
+    return img
+
+
+# def add_Gaussian_noise(img, noise_level1=2, noise_level2=25):
+#     noise_level = random.randint(noise_level1, noise_level2)
+#     rnum = np.random.rand()
+#     if rnum > 0.6:  # add color Gaussian noise
+#         img += np.random.normal(0, noise_level / 255.0, img.shape).astype(np.float32)
+#     elif rnum < 0.4:  # add grayscale Gaussian noise
+#         img += np.random.normal(0, noise_level / 255.0, (*img.shape[:2], 1)).astype(np.float32)
+#     else:  # add  noise
+#         L = noise_level2 / 255.
+#         D = np.diag(np.random.rand(3))
+#         U = orth(np.random.rand(3, 3))
+#         conv = np.dot(np.dot(np.transpose(U), D), U)
+#         img += np.random.multivariate_normal([0, 0, 0], np.abs(L ** 2 * conv), img.shape[:2]).astype(np.float32)
+#     img = np.clip(img, 0.0, 1.0)
+#     return img
+
+def add_Gaussian_noise(img, noise_level1=2, noise_level2=25):
+    noise_level = random.randint(noise_level1, noise_level2)
+    rnum = np.random.rand()
+    if rnum > 0.6:  # add color Gaussian noise
+        img = img + np.random.normal(0, noise_level / 255.0, img.shape).astype(np.float32)
+    elif rnum < 0.4:  # add grayscale Gaussian noise
+        img = img + np.random.normal(0, noise_level / 255.0, (*img.shape[:2], 1)).astype(np.float32)
+    else:  # add  noise
+        L = noise_level2 / 255.
+        D = np.diag(np.random.rand(3))
+        U = orth(np.random.rand(3, 3))
+        conv = np.dot(np.dot(np.transpose(U), D), U)
+        img = img + np.random.multivariate_normal([0, 0, 0], np.abs(L ** 2 * conv), img.shape[:2]).astype(np.float32)
+    img = np.clip(img, 0.0, 1.0)
+    return img
+
+
+def add_speckle_noise(img, noise_level1=2, noise_level2=25):
+    noise_level = random.randint(noise_level1, noise_level2)
+    img = np.clip(img, 0.0, 1.0)
+    rnum = random.random()
+    if rnum > 0.6:
+        img += img * np.random.normal(0, noise_level / 255.0, img.shape).astype(np.float32)
+    elif rnum < 0.4:
+        img += img * np.random.normal(0, noise_level / 255.0, (*img.shape[:2], 1)).astype(np.float32)
+    else:
+        L = noise_level2 / 255.
+        D = np.diag(np.random.rand(3))
+        U = orth(np.random.rand(3, 3))
+        conv = np.dot(np.dot(np.transpose(U), D), U)
+        img += img * np.random.multivariate_normal([0, 0, 0], np.abs(L ** 2 * conv), img.shape[:2]).astype(np.float32)
+    img = np.clip(img, 0.0, 1.0)
+    return img
+
+
+def add_Poisson_noise(img):
+    img = np.clip((img * 255.0).round(), 0, 255) / 255.
+    vals = 10 ** (2 * random.random() + 2.0)  # [2, 4]
+    if random.random() < 0.5:
+        img = np.random.poisson(img * vals).astype(np.float32) / vals
+    else:
+        img_gray = np.dot(img[..., :3], [0.299, 0.587, 0.114])
+        img_gray = np.clip((img_gray * 255.0).round(), 0, 255) / 255.
+        noise_gray = np.random.poisson(img_gray * vals).astype(np.float32) / vals - img_gray
+        img += noise_gray[:, :, np.newaxis]
+    img = np.clip(img, 0.0, 1.0)
+    return img
+
+
+def add_JPEG_noise(img):
+    quality_factor = random.randint(30, 95)
+    img = cv2.cvtColor(util.single2uint(img), cv2.COLOR_RGB2BGR)
+    result, encimg = cv2.imencode('.jpg', img, [int(cv2.IMWRITE_JPEG_QUALITY), quality_factor])
+    img = cv2.imdecode(encimg, 1)
+    img = cv2.cvtColor(util.uint2single(img), cv2.COLOR_BGR2RGB)
+    return img
+
+
+def random_crop(lq, hq, sf=4, lq_patchsize=64):
+    h, w = lq.shape[:2]
+    rnd_h = random.randint(0, h - lq_patchsize)
+    rnd_w = random.randint(0, w - lq_patchsize)
+    lq = lq[rnd_h:rnd_h + lq_patchsize, rnd_w:rnd_w + lq_patchsize, :]
+
+    rnd_h_H, rnd_w_H = int(rnd_h * sf), int(rnd_w * sf)
+    hq = hq[rnd_h_H:rnd_h_H + lq_patchsize * sf, rnd_w_H:rnd_w_H + lq_patchsize * sf, :]
+    return lq, hq
+
+
+def degradation_bsrgan(img, sf=4, lq_patchsize=72, isp_model=None):
+    """
+    This is the degradation model of BSRGAN from the paper
+    "Designing a Practical Degradation Model for Deep Blind Image Super-Resolution"
+    ----------
+    img: HXWXC, [0, 1], its size should be large than (lq_patchsizexsf)x(lq_patchsizexsf)
+    sf: scale factor
+    isp_model: camera ISP model
+    Returns
+    -------
+    img: low-quality patch, size: lq_patchsizeXlq_patchsizeXC, range: [0, 1]
+    hq: corresponding high-quality patch, size: (lq_patchsizexsf)X(lq_patchsizexsf)XC, range: [0, 1]
+    """
+    isp_prob, jpeg_prob, scale2_prob = 0.25, 0.9, 0.25
+    sf_ori = sf
+
+    h1, w1 = img.shape[:2]
+    img = img.copy()[:w1 - w1 % sf, :h1 - h1 % sf, ...]  # mod crop
+    h, w = img.shape[:2]
+
+    if h < lq_patchsize * sf or w < lq_patchsize * sf:
+        raise ValueError(f'img size ({h1}X{w1}) is too small!')
+
+    hq = img.copy()
+
+    if sf == 4 and random.random() < scale2_prob:  # downsample1
+        if np.random.rand() < 0.5:
+            img = cv2.resize(img, (int(1 / 2 * img.shape[1]), int(1 / 2 * img.shape[0])),
+                             interpolation=random.choice([1, 2, 3]))
+        else:
+            img = util.imresize_np(img, 1 / 2, True)
+        img = np.clip(img, 0.0, 1.0)
+        sf = 2
+
+    shuffle_order = random.sample(range(7), 7)
+    idx1, idx2 = shuffle_order.index(2), shuffle_order.index(3)
+    if idx1 > idx2:  # keep downsample3 last
+        shuffle_order[idx1], shuffle_order[idx2] = shuffle_order[idx2], shuffle_order[idx1]
+
+    for i in shuffle_order:
+
+        if i == 0:
+            img = add_blur(img, sf=sf)
+
+        elif i == 1:
+            img = add_blur(img, sf=sf)
+
+        elif i == 2:
+            a, b = img.shape[1], img.shape[0]
+            # downsample2
+            if random.random() < 0.75:
+                sf1 = random.uniform(1, 2 * sf)
+                img = cv2.resize(img, (int(1 / sf1 * img.shape[1]), int(1 / sf1 * img.shape[0])),
+                                 interpolation=random.choice([1, 2, 3]))
+            else:
+                k = fspecial('gaussian', 25, random.uniform(0.1, 0.6 * sf))
+                k_shifted = shift_pixel(k, sf)
+                k_shifted = k_shifted / k_shifted.sum()  # blur with shifted kernel
+                img = ndimage.filters.convolve(img, np.expand_dims(k_shifted, axis=2), mode='mirror')
+                img = img[0::sf, 0::sf, ...]  # nearest downsampling
+            img = np.clip(img, 0.0, 1.0)
+
+        elif i == 3:
+            # downsample3
+            img = cv2.resize(img, (int(1 / sf * a), int(1 / sf * b)), interpolation=random.choice([1, 2, 3]))
+            img = np.clip(img, 0.0, 1.0)
+
+        elif i == 4:
+            # add Gaussian noise
+            img = add_Gaussian_noise(img, noise_level1=2, noise_level2=25)
+
+        elif i == 5:
+            # add JPEG noise
+            if random.random() < jpeg_prob:
+                img = add_JPEG_noise(img)
+
+        elif i == 6:
+            # add processed camera sensor noise
+            if random.random() < isp_prob and isp_model is not None:
+                with torch.no_grad():
+                    img, hq = isp_model.forward(img.copy(), hq)
+
+    # add final JPEG compression noise
+    img = add_JPEG_noise(img)
+
+    # random crop
+    img, hq = random_crop(img, hq, sf_ori, lq_patchsize)
+
+    return img, hq
+
+
+# todo no isp_model?
+def degradation_bsrgan_variant(image, sf=4, isp_model=None):
+    """
+    This is the degradation model of BSRGAN from the paper
+    "Designing a Practical Degradation Model for Deep Blind Image Super-Resolution"
+    ----------
+    sf: scale factor
+    isp_model: camera ISP model
+    Returns
+    -------
+    img: low-quality patch, size: lq_patchsizeXlq_patchsizeXC, range: [0, 1]
+    hq: corresponding high-quality patch, size: (lq_patchsizexsf)X(lq_patchsizexsf)XC, range: [0, 1]
+    """
+    image = util.uint2single(image)
+    isp_prob, jpeg_prob, scale2_prob = 0.25, 0.9, 0.25
+    sf_ori = sf
+
+    h1, w1 = image.shape[:2]
+    image = image.copy()[:w1 - w1 % sf, :h1 - h1 % sf, ...]  # mod crop
+    h, w = image.shape[:2]
+
+    hq = image.copy()
+
+    if sf == 4 and random.random() < scale2_prob:  # downsample1
+        if np.random.rand() < 0.5:
+            image = cv2.resize(image, (int(1 / 2 * image.shape[1]), int(1 / 2 * image.shape[0])),
+                               interpolation=random.choice([1, 2, 3]))
+        else:
+            image = util.imresize_np(image, 1 / 2, True)
+        image = np.clip(image, 0.0, 1.0)
+        sf = 2
+
+    shuffle_order = random.sample(range(7), 7)
+    idx1, idx2 = shuffle_order.index(2), shuffle_order.index(3)
+    if idx1 > idx2:  # keep downsample3 last
+        shuffle_order[idx1], shuffle_order[idx2] = shuffle_order[idx2], shuffle_order[idx1]
+
+    for i in shuffle_order:
+
+        if i == 0:
+            image = add_blur(image, sf=sf)
+
+        elif i == 1:
+            image = add_blur(image, sf=sf)
+
+        elif i == 2:
+            a, b = image.shape[1], image.shape[0]
+            # downsample2
+            if random.random() < 0.75:
+                sf1 = random.uniform(1, 2 * sf)
+                image = cv2.resize(image, (int(1 / sf1 * image.shape[1]), int(1 / sf1 * image.shape[0])),
+                                   interpolation=random.choice([1, 2, 3]))
+            else:
+                k = fspecial('gaussian', 25, random.uniform(0.1, 0.6 * sf))
+                k_shifted = shift_pixel(k, sf)
+                k_shifted = k_shifted / k_shifted.sum()  # blur with shifted kernel
+                image = ndimage.filters.convolve(image, np.expand_dims(k_shifted, axis=2), mode='mirror')
+                image = image[0::sf, 0::sf, ...]  # nearest downsampling
+            image = np.clip(image, 0.0, 1.0)
+
+        elif i == 3:
+            # downsample3
+            image = cv2.resize(image, (int(1 / sf * a), int(1 / sf * b)), interpolation=random.choice([1, 2, 3]))
+            image = np.clip(image, 0.0, 1.0)
+
+        elif i == 4:
+            # add Gaussian noise
+            image = add_Gaussian_noise(image, noise_level1=2, noise_level2=25)
+
+        elif i == 5:
+            # add JPEG noise
+            if random.random() < jpeg_prob:
+                image = add_JPEG_noise(image)
+
+        # elif i == 6:
+        #     # add processed camera sensor noise
+        #     if random.random() < isp_prob and isp_model is not None:
+        #         with torch.no_grad():
+        #             img, hq = isp_model.forward(img.copy(), hq)
+
+    # add final JPEG compression noise
+    image = add_JPEG_noise(image)
+    image = util.single2uint(image)
+    example = {"image":image}
+    return example
+
+
+# TODO incase there is a pickle error one needs to replace a += x with a = a + x in add_speckle_noise etc...
