Upload 560 files

lavis/__init__.py

@@ -0,0 +1,31 @@
+"""
+ Copyright (c) 2022, salesforce.com, inc.
+ All rights reserved.
+ SPDX-License-Identifier: BSD-3-Clause
+ For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause
+"""
+
+import os
+import sys
+
+from omegaconf import OmegaConf
+
+from lavis.common.registry import registry
+
+from lavis.datasets.builders import *
+from lavis.models import *
+from lavis.processors import *
+from lavis.tasks import *
+
+
+root_dir = os.path.dirname(os.path.abspath(__file__))
+default_cfg = OmegaConf.load(os.path.join(root_dir, "configs/default.yaml"))
+
+registry.register_path("library_root", root_dir)
+repo_root = os.path.join(root_dir, "..")
+registry.register_path("repo_root", repo_root)
+cache_root = os.path.join(repo_root, default_cfg.env.cache_root)
+registry.register_path("cache_root", cache_root)
+
+registry.register("MAX_INT", sys.maxsize)
+registry.register("SPLIT_NAMES", ["train", "val", "test"])

lavis/common/annotator/canny/__init__.py

@@ -0,0 +1,6 @@
+import cv2
+
+
+class CannyDetector:
+    def __call__(self, img, low_threshold, high_threshold):
+        return cv2.Canny(img, low_threshold, high_threshold)

lavis/common/annotator/ckpts/download.sh

@@ -0,0 +1,5 @@
+#! /bin/bash
+
+wget https://huggingface.co/lllyasviel/ControlNet/resolve/main/annotator/ckpts/dpt_hybrid-midas-501f0c75.pt
+wget https://huggingface.co/lllyasviel/ControlNet/resolve/main/annotator/ckpts/network-bsds500.pth
+