+def degradation_bsrgan_plus(img, sf=4, shuffle_prob=0.5, use_sharp=True, lq_patchsize=64, isp_model=None):
+    """
+    This is an extended degradation model by combining
+    the degradation models of BSRGAN and Real-ESRGAN
+    ----------
+    img: HXWXC, [0, 1], its size should be large than (lq_patchsizexsf)x(lq_patchsizexsf)
+    sf: scale factor
+    use_shuffle: the degradation shuffle
+    use_sharp: sharpening the img
+    Returns
+    -------
+    img: low-quality patch, size: lq_patchsizeXlq_patchsizeXC, range: [0, 1]
+    hq: corresponding high-quality patch, size: (lq_patchsizexsf)X(lq_patchsizexsf)XC, range: [0, 1]
+    """
+
+    h1, w1 = img.shape[:2]
+    img = img.copy()[:w1 - w1 % sf, :h1 - h1 % sf, ...]  # mod crop
+    h, w = img.shape[:2]
+
+    if h < lq_patchsize * sf or w < lq_patchsize * sf:
+        raise ValueError(f'img size ({h1}X{w1}) is too small!')
+
+    if use_sharp:
+        img = add_sharpening(img)
+    hq = img.copy()
+
+    if random.random() < shuffle_prob:
+        shuffle_order = random.sample(range(13), 13)
+    else:
+        shuffle_order = list(range(13))
+        # local shuffle for noise, JPEG is always the last one
+        shuffle_order[2:6] = random.sample(shuffle_order[2:6], len(range(2, 6)))
+        shuffle_order[9:13] = random.sample(shuffle_order[9:13], len(range(9, 13)))
+
+    poisson_prob, speckle_prob, isp_prob = 0.1, 0.1, 0.1
+
+    for i in shuffle_order:
+        if i == 0:
+            img = add_blur(img, sf=sf)
+        elif i == 1:
+            img = add_resize(img, sf=sf)
+        elif i == 2:
+            img = add_Gaussian_noise(img, noise_level1=2, noise_level2=25)
+        elif i == 3:
+            if random.random() < poisson_prob:
+                img = add_Poisson_noise(img)
+        elif i == 4:
+            if random.random() < speckle_prob:
+                img = add_speckle_noise(img)
+        elif i == 5:
+            if random.random() < isp_prob and isp_model is not None:
+                with torch.no_grad():
+                    img, hq = isp_model.forward(img.copy(), hq)
+        elif i == 6:
+            img = add_JPEG_noise(img)
+        elif i == 7:
+            img = add_blur(img, sf=sf)
+        elif i == 8:
+            img = add_resize(img, sf=sf)
+        elif i == 9:
+            img = add_Gaussian_noise(img, noise_level1=2, noise_level2=25)
+        elif i == 10:
+            if random.random() < poisson_prob:
+                img = add_Poisson_noise(img)
+        elif i == 11:
+            if random.random() < speckle_prob:
+                img = add_speckle_noise(img)
+        elif i == 12:
+            if random.random() < isp_prob and isp_model is not None:
+                with torch.no_grad():
+                    img, hq = isp_model.forward(img.copy(), hq)
+        else:
+            print('check the shuffle!')
+
+    # resize to desired size
+    img = cv2.resize(img, (int(1 / sf * hq.shape[1]), int(1 / sf * hq.shape[0])),
+                     interpolation=random.choice([1, 2, 3]))
+
+    # add final JPEG compression noise
+    img = add_JPEG_noise(img)
+
+    # random crop
+    img, hq = random_crop(img, hq, sf, lq_patchsize)
+
+    return img, hq
+
+
+if __name__ == '__main__':
+	print("hey")
+	img = util.imread_uint('utils/test.png', 3)
+	print(img)
+	img = util.uint2single(img)
+	print(img)
+	img = img[:448, :448]
+	h = img.shape[0] // 4
+	print("resizing to", h)
+	sf = 4
+	deg_fn = partial(degradation_bsrgan_variant, sf=sf)
+	for i in range(20):
+		print(i)
+		img_lq = deg_fn(img)
+		print(img_lq)
+		img_lq_bicubic = albumentations.SmallestMaxSize(max_size=h, interpolation=cv2.INTER_CUBIC)(image=img)["image"]
+		print(img_lq.shape)
+		print("bicubic", img_lq_bicubic.shape)
+		print(img_hq.shape)
+		lq_nearest = cv2.resize(util.single2uint(img_lq), (int(sf * img_lq.shape[1]), int(sf * img_lq.shape[0])),
+		                        interpolation=0)
+		lq_bicubic_nearest = cv2.resize(util.single2uint(img_lq_bicubic), (int(sf * img_lq.shape[1]), int(sf * img_lq.shape[0])),
+		                        interpolation=0)
+		img_concat = np.concatenate([lq_bicubic_nearest, lq_nearest, util.single2uint(img_hq)], axis=1)
+		util.imsave(img_concat, str(i) + '.png')
+
+

ldm/modules/image_degradation/bsrgan_light.py

@@ -0,0 +1,650 @@
+# -*- coding: utf-8 -*-
+import numpy as np
+import cv2
+import torch
+
+from functools import partial
+import random
+from scipy import ndimage
+import scipy
+import scipy.stats as ss
+from scipy.interpolate import interp2d
+from scipy.linalg import orth
+import albumentations
+
+import ldm.modules.image_degradation.utils_image as util
+
+"""
+# --------------------------------------------
+# Super-Resolution
+# --------------------------------------------
+#
+# Kai Zhang (cskaizhang@gmail.com)
+# https://github.com/cszn
+# From 2019/03--2021/08
+# --------------------------------------------
+"""
+
+
+def modcrop_np(img, sf):
+    '''
+    Args:
+        img: numpy image, WxH or WxHxC
+        sf: scale factor
+    Return:
+        cropped image
+    '''
+    w, h = img.shape[:2]
+    im = np.copy(img)
+    return im[:w - w % sf, :h - h % sf, ...]
+
+
+"""
+# --------------------------------------------
+# anisotropic Gaussian kernels
+# --------------------------------------------
+"""
+
+
+def analytic_kernel(k):
+    """Calculate the X4 kernel from the X2 kernel (for proof see appendix in paper)"""
+    k_size = k.shape[0]
+    # Calculate the big kernels size
+    big_k = np.zeros((3 * k_size - 2, 3 * k_size - 2))
+    # Loop over the small kernel to fill the big one
+    for r in range(k_size):
+        for c in range(k_size):
+            big_k[2 * r:2 * r + k_size, 2 * c:2 * c + k_size] += k[r, c] * k
+    # Crop the edges of the big kernel to ignore very small values and increase run time of SR
+    crop = k_size // 2
+    cropped_big_k = big_k[crop:-crop, crop:-crop]
+    # Normalize to 1
+    return cropped_big_k / cropped_big_k.sum()
+
+
+def anisotropic_Gaussian(ksize=15, theta=np.pi, l1=6, l2=6):
+    """ generate an anisotropic Gaussian kernel
+    Args:
+        ksize : e.g., 15, kernel size
+        theta : [0,  pi], rotation angle range
+        l1    : [0.1,50], scaling of eigenvalues
+        l2    : [0.1,l1], scaling of eigenvalues
+        If l1 = l2, will get an isotropic Gaussian kernel.
+    Returns:
+        k     : kernel
+    """
+
+    v = np.dot(np.array([[np.cos(theta), -np.sin(theta)], [np.sin(theta), np.cos(theta)]]), np.array([1., 0.]))
+    V = np.array([[v[0], v[1]], [v[1], -v[0]]])
+    D = np.array([[l1, 0], [0, l2]])
+    Sigma = np.dot(np.dot(V, D), np.linalg.inv(V))
+    k = gm_blur_kernel(mean=[0, 0], cov=Sigma, size=ksize)
+
+    return k
+
+
+def gm_blur_kernel(mean, cov, size=15):
+    center = size / 2.0 + 0.5
+    k = np.zeros([size, size])
+    for y in range(size):
+        for x in range(size):
+            cy = y - center + 1
+            cx = x - center + 1
+            k[y, x] = ss.multivariate_normal.pdf([cx, cy], mean=mean, cov=cov)
+
+    k = k / np.sum(k)
+    return k
+
+
+def shift_pixel(x, sf, upper_left=True):
+    """shift pixel for super-resolution with different scale factors
+    Args:
+        x: WxHxC or WxH
+        sf: scale factor
+        upper_left: shift direction
+    """
+    h, w = x.shape[:2]
+    shift = (sf - 1) * 0.5
+    xv, yv = np.arange(0, w, 1.0), np.arange(0, h, 1.0)
+    if upper_left:
+        x1 = xv + shift
+        y1 = yv + shift
+    else:
+        x1 = xv - shift
+        y1 = yv - shift
+
+    x1 = np.clip(x1, 0, w - 1)
+    y1 = np.clip(y1, 0, h - 1)
+
+    if x.ndim == 2:
+        x = interp2d(xv, yv, x)(x1, y1)
+    if x.ndim == 3:
+        for i in range(x.shape[-1]):
+            x[:, :, i] = interp2d(xv, yv, x[:, :, i])(x1, y1)
+
+    return x
+
+
+def blur(x, k):
+    '''
+    x: image, NxcxHxW
+    k: kernel, Nx1xhxw
+    '''
+    n, c = x.shape[:2]
+    p1, p2 = (k.shape[-2] - 1) // 2, (k.shape[-1] - 1) // 2
+    x = torch.nn.functional.pad(x, pad=(p1, p2, p1, p2), mode='replicate')
+    k = k.repeat(1, c, 1, 1)
+    k = k.view(-1, 1, k.shape[2], k.shape[3])
+    x = x.view(1, -1, x.shape[2], x.shape[3])
+    x = torch.nn.functional.conv2d(x, k, bias=None, stride=1, padding=0, groups=n * c)
+    x = x.view(n, c, x.shape[2], x.shape[3])
+
+    return x
+
+
+def gen_kernel(k_size=np.array([15, 15]), scale_factor=np.array([4, 4]), min_var=0.6, max_var=10., noise_level=0):
+    """"
+    # modified version of https://github.com/assafshocher/BlindSR_dataset_generator
+    # Kai Zhang
+    # min_var = 0.175 * sf  # variance of the gaussian kernel will be sampled between min_var and max_var
+    # max_var = 2.5 * sf
+    """
+    # Set random eigen-vals (lambdas) and angle (theta) for COV matrix
+    lambda_1 = min_var + np.random.rand() * (max_var - min_var)
+    lambda_2 = min_var + np.random.rand() * (max_var - min_var)
+    theta = np.random.rand() * np.pi  # random theta
+    noise = -noise_level + np.random.rand(*k_size) * noise_level * 2
+
+    # Set COV matrix using Lambdas and Theta
+    LAMBDA = np.diag([lambda_1, lambda_2])
+    Q = np.array([[np.cos(theta), -np.sin(theta)],
+                  [np.sin(theta), np.cos(theta)]])
+    SIGMA = Q @ LAMBDA @ Q.T
+    INV_SIGMA = np.linalg.inv(SIGMA)[None, None, :, :]
+
+    # Set expectation position (shifting kernel for aligned image)
+    MU = k_size // 2 - 0.5 * (scale_factor - 1)  # - 0.5 * (scale_factor - k_size % 2)
+    MU = MU[None, None, :, None]
+
+    # Create meshgrid for Gaussian
+    [X, Y] = np.meshgrid(range(k_size[0]), range(k_size[1]))
+    Z = np.stack([X, Y], 2)[:, :, :, None]
+
+    # Calcualte Gaussian for every pixel of the kernel
+    ZZ = Z - MU
+    ZZ_t = ZZ.transpose(0, 1, 3, 2)
+    raw_kernel = np.exp(-0.5 * np.squeeze(ZZ_t @ INV_SIGMA @ ZZ)) * (1 + noise)
+
+    # shift the kernel so it will be centered
+    # raw_kernel_centered = kernel_shift(raw_kernel, scale_factor)
+
+    # Normalize the kernel and return
+    # kernel = raw_kernel_centered / np.sum(raw_kernel_centered)
+    kernel = raw_kernel / np.sum(raw_kernel)
+    return kernel
+
+
+def fspecial_gaussian(hsize, sigma):
+    hsize = [hsize, hsize]
+    siz = [(hsize[0] - 1.0) / 2.0, (hsize[1] - 1.0) / 2.0]
+    std = sigma
+    [x, y] = np.meshgrid(np.arange(-siz[1], siz[1] + 1), np.arange(-siz[0], siz[0] + 1))
+    arg = -(x * x + y * y) / (2 * std * std)
+    h = np.exp(arg)
+    h[h < scipy.finfo(float).eps * h.max()] = 0
+    sumh = h.sum()
+    if sumh != 0:
+        h = h / sumh
+    return h
+
+
+def fspecial_laplacian(alpha):
+    alpha = max([0, min([alpha, 1])])
+    h1 = alpha / (alpha + 1)
+    h2 = (1 - alpha) / (alpha + 1)
+    h = [[h1, h2, h1], [h2, -4 / (alpha + 1), h2], [h1, h2, h1]]
+    h = np.array(h)
+    return h
+
+
+def fspecial(filter_type, *args, **kwargs):
+    '''
+    python code from:
+    https://github.com/ronaldosena/imagens-medicas-2/blob/40171a6c259edec7827a6693a93955de2bd39e76/Aulas/aula_2_-_uniform_filter/matlab_fspecial.py
+    '''
+    if filter_type == 'gaussian':
+        return fspecial_gaussian(*args, **kwargs)
+    if filter_type == 'laplacian':
+        return fspecial_laplacian(*args, **kwargs)
+
+
+"""
+# --------------------------------------------
+# degradation models
+# --------------------------------------------
+"""
+
+
+def bicubic_degradation(x, sf=3):
+    '''
+    Args:
+        x: HxWxC image, [0, 1]
+        sf: down-scale factor
+    Return:
+        bicubicly downsampled LR image
+    '''
+    x = util.imresize_np(x, scale=1 / sf)
+    return x
+
+
+def srmd_degradation(x, k, sf=3):
+    ''' blur + bicubic downsampling
+    Args:
+        x: HxWxC image, [0, 1]
+        k: hxw, double
+        sf: down-scale factor
+    Return:
+        downsampled LR image
+    Reference:
+        @inproceedings{zhang2018learning,
+          title={Learning a single convolutional super-resolution network for multiple degradations},
+          author={Zhang, Kai and Zuo, Wangmeng and Zhang, Lei},
+          booktitle={IEEE Conference on Computer Vision and Pattern Recognition},
+          pages={3262--3271},
+          year={2018}
+        }
+    '''
+    x = ndimage.convolve(x, np.expand_dims(k, axis=2), mode='wrap')  # 'nearest' | 'mirror'
+    x = bicubic_degradation(x, sf=sf)
+    return x
+
+
+def dpsr_degradation(x, k, sf=3):
+    ''' bicubic downsampling + blur
+    Args:
+        x: HxWxC image, [0, 1]
+        k: hxw, double
+        sf: down-scale factor
+    Return:
+        downsampled LR image
+    Reference:
+        @inproceedings{zhang2019deep,
+          title={Deep Plug-and-Play Super-Resolution for Arbitrary Blur Kernels},
+          author={Zhang, Kai and Zuo, Wangmeng and Zhang, Lei},
+          booktitle={IEEE Conference on Computer Vision and Pattern Recognition},
+          pages={1671--1681},
+          year={2019}
+        }
+    '''
+    x = bicubic_degradation(x, sf=sf)
+    x = ndimage.convolve(x, np.expand_dims(k, axis=2), mode='wrap')
+    return x
+
+
+def classical_degradation(x, k, sf=3):
+    ''' blur + downsampling
+    Args:
+        x: HxWxC image, [0, 1]/[0, 255]
+        k: hxw, double
+        sf: down-scale factor
+    Return:
+        downsampled LR image
+    '''
+    x = ndimage.convolve(x, np.expand_dims(k, axis=2), mode='wrap')
+    # x = filters.correlate(x, np.expand_dims(np.flip(k), axis=2))
+    st = 0
+    return x[st::sf, st::sf, ...]
+
+
+def add_sharpening(img, weight=0.5, radius=50, threshold=10):
+    """USM sharpening. borrowed from real-ESRGAN
+    Input image: I; Blurry image: B.
+    1. K = I + weight * (I - B)
+    2. Mask = 1 if abs(I - B) > threshold, else: 0
+    3. Blur mask:
+    4. Out = Mask * K + (1 - Mask) * I
+    Args:
+        img (Numpy array): Input image, HWC, BGR; float32, [0, 1].
+        weight (float): Sharp weight. Default: 1.
+        radius (float): Kernel size of Gaussian blur. Default: 50.
+        threshold (int):
+    """
+    if radius % 2 == 0:
+        radius += 1
+    blur = cv2.GaussianBlur(img, (radius, radius), 0)
+    residual = img - blur
+    mask = np.abs(residual) * 255 > threshold
+    mask = mask.astype('float32')
+    soft_mask = cv2.GaussianBlur(mask, (radius, radius), 0)
+
+    K = img + weight * residual
+    K = np.clip(K, 0, 1)
+    return soft_mask * K + (1 - soft_mask) * img
+
+
+def add_blur(img, sf=4):
+    wd2 = 4.0 + sf
+    wd = 2.0 + 0.2 * sf
+
+    wd2 = wd2/4
+    wd = wd/4
+
+    if random.random() < 0.5:
+        l1 = wd2 * random.random()
+        l2 = wd2 * random.random()
+        k = anisotropic_Gaussian(ksize=random.randint(2, 11) + 3, theta=random.random() * np.pi, l1=l1, l2=l2)
+    else:
+        k = fspecial('gaussian', random.randint(2, 4) + 3, wd * random.random())
+    img = ndimage.convolve(img, np.expand_dims(k, axis=2), mode='mirror')
+
+    return img
+
+
+def add_resize(img, sf=4):
+    rnum = np.random.rand()
+    if rnum > 0.8:  # up
+        sf1 = random.uniform(1, 2)
+    elif rnum < 0.7:  # down
+        sf1 = random.uniform(0.5 / sf, 1)
+    else:
+        sf1 = 1.0
+    img = cv2.resize(img, (int(sf1 * img.shape[1]), int(sf1 * img.shape[0])), interpolation=random.choice([1, 2, 3]))
+    img = np.clip(img, 0.0, 1.0)
+
+    return img
+
+
+# def add_Gaussian_noise(img, noise_level1=2, noise_level2=25):
+#     noise_level = random.randint(noise_level1, noise_level2)
+#     rnum = np.random.rand()
+#     if rnum > 0.6:  # add color Gaussian noise
+#         img += np.random.normal(0, noise_level / 255.0, img.shape).astype(np.float32)
+#     elif rnum < 0.4:  # add grayscale Gaussian noise
+#         img += np.random.normal(0, noise_level / 255.0, (*img.shape[:2], 1)).astype(np.float32)
+#     else:  # add  noise
+#         L = noise_level2 / 255.
+#         D = np.diag(np.random.rand(3))
+#         U = orth(np.random.rand(3, 3))
+#         conv = np.dot(np.dot(np.transpose(U), D), U)
+#         img += np.random.multivariate_normal([0, 0, 0], np.abs(L ** 2 * conv), img.shape[:2]).astype(np.float32)
+#     img = np.clip(img, 0.0, 1.0)
+#     return img
+
+def add_Gaussian_noise(img, noise_level1=2, noise_level2=25):
+    noise_level = random.randint(noise_level1, noise_level2)
+    rnum = np.random.rand()
+    if rnum > 0.6:  # add color Gaussian noise
+        img = img + np.random.normal(0, noise_level / 255.0, img.shape).astype(np.float32)
+    elif rnum < 0.4:  # add grayscale Gaussian noise
+        img = img + np.random.normal(0, noise_level / 255.0, (*img.shape[:2], 1)).astype(np.float32)
+    else:  # add  noise
+        L = noise_level2 / 255.
+        D = np.diag(np.random.rand(3))
+        U = orth(np.random.rand(3, 3))
+        conv = np.dot(np.dot(np.transpose(U), D), U)
+        img = img + np.random.multivariate_normal([0, 0, 0], np.abs(L ** 2 * conv), img.shape[:2]).astype(np.float32)
+    img = np.clip(img, 0.0, 1.0)
+    return img
+
+
+def add_speckle_noise(img, noise_level1=2, noise_level2=25):
+    noise_level = random.randint(noise_level1, noise_level2)
+    img = np.clip(img, 0.0, 1.0)
+    rnum = random.random()
+    if rnum > 0.6:
+        img += img * np.random.normal(0, noise_level / 255.0, img.shape).astype(np.float32)
+    elif rnum < 0.4:
+        img += img * np.random.normal(0, noise_level / 255.0, (*img.shape[:2], 1)).astype(np.float32)
+    else:
+        L = noise_level2 / 255.
+        D = np.diag(np.random.rand(3))
+        U = orth(np.random.rand(3, 3))
+        conv = np.dot(np.dot(np.transpose(U), D), U)
+        img += img * np.random.multivariate_normal([0, 0, 0], np.abs(L ** 2 * conv), img.shape[:2]).astype(np.float32)
+    img = np.clip(img, 0.0, 1.0)
+    return img
+
+
+def add_Poisson_noise(img):
+    img = np.clip((img * 255.0).round(), 0, 255) / 255.
+    vals = 10 ** (2 * random.random() + 2.0)  # [2, 4]
+    if random.random() < 0.5:
+        img = np.random.poisson(img * vals).astype(np.float32) / vals
+    else:
+        img_gray = np.dot(img[..., :3], [0.299, 0.587, 0.114])
+        img_gray = np.clip((img_gray * 255.0).round(), 0, 255) / 255.