lavis/common/annotator/hed/__init__.py

@@ -0,0 +1,132 @@
+import numpy as np
+import cv2
+import os
+import torch
+from einops import rearrange
+from annotator.util import annotator_ckpts_path
+
+
+class Network(torch.nn.Module):
+    def __init__(self, model_path):
+        super().__init__()
+
+        self.netVggOne = torch.nn.Sequential(
+            torch.nn.Conv2d(in_channels=3, out_channels=64, kernel_size=3, stride=1, padding=1),
+            torch.nn.ReLU(inplace=False),
+            torch.nn.Conv2d(in_channels=64, out_channels=64, kernel_size=3, stride=1, padding=1),
+            torch.nn.ReLU(inplace=False)
+        )
+
+        self.netVggTwo = torch.nn.Sequential(
+            torch.nn.MaxPool2d(kernel_size=2, stride=2),
+            torch.nn.Conv2d(in_channels=64, out_channels=128, kernel_size=3, stride=1, padding=1),
+            torch.nn.ReLU(inplace=False),
+            torch.nn.Conv2d(in_channels=128, out_channels=128, kernel_size=3, stride=1, padding=1),
+            torch.nn.ReLU(inplace=False)
+        )
+
+        self.netVggThr = torch.nn.Sequential(
+            torch.nn.MaxPool2d(kernel_size=2, stride=2),
+            torch.nn.Conv2d(in_channels=128, out_channels=256, kernel_size=3, stride=1, padding=1),
+            torch.nn.ReLU(inplace=False),
+            torch.nn.Conv2d(in_channels=256, out_channels=256, kernel_size=3, stride=1, padding=1),
+            torch.nn.ReLU(inplace=False),
+            torch.nn.Conv2d(in_channels=256, out_channels=256, kernel_size=3, stride=1, padding=1),
+            torch.nn.ReLU(inplace=False)
+        )
+
+        self.netVggFou = torch.nn.Sequential(
+            torch.nn.MaxPool2d(kernel_size=2, stride=2),
+            torch.nn.Conv2d(in_channels=256, out_channels=512, kernel_size=3, stride=1, padding=1),
+            torch.nn.ReLU(inplace=False),
+            torch.nn.Conv2d(in_channels=512, out_channels=512, kernel_size=3, stride=1, padding=1),
+            torch.nn.ReLU(inplace=False),
+            torch.nn.Conv2d(in_channels=512, out_channels=512, kernel_size=3, stride=1, padding=1),
+            torch.nn.ReLU(inplace=False)
+        )
+
+        self.netVggFiv = torch.nn.Sequential(
+            torch.nn.MaxPool2d(kernel_size=2, stride=2),
+            torch.nn.Conv2d(in_channels=512, out_channels=512, kernel_size=3, stride=1, padding=1),
+            torch.nn.ReLU(inplace=False),
+            torch.nn.Conv2d(in_channels=512, out_channels=512, kernel_size=3, stride=1, padding=1),
+            torch.nn.ReLU(inplace=False),
+            torch.nn.Conv2d(in_channels=512, out_channels=512, kernel_size=3, stride=1, padding=1),
+            torch.nn.ReLU(inplace=False)
+        )
+
+        self.netScoreOne = torch.nn.Conv2d(in_channels=64, out_channels=1, kernel_size=1, stride=1, padding=0)
+        self.netScoreTwo = torch.nn.Conv2d(in_channels=128, out_channels=1, kernel_size=1, stride=1, padding=0)
+        self.netScoreThr = torch.nn.Conv2d(in_channels=256, out_channels=1, kernel_size=1, stride=1, padding=0)
+        self.netScoreFou = torch.nn.Conv2d(in_channels=512, out_channels=1, kernel_size=1, stride=1, padding=0)
+        self.netScoreFiv = torch.nn.Conv2d(in_channels=512, out_channels=1, kernel_size=1, stride=1, padding=0)
+
+        self.netCombine = torch.nn.Sequential(
+            torch.nn.Conv2d(in_channels=5, out_channels=1, kernel_size=1, stride=1, padding=0),
+            torch.nn.Sigmoid()
+        )
+
+        self.load_state_dict({strKey.replace('module', 'net'): tenWeight for strKey, tenWeight in torch.load(model_path).items()})
+
+    def forward(self, tenInput):
+        tenInput = tenInput * 255.0
+        tenInput = tenInput - torch.tensor(data=[104.00698793, 116.66876762, 122.67891434], dtype=tenInput.dtype, device=tenInput.device).view(1, 3, 1, 1)
+
+        tenVggOne = self.netVggOne(tenInput)
+        tenVggTwo = self.netVggTwo(tenVggOne)
+        tenVggThr = self.netVggThr(tenVggTwo)
+        tenVggFou = self.netVggFou(tenVggThr)
+        tenVggFiv = self.netVggFiv(tenVggFou)
+
+        tenScoreOne = self.netScoreOne(tenVggOne)
+        tenScoreTwo = self.netScoreTwo(tenVggTwo)
+        tenScoreThr = self.netScoreThr(tenVggThr)
+        tenScoreFou = self.netScoreFou(tenVggFou)
+        tenScoreFiv = self.netScoreFiv(tenVggFiv)
+
+        tenScoreOne = torch.nn.functional.interpolate(input=tenScoreOne, size=(tenInput.shape[2], tenInput.shape[3]), mode='bilinear', align_corners=False)
+        tenScoreTwo = torch.nn.functional.interpolate(input=tenScoreTwo, size=(tenInput.shape[2], tenInput.shape[3]), mode='bilinear', align_corners=False)
+        tenScoreThr = torch.nn.functional.interpolate(input=tenScoreThr, size=(tenInput.shape[2], tenInput.shape[3]), mode='bilinear', align_corners=False)
+        tenScoreFou = torch.nn.functional.interpolate(input=tenScoreFou, size=(tenInput.shape[2], tenInput.shape[3]), mode='bilinear', align_corners=False)
+        tenScoreFiv = torch.nn.functional.interpolate(input=tenScoreFiv, size=(tenInput.shape[2], tenInput.shape[3]), mode='bilinear', align_corners=False)
+
+        return self.netCombine(torch.cat([ tenScoreOne, tenScoreTwo, tenScoreThr, tenScoreFou, tenScoreFiv ], 1))
+
+
+class HEDdetector:
+    def __init__(self):
+        remote_model_path = "https://huggingface.co/lllyasviel/ControlNet/resolve/main/annotator/ckpts/network-bsds500.pth"
+        modelpath = os.path.join(annotator_ckpts_path, "network-bsds500.pth")
+        if not os.path.exists(modelpath):
+            from basicsr.utils.download_util import load_file_from_url
+            load_file_from_url(remote_model_path, model_dir=annotator_ckpts_path)
+        self.netNetwork = Network(modelpath).cuda().eval()
+
+    def __call__(self, input_image):
+        assert input_image.ndim == 3
+        input_image = input_image[:, :, ::-1].copy()
+        with torch.no_grad():
+            image_hed = torch.from_numpy(input_image).float().cuda()
+            image_hed = image_hed / 255.0
+            image_hed = rearrange(image_hed, 'h w c -> 1 c h w')
+            edge = self.netNetwork(image_hed)[0]
+            edge = (edge.cpu().numpy() * 255.0).clip(0, 255).astype(np.uint8)
+            return edge[0]
+
+
+def nms(x, t, s):
+    x = cv2.GaussianBlur(x.astype(np.float32), (0, 0), s)
+
+    f1 = np.array([[0, 0, 0], [1, 1, 1], [0, 0, 0]], dtype=np.uint8)
+    f2 = np.array([[0, 1, 0], [0, 1, 0], [0, 1, 0]], dtype=np.uint8)
+    f3 = np.array([[1, 0, 0], [0, 1, 0], [0, 0, 1]], dtype=np.uint8)
+    f4 = np.array([[0, 0, 1], [0, 1, 0], [1, 0, 0]], dtype=np.uint8)
+
+    y = np.zeros_like(x)
+
+    for f in [f1, f2, f3, f4]:
+        np.putmask(y, cv2.dilate(x, kernel=f) == x, x)
+
+    z = np.zeros_like(y, dtype=np.uint8)
+    z[y > t] = 255
+    return z