+        noise_gray = np.random.poisson(img_gray * vals).astype(np.float32) / vals - img_gray
+        img += noise_gray[:, :, np.newaxis]
+    img = np.clip(img, 0.0, 1.0)
+    return img
+
+
+def add_JPEG_noise(img):
+    quality_factor = random.randint(80, 95)
+    img = cv2.cvtColor(util.single2uint(img), cv2.COLOR_RGB2BGR)
+    result, encimg = cv2.imencode('.jpg', img, [int(cv2.IMWRITE_JPEG_QUALITY), quality_factor])
+    img = cv2.imdecode(encimg, 1)
+    img = cv2.cvtColor(util.uint2single(img), cv2.COLOR_BGR2RGB)
+    return img
+
+
+def random_crop(lq, hq, sf=4, lq_patchsize=64):
+    h, w = lq.shape[:2]
+    rnd_h = random.randint(0, h - lq_patchsize)
+    rnd_w = random.randint(0, w - lq_patchsize)
+    lq = lq[rnd_h:rnd_h + lq_patchsize, rnd_w:rnd_w + lq_patchsize, :]
+
+    rnd_h_H, rnd_w_H = int(rnd_h * sf), int(rnd_w * sf)
+    hq = hq[rnd_h_H:rnd_h_H + lq_patchsize * sf, rnd_w_H:rnd_w_H + lq_patchsize * sf, :]
+    return lq, hq
+
+
+def degradation_bsrgan(img, sf=4, lq_patchsize=72, isp_model=None):
+    """
+    This is the degradation model of BSRGAN from the paper
+    "Designing a Practical Degradation Model for Deep Blind Image Super-Resolution"
+    ----------
+    img: HXWXC, [0, 1], its size should be large than (lq_patchsizexsf)x(lq_patchsizexsf)
+    sf: scale factor
+    isp_model: camera ISP model
+    Returns
+    -------
+    img: low-quality patch, size: lq_patchsizeXlq_patchsizeXC, range: [0, 1]
+    hq: corresponding high-quality patch, size: (lq_patchsizexsf)X(lq_patchsizexsf)XC, range: [0, 1]
+    """
+    isp_prob, jpeg_prob, scale2_prob = 0.25, 0.9, 0.25
+    sf_ori = sf
+
+    h1, w1 = img.shape[:2]
+    img = img.copy()[:w1 - w1 % sf, :h1 - h1 % sf, ...]  # mod crop
+    h, w = img.shape[:2]
+
+    if h < lq_patchsize * sf or w < lq_patchsize * sf:
+        raise ValueError(f'img size ({h1}X{w1}) is too small!')
+
+    hq = img.copy()
+
+    if sf == 4 and random.random() < scale2_prob:  # downsample1
+        if np.random.rand() < 0.5:
+            img = cv2.resize(img, (int(1 / 2 * img.shape[1]), int(1 / 2 * img.shape[0])),
+                             interpolation=random.choice([1, 2, 3]))
+        else:
+            img = util.imresize_np(img, 1 / 2, True)
+        img = np.clip(img, 0.0, 1.0)
+        sf = 2
+
+    shuffle_order = random.sample(range(7), 7)
+    idx1, idx2 = shuffle_order.index(2), shuffle_order.index(3)
+    if idx1 > idx2:  # keep downsample3 last
+        shuffle_order[idx1], shuffle_order[idx2] = shuffle_order[idx2], shuffle_order[idx1]
+
+    for i in shuffle_order:
+
+        if i == 0:
+            img = add_blur(img, sf=sf)
+
+        elif i == 1:
+            img = add_blur(img, sf=sf)
+
+        elif i == 2:
+            a, b = img.shape[1], img.shape[0]
+            # downsample2
+            if random.random() < 0.75:
+                sf1 = random.uniform(1, 2 * sf)
+                img = cv2.resize(img, (int(1 / sf1 * img.shape[1]), int(1 / sf1 * img.shape[0])),
+                                 interpolation=random.choice([1, 2, 3]))
+            else:
+                k = fspecial('gaussian', 25, random.uniform(0.1, 0.6 * sf))
+                k_shifted = shift_pixel(k, sf)
+                k_shifted = k_shifted / k_shifted.sum()  # blur with shifted kernel
+                img = ndimage.convolve(img, np.expand_dims(k_shifted, axis=2), mode='mirror')
+                img = img[0::sf, 0::sf, ...]  # nearest downsampling
+            img = np.clip(img, 0.0, 1.0)
+
+        elif i == 3:
+            # downsample3
+            img = cv2.resize(img, (int(1 / sf * a), int(1 / sf * b)), interpolation=random.choice([1, 2, 3]))
+            img = np.clip(img, 0.0, 1.0)
+
+        elif i == 4:
+            # add Gaussian noise
+            img = add_Gaussian_noise(img, noise_level1=2, noise_level2=8)
+
+        elif i == 5:
+            # add JPEG noise
+            if random.random() < jpeg_prob:
+                img = add_JPEG_noise(img)
+
+        elif i == 6:
+            # add processed camera sensor noise
+            if random.random() < isp_prob and isp_model is not None:
+                with torch.no_grad():
+                    img, hq = isp_model.forward(img.copy(), hq)
+
+    # add final JPEG compression noise
+    img = add_JPEG_noise(img)
+
+    # random crop
+    img, hq = random_crop(img, hq, sf_ori, lq_patchsize)
+
+    return img, hq
+
+
+# todo no isp_model?
+def degradation_bsrgan_variant(image, sf=4, isp_model=None):
+    """
+    This is the degradation model of BSRGAN from the paper
+    "Designing a Practical Degradation Model for Deep Blind Image Super-Resolution"
+    ----------
+    sf: scale factor
+    isp_model: camera ISP model
+    Returns
+    -------
+    img: low-quality patch, size: lq_patchsizeXlq_patchsizeXC, range: [0, 1]
+    hq: corresponding high-quality patch, size: (lq_patchsizexsf)X(lq_patchsizexsf)XC, range: [0, 1]
+    """
+    image = util.uint2single(image)
+    isp_prob, jpeg_prob, scale2_prob = 0.25, 0.9, 0.25
+    sf_ori = sf
+
+    h1, w1 = image.shape[:2]
+    image = image.copy()[:w1 - w1 % sf, :h1 - h1 % sf, ...]  # mod crop
+    h, w = image.shape[:2]
+
+    hq = image.copy()
+
+    if sf == 4 and random.random() < scale2_prob:  # downsample1
+        if np.random.rand() < 0.5:
+            image = cv2.resize(image, (int(1 / 2 * image.shape[1]), int(1 / 2 * image.shape[0])),
+                               interpolation=random.choice([1, 2, 3]))
+        else:
+            image = util.imresize_np(image, 1 / 2, True)
+        image = np.clip(image, 0.0, 1.0)
+        sf = 2
+
+    shuffle_order = random.sample(range(7), 7)
+    idx1, idx2 = shuffle_order.index(2), shuffle_order.index(3)
+    if idx1 > idx2:  # keep downsample3 last
+        shuffle_order[idx1], shuffle_order[idx2] = shuffle_order[idx2], shuffle_order[idx1]
+
+    for i in shuffle_order:
+
+        if i == 0:
+            image = add_blur(image, sf=sf)
+
+        # elif i == 1:
+        #     image = add_blur(image, sf=sf)
+
+        if i == 0:
+            pass
+
+        elif i == 2:
+            a, b = image.shape[1], image.shape[0]
+            # downsample2
+            if random.random() < 0.8:
+                sf1 = random.uniform(1, 2 * sf)
+                image = cv2.resize(image, (int(1 / sf1 * image.shape[1]), int(1 / sf1 * image.shape[0])),
+                                   interpolation=random.choice([1, 2, 3]))
+            else:
+                k = fspecial('gaussian', 25, random.uniform(0.1, 0.6 * sf))
+                k_shifted = shift_pixel(k, sf)
+                k_shifted = k_shifted / k_shifted.sum()  # blur with shifted kernel
+                image = ndimage.convolve(image, np.expand_dims(k_shifted, axis=2), mode='mirror')
+                image = image[0::sf, 0::sf, ...]  # nearest downsampling
+
+            image = np.clip(image, 0.0, 1.0)
+
+        elif i == 3:
+            # downsample3
+            image = cv2.resize(image, (int(1 / sf * a), int(1 / sf * b)), interpolation=random.choice([1, 2, 3]))
+            image = np.clip(image, 0.0, 1.0)
+
+        elif i == 4:
+            # add Gaussian noise
+            image = add_Gaussian_noise(image, noise_level1=1, noise_level2=2)
+
+        elif i == 5:
+            # add JPEG noise
+            if random.random() < jpeg_prob:
+                image = add_JPEG_noise(image)
+        #
+        # elif i == 6:
+        #     # add processed camera sensor noise
+        #     if random.random() < isp_prob and isp_model is not None:
+        #         with torch.no_grad():
+        #             img, hq = isp_model.forward(img.copy(), hq)
+
+    # add final JPEG compression noise
+    image = add_JPEG_noise(image)
+    image = util.single2uint(image)
+    example = {"image": image}
+    return example
+
+
+
+
+if __name__ == '__main__':
+    print("hey")
+    img = util.imread_uint('utils/test.png', 3)
+    img = img[:448, :448]
+    h = img.shape[0] // 4
+    print("resizing to", h)
+    sf = 4
+    deg_fn = partial(degradation_bsrgan_variant, sf=sf)
+    for i in range(20):
+        print(i)
+        img_hq = img
+        img_lq = deg_fn(img)["image"]
+        img_hq, img_lq = util.uint2single(img_hq), util.uint2single(img_lq)
+        print(img_lq)
+        img_lq_bicubic = albumentations.SmallestMaxSize(max_size=h, interpolation=cv2.INTER_CUBIC)(image=img_hq)["image"]
+        print(img_lq.shape)
+        print("bicubic", img_lq_bicubic.shape)
+        print(img_hq.shape)
+        lq_nearest = cv2.resize(util.single2uint(img_lq), (int(sf * img_lq.shape[1]), int(sf * img_lq.shape[0])),
+                                interpolation=0)
+        lq_bicubic_nearest = cv2.resize(util.single2uint(img_lq_bicubic),
+                                        (int(sf * img_lq.shape[1]), int(sf * img_lq.shape[0])),
+                                        interpolation=0)
+        img_concat = np.concatenate([lq_bicubic_nearest, lq_nearest, util.single2uint(img_hq)], axis=1)
+        util.imsave(img_concat, str(i) + '.png')

ldm/modules/image_degradation/utils_image.py

@@ -0,0 +1,916 @@
+import os
+import math
+import random
+import numpy as np
+import torch
+import cv2
+from torchvision.utils import make_grid
+from datetime import datetime
+#import matplotlib.pyplot as plt   # TODO: check with Dominik, also bsrgan.py vs bsrgan_light.py
+
+
+os.environ["KMP_DUPLICATE_LIB_OK"]="TRUE"
+
+
+'''
+# --------------------------------------------
+# Kai Zhang (github: https://github.com/cszn)
+# 03/Mar/2019
+# --------------------------------------------
+# https://github.com/twhui/SRGAN-pyTorch
+# https://github.com/xinntao/BasicSR
+# --------------------------------------------
+'''
+
+
+IMG_EXTENSIONS = ['.jpg', '.JPG', '.jpeg', '.JPEG', '.png', '.PNG', '.ppm', '.PPM', '.bmp', '.BMP', '.tif']
+
+
+def is_image_file(filename):
+    return any(filename.endswith(extension) for extension in IMG_EXTENSIONS)
+
+
+def get_timestamp():
+    return datetime.now().strftime('%y%m%d-%H%M%S')
+
+
+def imshow(x, title=None, cbar=False, figsize=None):
+    plt.figure(figsize=figsize)
+    plt.imshow(np.squeeze(x), interpolation='nearest', cmap='gray')
+    if title:
+        plt.title(title)
+    if cbar:
+        plt.colorbar()
+    plt.show()
+
+
+def surf(Z, cmap='rainbow', figsize=None):
+    plt.figure(figsize=figsize)
+    ax3 = plt.axes(projection='3d')
+
+    w, h = Z.shape[:2]
+    xx = np.arange(0,w,1)
+    yy = np.arange(0,h,1)
+    X, Y = np.meshgrid(xx, yy)
+    ax3.plot_surface(X,Y,Z,cmap=cmap)
+    #ax3.contour(X,Y,Z, zdim='z',offset=-2，cmap=cmap)
+    plt.show()
+
+
+'''
+# --------------------------------------------
+# get image pathes
+# --------------------------------------------
+'''
+
+
+def get_image_paths(dataroot):
+    paths = None  # return None if dataroot is None
+    if dataroot is not None:
+        paths = sorted(_get_paths_from_images(dataroot))
+    return paths
+
+
+def _get_paths_from_images(path):
+    assert os.path.isdir(path), '{:s} is not a valid directory'.format(path)
+    images = []
+    for dirpath, _, fnames in sorted(os.walk(path)):
+        for fname in sorted(fnames):
+            if is_image_file(fname):
+                img_path = os.path.join(dirpath, fname)
+                images.append(img_path)
+    assert images, '{:s} has no valid image file'.format(path)
+    return images
+
+
+'''
+# --------------------------------------------
+# split large images into small images 
+# --------------------------------------------
+'''
+
+
+def patches_from_image(img, p_size=512, p_overlap=64, p_max=800):
+    w, h = img.shape[:2]
+    patches = []
+    if w > p_max and h > p_max:
+        w1 = list(np.arange(0, w-p_size, p_size-p_overlap, dtype=np.int))
+        h1 = list(np.arange(0, h-p_size, p_size-p_overlap, dtype=np.int))
+        w1.append(w-p_size)
+        h1.append(h-p_size)
+#        print(w1)
+#        print(h1)
+        for i in w1:
+            for j in h1:
+                patches.append(img[i:i+p_size, j:j+p_size,:])
+    else:
+        patches.append(img)
+
+    return patches
+
+
+def imssave(imgs, img_path):
+    """
+    imgs: list, N images of size WxHxC
+    """
+    img_name, ext = os.path.splitext(os.path.basename(img_path))
+
+    for i, img in enumerate(imgs):
+        if img.ndim == 3:
+            img = img[:, :, [2, 1, 0]]
+        new_path = os.path.join(os.path.dirname(img_path), img_name+str('_s{:04d}'.format(i))+'.png')
+        cv2.imwrite(new_path, img)
+
+
+def split_imageset(original_dataroot, taget_dataroot, n_channels=3, p_size=800, p_overlap=96, p_max=1000):
+    """
+    split the large images from original_dataroot into small overlapped images with size (p_size)x(p_size),
+    and save them into taget_dataroot; only the images with larger size than (p_max)x(p_max)
+    will be splitted.
+    Args:
+        original_dataroot:
+        taget_dataroot:
+        p_size: size of small images
+        p_overlap: patch size in training is a good choice
+        p_max: images with smaller size than (p_max)x(p_max) keep unchanged.
+    """
+    paths = get_image_paths(original_dataroot)
+    for img_path in paths:
+        # img_name, ext = os.path.splitext(os.path.basename(img_path))
+        img = imread_uint(img_path, n_channels=n_channels)
+        patches = patches_from_image(img, p_size, p_overlap, p_max)
+        imssave(patches, os.path.join(taget_dataroot,os.path.basename(img_path)))
+        #if original_dataroot == taget_dataroot:
+        #del img_path
+
+'''
+# --------------------------------------------
+# makedir
+# --------------------------------------------
+'''
+
+
+def mkdir(path):
+    if not os.path.exists(path):
+        os.makedirs(path)
+
+
+def mkdirs(paths):
+    if isinstance(paths, str):
+        mkdir(paths)
+    else:
+        for path in paths:
+            mkdir(path)
+
+
+def mkdir_and_rename(path):
+    if os.path.exists(path):
+        new_name = path + '_archived_' + get_timestamp()
+        print('Path already exists. Rename it to [{:s}]'.format(new_name))
+        os.rename(path, new_name)
+    os.makedirs(path)
+
+
+'''
+# --------------------------------------------
+# read image from path
+# opencv is fast, but read BGR numpy image
+# --------------------------------------------
+'''
+
+
+# --------------------------------------------
+# get uint8 image of size HxWxn_channles (RGB)
+# --------------------------------------------
+def imread_uint(path, n_channels=3):
+    #  input: path
+    # output: HxWx3(RGB or GGG), or HxWx1 (G)
+    if n_channels == 1:
+        img = cv2.imread(path, 0)  # cv2.IMREAD_GRAYSCALE
+        img = np.expand_dims(img, axis=2)  # HxWx1
+    elif n_channels == 3:
+        img = cv2.imread(path, cv2.IMREAD_UNCHANGED)  # BGR or G
+        if img.ndim == 2:
+            img = cv2.cvtColor(img, cv2.COLOR_GRAY2RGB)  # GGG
+        else:
+            img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)  # RGB
+    return img
+
+
+# --------------------------------------------
+# matlab's imwrite
+# --------------------------------------------
+def imsave(img, img_path):
+    img = np.squeeze(img)
+    if img.ndim == 3:
+        img = img[:, :, [2, 1, 0]]
+    cv2.imwrite(img_path, img)
+
+def imwrite(img, img_path):
+    img = np.squeeze(img)
+    if img.ndim == 3:
+        img = img[:, :, [2, 1, 0]]
+    cv2.imwrite(img_path, img)
+
+
+
+# --------------------------------------------
+# get single image of size HxWxn_channles (BGR)
+# --------------------------------------------
+def read_img(path):
+    # read image by cv2
+    # return: Numpy float32, HWC, BGR, [0,1]
+    img = cv2.imread(path, cv2.IMREAD_UNCHANGED)  # cv2.IMREAD_GRAYSCALE
+    img = img.astype(np.float32) / 255.
+    if img.ndim == 2:
+        img = np.expand_dims(img, axis=2)
+    # some images have 4 channels
+    if img.shape[2] > 3:
+        img = img[:, :, :3]
+    return img
+
+
+'''
+# --------------------------------------------
+# image format conversion
+# --------------------------------------------
+# numpy(single) <--->  numpy(unit)
+# numpy(single) <--->  tensor
+# numpy(unit)   <--->  tensor
+# --------------------------------------------
+'''
+
+
+# --------------------------------------------
+# numpy(single) [0, 1] <--->  numpy(unit)
+# --------------------------------------------
+
+
+def uint2single(img):
+
+    return np.float32(img/255.)
+
+
+def single2uint(img):
+
+    return np.uint8((img.clip(0, 1)*255.).round())
+
+
+def uint162single(img):
+
+    return np.float32(img/65535.)
+
+
+def single2uint16(img):
+
+    return np.uint16((img.clip(0, 1)*65535.).round())
+
+
+# --------------------------------------------
+# numpy(unit) (HxWxC or HxW) <--->  tensor
+# --------------------------------------------
+
+
+# convert uint to 4-dimensional torch tensor
+def uint2tensor4(img):
+    if img.ndim == 2:
+        img = np.expand_dims(img, axis=2)
+    return torch.from_numpy(np.ascontiguousarray(img)).permute(2, 0, 1).float().div(255.).unsqueeze(0)
+
+
+# convert uint to 3-dimensional torch tensor
+def uint2tensor3(img):
+    if img.ndim == 2:
+        img = np.expand_dims(img, axis=2)
+    return torch.from_numpy(np.ascontiguousarray(img)).permute(2, 0, 1).float().div(255.)
+
+
+# convert 2/3/4-dimensional torch tensor to uint
+def tensor2uint(img):
+    img = img.data.squeeze().float().clamp_(0, 1).cpu().numpy()
+    if img.ndim == 3:
+        img = np.transpose(img, (1, 2, 0))
+    return np.uint8((img*255.0).round())
+
+
+# --------------------------------------------
+# numpy(single) (HxWxC) <--->  tensor
+# --------------------------------------------
+
+
+# convert single (HxWxC) to 3-dimensional torch tensor
+def single2tensor3(img):
+    return torch.from_numpy(np.ascontiguousarray(img)).permute(2, 0, 1).float()
+
+
+# convert single (HxWxC) to 4-dimensional torch tensor
+def single2tensor4(img):
+    return torch.from_numpy(np.ascontiguousarray(img)).permute(2, 0, 1).float().unsqueeze(0)
+
+
+# convert torch tensor to single
+def tensor2single(img):
+    img = img.data.squeeze().float().cpu().numpy()
+    if img.ndim == 3:
+        img = np.transpose(img, (1, 2, 0))
+
+    return img
+
+# convert torch tensor to single
+def tensor2single3(img):
+    img = img.data.squeeze().float().cpu().numpy()
+    if img.ndim == 3:
+        img = np.transpose(img, (1, 2, 0))
+    elif img.ndim == 2:
+        img = np.expand_dims(img, axis=2)
+    return img
+
+
+def single2tensor5(img):
+    return torch.from_numpy(np.ascontiguousarray(img)).permute(2, 0, 1, 3).float().unsqueeze(0)
+
+
+def single32tensor5(img):
+    return torch.from_numpy(np.ascontiguousarray(img)).float().unsqueeze(0).unsqueeze(0)
+
+
+def single42tensor4(img):
+    return torch.from_numpy(np.ascontiguousarray(img)).permute(2, 0, 1, 3).float()
+
+
+# from skimage.io import imread, imsave
+def tensor2img(tensor, out_type=np.uint8, min_max=(0, 1)):
+    '''
+    Converts a torch Tensor into an image Numpy array of BGR channel order
+    Input: 4D(B,(3/1),H,W), 3D(C,H,W), or 2D(H,W), any range, RGB channel order
+    Output: 3D(H,W,C) or 2D(H,W), [0,255], np.uint8 (default)
+    '''
+    tensor = tensor.squeeze().float().cpu().clamp_(*min_max)  # squeeze first, then clamp
+    tensor = (tensor - min_max[0]) / (min_max[1] - min_max[0])  # to range [0,1]
+    n_dim = tensor.dim()
+    if n_dim == 4:
+        n_img = len(tensor)
+        img_np = make_grid(tensor, nrow=int(math.sqrt(n_img)), normalize=False).numpy()
+        img_np = np.transpose(img_np[[2, 1, 0], :, :], (1, 2, 0))  # HWC, BGR
+    elif n_dim == 3:
+        img_np = tensor.numpy()
+        img_np = np.transpose(img_np[[2, 1, 0], :, :], (1, 2, 0))  # HWC, BGR
+    elif n_dim == 2:
+        img_np = tensor.numpy()
+    else:
+        raise TypeError(
+            'Only support 4D, 3D and 2D tensor. But received with dimension: {:d}'.format(n_dim))
+    if out_type == np.uint8:
+        img_np = (img_np * 255.0).round()
+        # Important. Unlike matlab, numpy.unit8() WILL NOT round by default.
+    return img_np.astype(out_type)
+
+
+'''
+# --------------------------------------------
+# Augmentation, flipe and/or rotate
+# --------------------------------------------
+# The following two are enough.