lavis/common/annotator/midas/__init__.py

@@ -0,0 +1,38 @@
+import cv2
+import numpy as np
+import torch
+
+from einops import rearrange
+from .api import MiDaSInference
+
+
+class MidasDetector:
+    def __init__(self):
+        self.model = MiDaSInference(model_type="dpt_hybrid").cuda()
+
+    def __call__(self, input_image, a=np.pi * 2.0, bg_th=0.1):
+        assert input_image.ndim == 3
+        image_depth = input_image
+        with torch.no_grad():
+            image_depth = torch.from_numpy(image_depth).float().cuda()
+            image_depth = image_depth / 127.5 - 1.0
+            image_depth = rearrange(image_depth, 'h w c -> 1 c h w')
+            depth = self.model(image_depth)[0]
+
+            depth_pt = depth.clone()
+            depth_pt -= torch.min(depth_pt)
+            depth_pt /= torch.max(depth_pt)
+            depth_pt = depth_pt.cpu().numpy()
+            depth_image = (depth_pt * 255.0).clip(0, 255).astype(np.uint8)
+
+            depth_np = depth.cpu().numpy()
+            x = cv2.Sobel(depth_np, cv2.CV_32F, 1, 0, ksize=3)
+            y = cv2.Sobel(depth_np, cv2.CV_32F, 0, 1, ksize=3)
+            z = np.ones_like(x) * a
+            x[depth_pt < bg_th] = 0
+            y[depth_pt < bg_th] = 0
+            normal = np.stack([x, y, z], axis=2)
+            normal /= np.sum(normal ** 2.0, axis=2, keepdims=True) ** 0.5
+            normal_image = (normal * 127.5 + 127.5).clip(0, 255).astype(np.uint8)
+
+            return depth_image, normal_image