+# (1) augmet_img: numpy image of WxHxC or WxH
+# (2) augment_img_tensor4: tensor image 1xCxWxH
+# --------------------------------------------
+'''
+
+
+def augment_img(img, mode=0):
+    '''Kai Zhang (github: https://github.com/cszn)
+    '''
+    if mode == 0:
+        return img
+    elif mode == 1:
+        return np.flipud(np.rot90(img))
+    elif mode == 2:
+        return np.flipud(img)
+    elif mode == 3:
+        return np.rot90(img, k=3)
+    elif mode == 4:
+        return np.flipud(np.rot90(img, k=2))
+    elif mode == 5:
+        return np.rot90(img)
+    elif mode == 6:
+        return np.rot90(img, k=2)
+    elif mode == 7:
+        return np.flipud(np.rot90(img, k=3))
+
+
+def augment_img_tensor4(img, mode=0):
+    '''Kai Zhang (github: https://github.com/cszn)
+    '''
+    if mode == 0:
+        return img
+    elif mode == 1:
+        return img.rot90(1, [2, 3]).flip([2])
+    elif mode == 2:
+        return img.flip([2])
+    elif mode == 3:
+        return img.rot90(3, [2, 3])
+    elif mode == 4:
+        return img.rot90(2, [2, 3]).flip([2])
+    elif mode == 5:
+        return img.rot90(1, [2, 3])
+    elif mode == 6:
+        return img.rot90(2, [2, 3])
+    elif mode == 7:
+        return img.rot90(3, [2, 3]).flip([2])
+
+
+def augment_img_tensor(img, mode=0):
+    '''Kai Zhang (github: https://github.com/cszn)
+    '''
+    img_size = img.size()
+    img_np = img.data.cpu().numpy()
+    if len(img_size) == 3:
+        img_np = np.transpose(img_np, (1, 2, 0))
+    elif len(img_size) == 4:
+        img_np = np.transpose(img_np, (2, 3, 1, 0))
+    img_np = augment_img(img_np, mode=mode)
+    img_tensor = torch.from_numpy(np.ascontiguousarray(img_np))
+    if len(img_size) == 3:
+        img_tensor = img_tensor.permute(2, 0, 1)
+    elif len(img_size) == 4:
+        img_tensor = img_tensor.permute(3, 2, 0, 1)
+
+    return img_tensor.type_as(img)
+
+
+def augment_img_np3(img, mode=0):
+    if mode == 0:
+        return img
+    elif mode == 1:
+        return img.transpose(1, 0, 2)
+    elif mode == 2:
+        return img[::-1, :, :]
+    elif mode == 3:
+        img = img[::-1, :, :]
+        img = img.transpose(1, 0, 2)
+        return img
+    elif mode == 4:
+        return img[:, ::-1, :]
+    elif mode == 5:
+        img = img[:, ::-1, :]
+        img = img.transpose(1, 0, 2)
+        return img
+    elif mode == 6:
+        img = img[:, ::-1, :]
+        img = img[::-1, :, :]
+        return img
+    elif mode == 7:
+        img = img[:, ::-1, :]
+        img = img[::-1, :, :]
+        img = img.transpose(1, 0, 2)
+        return img
+
+
+def augment_imgs(img_list, hflip=True, rot=True):
+    # horizontal flip OR rotate
+    hflip = hflip and random.random() < 0.5
+    vflip = rot and random.random() < 0.5
+    rot90 = rot and random.random() < 0.5
+
+    def _augment(img):
+        if hflip:
+            img = img[:, ::-1, :]
+        if vflip:
+            img = img[::-1, :, :]
+        if rot90:
+            img = img.transpose(1, 0, 2)
+        return img
+
+    return [_augment(img) for img in img_list]
+
+
+'''
+# --------------------------------------------
+# modcrop and shave
+# --------------------------------------------
+'''
+
+
+def modcrop(img_in, scale):
+    # img_in: Numpy, HWC or HW
+    img = np.copy(img_in)
+    if img.ndim == 2:
+        H, W = img.shape
+        H_r, W_r = H % scale, W % scale
+        img = img[:H - H_r, :W - W_r]
+    elif img.ndim == 3:
+        H, W, C = img.shape
+        H_r, W_r = H % scale, W % scale
+        img = img[:H - H_r, :W - W_r, :]
+    else:
+        raise ValueError('Wrong img ndim: [{:d}].'.format(img.ndim))
+    return img
+
+
+def shave(img_in, border=0):
+    # img_in: Numpy, HWC or HW
+    img = np.copy(img_in)
+    h, w = img.shape[:2]
+    img = img[border:h-border, border:w-border]
+    return img
+
+
+'''
+# --------------------------------------------
+# image processing process on numpy image
+# channel_convert(in_c, tar_type, img_list):
+# rgb2ycbcr(img, only_y=True):
+# bgr2ycbcr(img, only_y=True):
+# ycbcr2rgb(img):
+# --------------------------------------------
+'''
+
+
+def rgb2ycbcr(img, only_y=True):
+    '''same as matlab rgb2ycbcr
+    only_y: only return Y channel
+    Input:
+        uint8, [0, 255]
+        float, [0, 1]
+    '''
+    in_img_type = img.dtype
+    img.astype(np.float32)
+    if in_img_type != np.uint8:
+        img *= 255.
+    # convert
+    if only_y:
+        rlt = np.dot(img, [65.481, 128.553, 24.966]) / 255.0 + 16.0
+    else:
+        rlt = np.matmul(img, [[65.481, -37.797, 112.0], [128.553, -74.203, -93.786],
+                              [24.966, 112.0, -18.214]]) / 255.0 + [16, 128, 128]
+    if in_img_type == np.uint8:
+        rlt = rlt.round()
+    else:
+        rlt /= 255.
+    return rlt.astype(in_img_type)
+
+
+def ycbcr2rgb(img):
+    '''same as matlab ycbcr2rgb
+    Input:
+        uint8, [0, 255]
+        float, [0, 1]
+    '''
+    in_img_type = img.dtype
+    img.astype(np.float32)
+    if in_img_type != np.uint8:
+        img *= 255.
+    # convert
+    rlt = np.matmul(img, [[0.00456621, 0.00456621, 0.00456621], [0, -0.00153632, 0.00791071],
+                          [0.00625893, -0.00318811, 0]]) * 255.0 + [-222.921, 135.576, -276.836]
+    if in_img_type == np.uint8:
+        rlt = rlt.round()
+    else:
+        rlt /= 255.
+    return rlt.astype(in_img_type)
+
+
+def bgr2ycbcr(img, only_y=True):
+    '''bgr version of rgb2ycbcr
+    only_y: only return Y channel
+    Input:
+        uint8, [0, 255]
+        float, [0, 1]
+    '''
+    in_img_type = img.dtype
+    img.astype(np.float32)
+    if in_img_type != np.uint8:
+        img *= 255.
+    # convert
+    if only_y:
+        rlt = np.dot(img, [24.966, 128.553, 65.481]) / 255.0 + 16.0
+    else:
+        rlt = np.matmul(img, [[24.966, 112.0, -18.214], [128.553, -74.203, -93.786],
+                              [65.481, -37.797, 112.0]]) / 255.0 + [16, 128, 128]
+    if in_img_type == np.uint8:
+        rlt = rlt.round()
+    else:
+        rlt /= 255.
+    return rlt.astype(in_img_type)
+
+
+def channel_convert(in_c, tar_type, img_list):
+    # conversion among BGR, gray and y
+    if in_c == 3 and tar_type == 'gray':  # BGR to gray
+        gray_list = [cv2.cvtColor(img, cv2.COLOR_BGR2GRAY) for img in img_list]
+        return [np.expand_dims(img, axis=2) for img in gray_list]
+    elif in_c == 3 and tar_type == 'y':  # BGR to y
+        y_list = [bgr2ycbcr(img, only_y=True) for img in img_list]
+        return [np.expand_dims(img, axis=2) for img in y_list]
+    elif in_c == 1 and tar_type == 'RGB':  # gray/y to BGR
+        return [cv2.cvtColor(img, cv2.COLOR_GRAY2BGR) for img in img_list]
+    else:
+        return img_list
+
+
+'''
+# --------------------------------------------
+# metric, PSNR and SSIM
+# --------------------------------------------
+'''
+
+
+# --------------------------------------------
+# PSNR
+# --------------------------------------------
+def calculate_psnr(img1, img2, border=0):
+    # img1 and img2 have range [0, 255]
+    #img1 = img1.squeeze()
+    #img2 = img2.squeeze()
+    if not img1.shape == img2.shape:
+        raise ValueError('Input images must have the same dimensions.')
+    h, w = img1.shape[:2]
+    img1 = img1[border:h-border, border:w-border]
+    img2 = img2[border:h-border, border:w-border]
+
+    img1 = img1.astype(np.float64)
+    img2 = img2.astype(np.float64)
+    mse = np.mean((img1 - img2)**2)
+    if mse == 0:
+        return float('inf')
+    return 20 * math.log10(255.0 / math.sqrt(mse))
+
+
+# --------------------------------------------
+# SSIM
+# --------------------------------------------
+def calculate_ssim(img1, img2, border=0):
+    '''calculate SSIM
+    the same outputs as MATLAB's
+    img1, img2: [0, 255]
+    '''
+    #img1 = img1.squeeze()
+    #img2 = img2.squeeze()
+    if not img1.shape == img2.shape:
+        raise ValueError('Input images must have the same dimensions.')
+    h, w = img1.shape[:2]
+    img1 = img1[border:h-border, border:w-border]
+    img2 = img2[border:h-border, border:w-border]
+
+    if img1.ndim == 2:
+        return ssim(img1, img2)
+    elif img1.ndim == 3:
+        if img1.shape[2] == 3:
+            ssims = []
+            for i in range(3):
+                ssims.append(ssim(img1[:,:,i], img2[:,:,i]))
+            return np.array(ssims).mean()
+        elif img1.shape[2] == 1:
+            return ssim(np.squeeze(img1), np.squeeze(img2))
+    else:
+        raise ValueError('Wrong input image dimensions.')
+
+
+def ssim(img1, img2):
+    C1 = (0.01 * 255)**2
+    C2 = (0.03 * 255)**2
+
+    img1 = img1.astype(np.float64)
+    img2 = img2.astype(np.float64)
+    kernel = cv2.getGaussianKernel(11, 1.5)
+    window = np.outer(kernel, kernel.transpose())
+
+    mu1 = cv2.filter2D(img1, -1, window)[5:-5, 5:-5]  # valid
+    mu2 = cv2.filter2D(img2, -1, window)[5:-5, 5:-5]
+    mu1_sq = mu1**2
+    mu2_sq = mu2**2
+    mu1_mu2 = mu1 * mu2
+    sigma1_sq = cv2.filter2D(img1**2, -1, window)[5:-5, 5:-5] - mu1_sq
+    sigma2_sq = cv2.filter2D(img2**2, -1, window)[5:-5, 5:-5] - mu2_sq
+    sigma12 = cv2.filter2D(img1 * img2, -1, window)[5:-5, 5:-5] - mu1_mu2
+
+    ssim_map = ((2 * mu1_mu2 + C1) * (2 * sigma12 + C2)) / ((mu1_sq + mu2_sq + C1) *
+                                                            (sigma1_sq + sigma2_sq + C2))
+    return ssim_map.mean()
+
+
+'''
+# --------------------------------------------
+# matlab's bicubic imresize (numpy and torch) [0, 1]
+# --------------------------------------------
+'''
+
+
+# matlab 'imresize' function, now only support 'bicubic'
+def cubic(x):
+    absx = torch.abs(x)
+    absx2 = absx**2
+    absx3 = absx**3
+    return (1.5*absx3 - 2.5*absx2 + 1) * ((absx <= 1).type_as(absx)) + \
+        (-0.5*absx3 + 2.5*absx2 - 4*absx + 2) * (((absx > 1)*(absx <= 2)).type_as(absx))
+
+
+def calculate_weights_indices(in_length, out_length, scale, kernel, kernel_width, antialiasing):
+    if (scale < 1) and (antialiasing):
+        # Use a modified kernel to simultaneously interpolate and antialias- larger kernel width
+        kernel_width = kernel_width / scale
+
+    # Output-space coordinates
+    x = torch.linspace(1, out_length, out_length)
+
+    # Input-space coordinates. Calculate the inverse mapping such that 0.5
+    # in output space maps to 0.5 in input space, and 0.5+scale in output
+    # space maps to 1.5 in input space.
+    u = x / scale + 0.5 * (1 - 1 / scale)
+
+    # What is the left-most pixel that can be involved in the computation?
+    left = torch.floor(u - kernel_width / 2)
+
+    # What is the maximum number of pixels that can be involved in the
+    # computation?  Note: it's OK to use an extra pixel here; if the
+    # corresponding weights are all zero, it will be eliminated at the end
+    # of this function.
+    P = math.ceil(kernel_width) + 2
+
+    # The indices of the input pixels involved in computing the k-th output
+    # pixel are in row k of the indices matrix.
+    indices = left.view(out_length, 1).expand(out_length, P) + torch.linspace(0, P - 1, P).view(
+        1, P).expand(out_length, P)
+
+    # The weights used to compute the k-th output pixel are in row k of the
+    # weights matrix.
+    distance_to_center = u.view(out_length, 1).expand(out_length, P) - indices
+    # apply cubic kernel
+    if (scale < 1) and (antialiasing):
+        weights = scale * cubic(distance_to_center * scale)
+    else:
+        weights = cubic(distance_to_center)
+    # Normalize the weights matrix so that each row sums to 1.
+    weights_sum = torch.sum(weights, 1).view(out_length, 1)
+    weights = weights / weights_sum.expand(out_length, P)
+
+    # If a column in weights is all zero, get rid of it. only consider the first and last column.
+    weights_zero_tmp = torch.sum((weights == 0), 0)
+    if not math.isclose(weights_zero_tmp[0], 0, rel_tol=1e-6):
+        indices = indices.narrow(1, 1, P - 2)
+        weights = weights.narrow(1, 1, P - 2)
+    if not math.isclose(weights_zero_tmp[-1], 0, rel_tol=1e-6):
+        indices = indices.narrow(1, 0, P - 2)
+        weights = weights.narrow(1, 0, P - 2)
+    weights = weights.contiguous()
+    indices = indices.contiguous()
+    sym_len_s = -indices.min() + 1
+    sym_len_e = indices.max() - in_length
+    indices = indices + sym_len_s - 1
+    return weights, indices, int(sym_len_s), int(sym_len_e)
+
+
+# --------------------------------------------
+# imresize for tensor image [0, 1]
+# --------------------------------------------
+def imresize(img, scale, antialiasing=True):
+    # Now the scale should be the same for H and W
+    # input: img: pytorch tensor, CHW or HW [0,1]
+    # output: CHW or HW [0,1] w/o round
+    need_squeeze = True if img.dim() == 2 else False
+    if need_squeeze:
+        img.unsqueeze_(0)
+    in_C, in_H, in_W = img.size()
+    out_C, out_H, out_W = in_C, math.ceil(in_H * scale), math.ceil(in_W * scale)
+    kernel_width = 4
+    kernel = 'cubic'
+
+    # Return the desired dimension order for performing the resize.  The
+    # strategy is to perform the resize first along the dimension with the
+    # smallest scale factor.
+    # Now we do not support this.
+
+    # get weights and indices
+    weights_H, indices_H, sym_len_Hs, sym_len_He = calculate_weights_indices(
+        in_H, out_H, scale, kernel, kernel_width, antialiasing)
+    weights_W, indices_W, sym_len_Ws, sym_len_We = calculate_weights_indices(
+        in_W, out_W, scale, kernel, kernel_width, antialiasing)
+    # process H dimension
+    # symmetric copying
+    img_aug = torch.FloatTensor(in_C, in_H + sym_len_Hs + sym_len_He, in_W)
+    img_aug.narrow(1, sym_len_Hs, in_H).copy_(img)
+
+    sym_patch = img[:, :sym_len_Hs, :]
+    inv_idx = torch.arange(sym_patch.size(1) - 1, -1, -1).long()
+    sym_patch_inv = sym_patch.index_select(1, inv_idx)
+    img_aug.narrow(1, 0, sym_len_Hs).copy_(sym_patch_inv)
+
+    sym_patch = img[:, -sym_len_He:, :]
+    inv_idx = torch.arange(sym_patch.size(1) - 1, -1, -1).long()
+    sym_patch_inv = sym_patch.index_select(1, inv_idx)
+    img_aug.narrow(1, sym_len_Hs + in_H, sym_len_He).copy_(sym_patch_inv)
+
+    out_1 = torch.FloatTensor(in_C, out_H, in_W)
+    kernel_width = weights_H.size(1)
+    for i in range(out_H):
+        idx = int(indices_H[i][0])
+        for j in range(out_C):
+            out_1[j, i, :] = img_aug[j, idx:idx + kernel_width, :].transpose(0, 1).mv(weights_H[i])
+
+    # process W dimension
+    # symmetric copying
+    out_1_aug = torch.FloatTensor(in_C, out_H, in_W + sym_len_Ws + sym_len_We)
+    out_1_aug.narrow(2, sym_len_Ws, in_W).copy_(out_1)
+
+    sym_patch = out_1[:, :, :sym_len_Ws]
+    inv_idx = torch.arange(sym_patch.size(2) - 1, -1, -1).long()
+    sym_patch_inv = sym_patch.index_select(2, inv_idx)
+    out_1_aug.narrow(2, 0, sym_len_Ws).copy_(sym_patch_inv)
+
+    sym_patch = out_1[:, :, -sym_len_We:]
+    inv_idx = torch.arange(sym_patch.size(2) - 1, -1, -1).long()
+    sym_patch_inv = sym_patch.index_select(2, inv_idx)
+    out_1_aug.narrow(2, sym_len_Ws + in_W, sym_len_We).copy_(sym_patch_inv)
+
+    out_2 = torch.FloatTensor(in_C, out_H, out_W)
+    kernel_width = weights_W.size(1)
+    for i in range(out_W):
+        idx = int(indices_W[i][0])
+        for j in range(out_C):
+            out_2[j, :, i] = out_1_aug[j, :, idx:idx + kernel_width].mv(weights_W[i])
+    if need_squeeze:
+        out_2.squeeze_()
+    return out_2
+
+
+# --------------------------------------------
+# imresize for numpy image [0, 1]
+# --------------------------------------------
+def imresize_np(img, scale, antialiasing=True):
+    # Now the scale should be the same for H and W
+    # input: img: Numpy, HWC or HW [0,1]
+    # output: HWC or HW [0,1] w/o round
+    img = torch.from_numpy(img)
+    need_squeeze = True if img.dim() == 2 else False
+    if need_squeeze:
+        img.unsqueeze_(2)
+
+    in_H, in_W, in_C = img.size()
+    out_C, out_H, out_W = in_C, math.ceil(in_H * scale), math.ceil(in_W * scale)
+    kernel_width = 4
+    kernel = 'cubic'
+
+    # Return the desired dimension order for performing the resize.  The
+    # strategy is to perform the resize first along the dimension with the
+    # smallest scale factor.
+    # Now we do not support this.
+
+    # get weights and indices
+    weights_H, indices_H, sym_len_Hs, sym_len_He = calculate_weights_indices(
+        in_H, out_H, scale, kernel, kernel_width, antialiasing)
+    weights_W, indices_W, sym_len_Ws, sym_len_We = calculate_weights_indices(
+        in_W, out_W, scale, kernel, kernel_width, antialiasing)
+    # process H dimension
+    # symmetric copying
+    img_aug = torch.FloatTensor(in_H + sym_len_Hs + sym_len_He, in_W, in_C)
+    img_aug.narrow(0, sym_len_Hs, in_H).copy_(img)
+
+    sym_patch = img[:sym_len_Hs, :, :]
+    inv_idx = torch.arange(sym_patch.size(0) - 1, -1, -1).long()
+    sym_patch_inv = sym_patch.index_select(0, inv_idx)
+    img_aug.narrow(0, 0, sym_len_Hs).copy_(sym_patch_inv)
+
+    sym_patch = img[-sym_len_He:, :, :]
+    inv_idx = torch.arange(sym_patch.size(0) - 1, -1, -1).long()
+    sym_patch_inv = sym_patch.index_select(0, inv_idx)
+    img_aug.narrow(0, sym_len_Hs + in_H, sym_len_He).copy_(sym_patch_inv)
+
+    out_1 = torch.FloatTensor(out_H, in_W, in_C)
+    kernel_width = weights_H.size(1)
+    for i in range(out_H):
+        idx = int(indices_H[i][0])
+        for j in range(out_C):
+            out_1[i, :, j] = img_aug[idx:idx + kernel_width, :, j].transpose(0, 1).mv(weights_H[i])
+
+    # process W dimension
+    # symmetric copying
+    out_1_aug = torch.FloatTensor(out_H, in_W + sym_len_Ws + sym_len_We, in_C)
+    out_1_aug.narrow(1, sym_len_Ws, in_W).copy_(out_1)
+
+    sym_patch = out_1[:, :sym_len_Ws, :]
+    inv_idx = torch.arange(sym_patch.size(1) - 1, -1, -1).long()
+    sym_patch_inv = sym_patch.index_select(1, inv_idx)
+    out_1_aug.narrow(1, 0, sym_len_Ws).copy_(sym_patch_inv)
+
+    sym_patch = out_1[:, -sym_len_We:, :]
+    inv_idx = torch.arange(sym_patch.size(1) - 1, -1, -1).long()
+    sym_patch_inv = sym_patch.index_select(1, inv_idx)
+    out_1_aug.narrow(1, sym_len_Ws + in_W, sym_len_We).copy_(sym_patch_inv)
+
+    out_2 = torch.FloatTensor(out_H, out_W, in_C)
+    kernel_width = weights_W.size(1)
+    for i in range(out_W):
+        idx = int(indices_W[i][0])
+        for j in range(out_C):
+            out_2[:, i, j] = out_1_aug[:, idx:idx + kernel_width, j].mv(weights_W[i])
+    if need_squeeze:
+        out_2.squeeze_()
+
+    return out_2.numpy()
+
+
+if __name__ == '__main__':
+    print('---')
+#    img = imread_uint('test.bmp', 3)
+#    img = uint2single(img)
+#    img_bicubic = imresize_np(img, 1/4)

ldm/modules/losses/__init__.py

@@ -0,0 +1 @@
+from ldm.modules.losses.contperceptual import LPIPSWithDiscriminator

ldm/modules/losses/contperceptual.py

@@ -0,0 +1,111 @@
+import torch
+import torch.nn as nn
+
+from taming.modules.losses.vqperceptual import *  # TODO: taming dependency yes/no?