lavis/common/annotator/midas/api.py

@@ -0,0 +1,169 @@
+# based on https://github.com/isl-org/MiDaS
+
+import cv2
+import os
+import torch
+import torch.nn as nn
+from torchvision.transforms import Compose
+
+from .midas.dpt_depth import DPTDepthModel
+from .midas.midas_net import MidasNet
+from .midas.midas_net_custom import MidasNet_small
+from .midas.transforms import Resize, NormalizeImage, PrepareForNet
+from annotator.util import annotator_ckpts_path
+
+
+ISL_PATHS = {
+    "dpt_large": os.path.join(annotator_ckpts_path, "dpt_large-midas-2f21e586.pt"),
+    "dpt_hybrid": os.path.join(annotator_ckpts_path, "dpt_hybrid-midas-501f0c75.pt"),
+    "midas_v21": "",
+    "midas_v21_small": "",
+}
+
+remote_model_path = "https://huggingface.co/lllyasviel/ControlNet/resolve/main/annotator/ckpts/dpt_hybrid-midas-501f0c75.pt"
+
+
+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 load_midas_transform(model_type):
+    # https://github.com/isl-org/MiDaS/blob/master/run.py
+    # load transform only
+    if model_type == "dpt_large":  # DPT-Large
+        net_w, net_h = 384, 384
+        resize_mode = "minimal"
+        normalization = NormalizeImage(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5])
+
+    elif model_type == "dpt_hybrid":  # DPT-Hybrid
+        net_w, net_h = 384, 384
+        resize_mode = "minimal"
+        normalization = NormalizeImage(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5])
+
+    elif model_type == "midas_v21":
+        net_w, net_h = 384, 384
+        resize_mode = "upper_bound"
+        normalization = NormalizeImage(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
+
+    elif model_type == "midas_v21_small":
+        net_w, net_h = 256, 256
+        resize_mode = "upper_bound"
+        normalization = NormalizeImage(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
+
+    else:
+        assert False, f"model_type '{model_type}' not implemented, use: --model_type large"
+
+    transform = Compose(
+        [
+            Resize(
+                net_w,
+                net_h,
+                resize_target=None,
+                keep_aspect_ratio=True,
+                ensure_multiple_of=32,
+                resize_method=resize_mode,
+                image_interpolation_method=cv2.INTER_CUBIC,
+            ),
+            normalization,
+            PrepareForNet(),
+        ]
+    )
+
+    return transform
+
+
+def load_model(model_type):
+    # https://github.com/isl-org/MiDaS/blob/master/run.py
+    # load network
+    model_path = ISL_PATHS[model_type]
+    if model_type == "dpt_large":  # DPT-Large
+        model = DPTDepthModel(
+            path=model_path,
+            backbone="vitl16_384",
+            non_negative=True,
+        )
+        net_w, net_h = 384, 384
+        resize_mode = "minimal"
+        normalization = NormalizeImage(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5])
+
+    elif model_type == "dpt_hybrid":  # DPT-Hybrid
+        if not os.path.exists(model_path):
+            from basicsr.utils.download_util import load_file_from_url
+            load_file_from_url(remote_model_path, model_dir=annotator_ckpts_path)
+
+        model = DPTDepthModel(
+            path=model_path,
+            backbone="vitb_rn50_384",
+            non_negative=True,
+        )
+        net_w, net_h = 384, 384
+        resize_mode = "minimal"
+        normalization = NormalizeImage(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5])
+
+    elif model_type == "midas_v21":
+        model = MidasNet(model_path, non_negative=True)
+        net_w, net_h = 384, 384
+        resize_mode = "upper_bound"
+        normalization = NormalizeImage(
+            mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]
+        )
+
+    elif model_type == "midas_v21_small":
+        model = MidasNet_small(model_path, features=64, backbone="efficientnet_lite3", exportable=True,
+                               non_negative=True, blocks={'expand': True})
+        net_w, net_h = 256, 256
+        resize_mode = "upper_bound"
+        normalization = NormalizeImage(
+            mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]
+        )
+
+    else:
+        print(f"model_type '{model_type}' not implemented, use: --model_type large")
+        assert False
+
+    transform = Compose(
+        [
+            Resize(
+                net_w,
+                net_h,
+                resize_target=None,
+                keep_aspect_ratio=True,
+                ensure_multiple_of=32,
+                resize_method=resize_mode,
+                image_interpolation_method=cv2.INTER_CUBIC,
+            ),
+            normalization,
+            PrepareForNet(),
+        ]
+    )
+
+    return model.eval(), transform
+
+
+class MiDaSInference(nn.Module):
+    MODEL_TYPES_TORCH_HUB = [
+        "DPT_Large",
+        "DPT_Hybrid",
+        "MiDaS_small"
+    ]
+    MODEL_TYPES_ISL = [
+        "dpt_large",
+        "dpt_hybrid",
+        "midas_v21",
+        "midas_v21_small",
+    ]
+
+    def __init__(self, model_type):
+        super().__init__()
+        assert (model_type in self.MODEL_TYPES_ISL)
+        model, _ = load_model(model_type)
+        self.model = model
+        self.model.train = disabled_train
+
+    def forward(self, x):
+        with torch.no_grad():
+            prediction = self.model(x)
+        return prediction
+

lavis/common/annotator/midas/midas/base_model.py

@@ -0,0 +1,16 @@
+import torch
+
+
+class BaseModel(torch.nn.Module):
+    def load(self, path):
+        """Load model from file.
+
+        Args:
+            path (str): file path
+        """
+        parameters = torch.load(path, map_location=torch.device('cpu'))
+
+        if "optimizer" in parameters:
+            parameters = parameters["model"]
+
+        self.load_state_dict(parameters)