+
+
+class LPIPSWithDiscriminator(nn.Module):
+    def __init__(self, disc_start, logvar_init=0.0, kl_weight=1.0, pixelloss_weight=1.0,
+                 disc_num_layers=3, disc_in_channels=3, disc_factor=1.0, disc_weight=1.0,
+                 perceptual_weight=1.0, use_actnorm=False, disc_conditional=False,
+                 disc_loss="hinge"):
+
+        super().__init__()
+        assert disc_loss in ["hinge", "vanilla"]
+        self.kl_weight = kl_weight
+        self.pixel_weight = pixelloss_weight
+        self.perceptual_loss = LPIPS().eval()
+        self.perceptual_weight = perceptual_weight
+        # output log variance
+        self.logvar = nn.Parameter(torch.ones(size=()) * logvar_init)
+
+        self.discriminator = NLayerDiscriminator(input_nc=disc_in_channels,
+                                                 n_layers=disc_num_layers,
+                                                 use_actnorm=use_actnorm
+                                                 ).apply(weights_init)
+        self.discriminator_iter_start = disc_start
+        self.disc_loss = hinge_d_loss if disc_loss == "hinge" else vanilla_d_loss
+        self.disc_factor = disc_factor
+        self.discriminator_weight = disc_weight
+        self.disc_conditional = disc_conditional
+
+    def calculate_adaptive_weight(self, nll_loss, g_loss, last_layer=None):
+        if last_layer is not None:
+            nll_grads = torch.autograd.grad(nll_loss, last_layer, retain_graph=True)[0]
+            g_grads = torch.autograd.grad(g_loss, last_layer, retain_graph=True)[0]
+        else:
+            nll_grads = torch.autograd.grad(nll_loss, self.last_layer[0], retain_graph=True)[0]
+            g_grads = torch.autograd.grad(g_loss, self.last_layer[0], retain_graph=True)[0]
+
+        d_weight = torch.norm(nll_grads) / (torch.norm(g_grads) + 1e-4)
+        d_weight = torch.clamp(d_weight, 0.0, 1e4).detach()
+        d_weight = d_weight * self.discriminator_weight
+        return d_weight
+
+    def forward(self, inputs, reconstructions, posteriors, optimizer_idx,
+                global_step, last_layer=None, cond=None, split="train",
+                weights=None):
+        rec_loss = torch.abs(inputs.contiguous() - reconstructions.contiguous())
+        if self.perceptual_weight > 0:
+            p_loss = self.perceptual_loss(inputs.contiguous(), reconstructions.contiguous())
+            rec_loss = rec_loss + self.perceptual_weight * p_loss
+
+        nll_loss = rec_loss / torch.exp(self.logvar) + self.logvar
+        weighted_nll_loss = nll_loss
+        if weights is not None:
+            weighted_nll_loss = weights*nll_loss
+        weighted_nll_loss = torch.sum(weighted_nll_loss) / weighted_nll_loss.shape[0]
+        nll_loss = torch.sum(nll_loss) / nll_loss.shape[0]
+        kl_loss = posteriors.kl()
+        kl_loss = torch.sum(kl_loss) / kl_loss.shape[0]
+
+        # now the GAN part
+        if optimizer_idx == 0:
+            # generator update
+            if cond is None:
+                assert not self.disc_conditional
+                logits_fake = self.discriminator(reconstructions.contiguous())
+            else:
+                assert self.disc_conditional
+                logits_fake = self.discriminator(torch.cat((reconstructions.contiguous(), cond), dim=1))
+            g_loss = -torch.mean(logits_fake)
+
+            if self.disc_factor > 0.0:
+                try:
+                    d_weight = self.calculate_adaptive_weight(nll_loss, g_loss, last_layer=last_layer)
+                except RuntimeError:
+                    assert not self.training
+                    d_weight = torch.tensor(0.0)
+            else:
+                d_weight = torch.tensor(0.0)
+
+            disc_factor = adopt_weight(self.disc_factor, global_step, threshold=self.discriminator_iter_start)
+            loss = weighted_nll_loss + self.kl_weight * kl_loss + d_weight * disc_factor * g_loss
+
+            log = {"{}/total_loss".format(split): loss.clone().detach().mean(), "{}/logvar".format(split): self.logvar.detach(),
+                   "{}/kl_loss".format(split): kl_loss.detach().mean(), "{}/nll_loss".format(split): nll_loss.detach().mean(),
+                   "{}/rec_loss".format(split): rec_loss.detach().mean(),
+                   "{}/d_weight".format(split): d_weight.detach(),
+                   "{}/disc_factor".format(split): torch.tensor(disc_factor),
+                   "{}/g_loss".format(split): g_loss.detach().mean(),
+                   }
+            return loss, log
+
+        if optimizer_idx == 1:
+            # second pass for discriminator update
+            if cond is None:
+                logits_real = self.discriminator(inputs.contiguous().detach())
+                logits_fake = self.discriminator(reconstructions.contiguous().detach())
+            else:
+                logits_real = self.discriminator(torch.cat((inputs.contiguous().detach(), cond), dim=1))
+                logits_fake = self.discriminator(torch.cat((reconstructions.contiguous().detach(), cond), dim=1))
+
+            disc_factor = adopt_weight(self.disc_factor, global_step, threshold=self.discriminator_iter_start)
+            d_loss = disc_factor * self.disc_loss(logits_real, logits_fake)
+
+            log = {"{}/disc_loss".format(split): d_loss.clone().detach().mean(),
+                   "{}/logits_real".format(split): logits_real.detach().mean(),
+                   "{}/logits_fake".format(split): logits_fake.detach().mean()
+                   }
+            return d_loss, log
+

ldm/modules/losses/vqperceptual.py

@@ -0,0 +1,167 @@
+import torch
+from torch import nn
+import torch.nn.functional as F
+from einops import repeat
+
+from taming.modules.discriminator.model import NLayerDiscriminator, weights_init
+from taming.modules.losses.lpips import LPIPS
+from taming.modules.losses.vqperceptual import hinge_d_loss, vanilla_d_loss
+
+
+def hinge_d_loss_with_exemplar_weights(logits_real, logits_fake, weights):
+    assert weights.shape[0] == logits_real.shape[0] == logits_fake.shape[0]
+    loss_real = torch.mean(F.relu(1. - logits_real), dim=[1,2,3])
+    loss_fake = torch.mean(F.relu(1. + logits_fake), dim=[1,2,3])
+    loss_real = (weights * loss_real).sum() / weights.sum()
+    loss_fake = (weights * loss_fake).sum() / weights.sum()
+    d_loss = 0.5 * (loss_real + loss_fake)
+    return d_loss
+
+def adopt_weight(weight, global_step, threshold=0, value=0.):
+    if global_step < threshold:
+        weight = value
+    return weight
+
+
+def measure_perplexity(predicted_indices, n_embed):
+    # src: https://github.com/karpathy/deep-vector-quantization/blob/main/model.py
+    # eval cluster perplexity. when perplexity == num_embeddings then all clusters are used exactly equally
+    encodings = F.one_hot(predicted_indices, n_embed).float().reshape(-1, n_embed)
+    avg_probs = encodings.mean(0)
+    perplexity = (-(avg_probs * torch.log(avg_probs + 1e-10)).sum()).exp()
+    cluster_use = torch.sum(avg_probs > 0)
+    return perplexity, cluster_use
+
+def l1(x, y):
+    return torch.abs(x-y)
+
+
+def l2(x, y):
+    return torch.pow((x-y), 2)
+
+
+class VQLPIPSWithDiscriminator(nn.Module):
+    def __init__(self, disc_start, codebook_weight=1.0, pixelloss_weight=1.0,
+                 disc_num_layers=3, disc_in_channels=3, disc_factor=1.0, disc_weight=1.0,
+                 perceptual_weight=1.0, use_actnorm=False, disc_conditional=False,
+                 disc_ndf=64, disc_loss="hinge", n_classes=None, perceptual_loss="lpips",
+                 pixel_loss="l1"):
+        super().__init__()
+        assert disc_loss in ["hinge", "vanilla"]
+        assert perceptual_loss in ["lpips", "clips", "dists"]
+        assert pixel_loss in ["l1", "l2"]
+        self.codebook_weight = codebook_weight
+        self.pixel_weight = pixelloss_weight
+        if perceptual_loss == "lpips":
+            print(f"{self.__class__.__name__}: Running with LPIPS.")
+            self.perceptual_loss = LPIPS().eval()
+        else:
+            raise ValueError(f"Unknown perceptual loss: >> {perceptual_loss} <<")
+        self.perceptual_weight = perceptual_weight
+
+        if pixel_loss == "l1":
+            self.pixel_loss = l1
+        else:
+            self.pixel_loss = l2
+
+        self.discriminator = NLayerDiscriminator(input_nc=disc_in_channels,
+                                                 n_layers=disc_num_layers,
+                                                 use_actnorm=use_actnorm,
+                                                 ndf=disc_ndf
+                                                 ).apply(weights_init)
+        self.discriminator_iter_start = disc_start
+        if disc_loss == "hinge":
+            self.disc_loss = hinge_d_loss
+        elif disc_loss == "vanilla":
+            self.disc_loss = vanilla_d_loss
+        else:
+            raise ValueError(f"Unknown GAN loss '{disc_loss}'.")
+        print(f"VQLPIPSWithDiscriminator running with {disc_loss} loss.")
+        self.disc_factor = disc_factor
+        self.discriminator_weight = disc_weight
+        self.disc_conditional = disc_conditional
+        self.n_classes = n_classes
+
+    def calculate_adaptive_weight(self, nll_loss, g_loss, last_layer=None):
+        if last_layer is not None:
+            nll_grads = torch.autograd.grad(nll_loss, last_layer, retain_graph=True)[0]
+            g_grads = torch.autograd.grad(g_loss, last_layer, retain_graph=True)[0]
+        else:
+            nll_grads = torch.autograd.grad(nll_loss, self.last_layer[0], retain_graph=True)[0]
+            g_grads = torch.autograd.grad(g_loss, self.last_layer[0], retain_graph=True)[0]
+
+        d_weight = torch.norm(nll_grads) / (torch.norm(g_grads) + 1e-4)
+        d_weight = torch.clamp(d_weight, 0.0, 1e4).detach()
+        d_weight = d_weight * self.discriminator_weight
+        return d_weight
+
+    def forward(self, codebook_loss, inputs, reconstructions, optimizer_idx,
+                global_step, last_layer=None, cond=None, split="train", predicted_indices=None):
+        if not exists(codebook_loss):
+            codebook_loss = torch.tensor([0.]).to(inputs.device)
+        #rec_loss = torch.abs(inputs.contiguous() - reconstructions.contiguous())
+        rec_loss = self.pixel_loss(inputs.contiguous(), reconstructions.contiguous())
+        if self.perceptual_weight > 0:
+            p_loss = self.perceptual_loss(inputs.contiguous(), reconstructions.contiguous())
+            rec_loss = rec_loss + self.perceptual_weight * p_loss
+        else:
+            p_loss = torch.tensor([0.0])
+
+        nll_loss = rec_loss
+        #nll_loss = torch.sum(nll_loss) / nll_loss.shape[0]
+        nll_loss = torch.mean(nll_loss)
+
+        # now the GAN part
+        if optimizer_idx == 0:
+            # generator update
+            if cond is None:
+                assert not self.disc_conditional
+                logits_fake = self.discriminator(reconstructions.contiguous())
+            else:
+                assert self.disc_conditional
+                logits_fake = self.discriminator(torch.cat((reconstructions.contiguous(), cond), dim=1))
+            g_loss = -torch.mean(logits_fake)
+
+            try:
+                d_weight = self.calculate_adaptive_weight(nll_loss, g_loss, last_layer=last_layer)
+            except RuntimeError:
+                assert not self.training
+                d_weight = torch.tensor(0.0)
+
+            disc_factor = adopt_weight(self.disc_factor, global_step, threshold=self.discriminator_iter_start)
+            loss = nll_loss + d_weight * disc_factor * g_loss + self.codebook_weight * codebook_loss.mean()
+
+            log = {"{}/total_loss".format(split): loss.clone().detach().mean(),
+                   "{}/quant_loss".format(split): codebook_loss.detach().mean(),
+                   "{}/nll_loss".format(split): nll_loss.detach().mean(),
+                   "{}/rec_loss".format(split): rec_loss.detach().mean(),
+                   "{}/p_loss".format(split): p_loss.detach().mean(),
+                   "{}/d_weight".format(split): d_weight.detach(),
+                   "{}/disc_factor".format(split): torch.tensor(disc_factor),
+                   "{}/g_loss".format(split): g_loss.detach().mean(),
+                   }
+            if predicted_indices is not None:
+                assert self.n_classes is not None
+                with torch.no_grad():
+                    perplexity, cluster_usage = measure_perplexity(predicted_indices, self.n_classes)
+                log[f"{split}/perplexity"] = perplexity
+                log[f"{split}/cluster_usage"] = cluster_usage
+            return loss, log
+
+        if optimizer_idx == 1:
+            # second pass for discriminator update
+            if cond is None:
+                logits_real = self.discriminator(inputs.contiguous().detach())
+                logits_fake = self.discriminator(reconstructions.contiguous().detach())
+            else:
+                logits_real = self.discriminator(torch.cat((inputs.contiguous().detach(), cond), dim=1))
+                logits_fake = self.discriminator(torch.cat((reconstructions.contiguous().detach(), cond), dim=1))
+
+            disc_factor = adopt_weight(self.disc_factor, global_step, threshold=self.discriminator_iter_start)
+            d_loss = disc_factor * self.disc_loss(logits_real, logits_fake)
+
+            log = {"{}/disc_loss".format(split): d_loss.clone().detach().mean(),
+                   "{}/logits_real".format(split): logits_real.detach().mean(),
+                   "{}/logits_fake".format(split): logits_fake.detach().mean()
+                   }
+            return d_loss, log