lavis/common/annotator/midas/midas/blocks.py

@@ -0,0 +1,342 @@
+import torch
+import torch.nn as nn
+
+from .vit import (
+    _make_pretrained_vitb_rn50_384,
+    _make_pretrained_vitl16_384,
+    _make_pretrained_vitb16_384,
+    forward_vit,
+)
+
+def _make_encoder(backbone, features, use_pretrained, groups=1, expand=False, exportable=True, hooks=None, use_vit_only=False, use_readout="ignore",):
+    if backbone == "vitl16_384":
+        pretrained = _make_pretrained_vitl16_384(
+            use_pretrained, hooks=hooks, use_readout=use_readout
+        )
+        scratch = _make_scratch(
+            [256, 512, 1024, 1024], features, groups=groups, expand=expand
+        )  # ViT-L/16 - 85.0% Top1 (backbone)
+    elif backbone == "vitb_rn50_384":
+        pretrained = _make_pretrained_vitb_rn50_384(
+            use_pretrained,
+            hooks=hooks,
+            use_vit_only=use_vit_only,
+            use_readout=use_readout,
+        )
+        scratch = _make_scratch(
+            [256, 512, 768, 768], features, groups=groups, expand=expand
+        )  # ViT-H/16 - 85.0% Top1 (backbone)
+    elif backbone == "vitb16_384":
+        pretrained = _make_pretrained_vitb16_384(
+            use_pretrained, hooks=hooks, use_readout=use_readout
+        )
+        scratch = _make_scratch(
+            [96, 192, 384, 768], features, groups=groups, expand=expand
+        )  # ViT-B/16 - 84.6% Top1 (backbone)
+    elif backbone == "resnext101_wsl":
+        pretrained = _make_pretrained_resnext101_wsl(use_pretrained)
+        scratch = _make_scratch([256, 512, 1024, 2048], features, groups=groups, expand=expand)     # efficientnet_lite3  
+    elif backbone == "efficientnet_lite3":
+        pretrained = _make_pretrained_efficientnet_lite3(use_pretrained, exportable=exportable)
+        scratch = _make_scratch([32, 48, 136, 384], features, groups=groups, expand=expand)  # efficientnet_lite3     
+    else:
+        print(f"Backbone '{backbone}' not implemented")
+        assert False
+        
+    return pretrained, scratch
+
+
+def _make_scratch(in_shape, out_shape, groups=1, expand=False):
+    scratch = nn.Module()
+
+    out_shape1 = out_shape
+    out_shape2 = out_shape
+    out_shape3 = out_shape
+    out_shape4 = out_shape
+    if expand==True:
+        out_shape1 = out_shape
+        out_shape2 = out_shape*2
+        out_shape3 = out_shape*4
+        out_shape4 = out_shape*8
+
+    scratch.layer1_rn = nn.Conv2d(
+        in_shape[0], out_shape1, kernel_size=3, stride=1, padding=1, bias=False, groups=groups
+    )
+    scratch.layer2_rn = nn.Conv2d(
+        in_shape[1], out_shape2, kernel_size=3, stride=1, padding=1, bias=False, groups=groups
+    )
+    scratch.layer3_rn = nn.Conv2d(
+        in_shape[2], out_shape3, kernel_size=3, stride=1, padding=1, bias=False, groups=groups
+    )
+    scratch.layer4_rn = nn.Conv2d(
+        in_shape[3], out_shape4, kernel_size=3, stride=1, padding=1, bias=False, groups=groups
+    )
+
+    return scratch
+
+
+def _make_pretrained_efficientnet_lite3(use_pretrained, exportable=False):
+    efficientnet = torch.hub.load(
+        "rwightman/gen-efficientnet-pytorch",
+        "tf_efficientnet_lite3",
+        pretrained=use_pretrained,
+        exportable=exportable
+    )
+    return _make_efficientnet_backbone(efficientnet)
+
+
+def _make_efficientnet_backbone(effnet):
+    pretrained = nn.Module()
+
+    pretrained.layer1 = nn.Sequential(
+        effnet.conv_stem, effnet.bn1, effnet.act1, *effnet.blocks[0:2]
+    )
+    pretrained.layer2 = nn.Sequential(*effnet.blocks[2:3])
+    pretrained.layer3 = nn.Sequential(*effnet.blocks[3:5])
+    pretrained.layer4 = nn.Sequential(*effnet.blocks[5:9])
+
+    return pretrained
+    
+
+def _make_resnet_backbone(resnet):
+    pretrained = nn.Module()
+    pretrained.layer1 = nn.Sequential(
+        resnet.conv1, resnet.bn1, resnet.relu, resnet.maxpool, resnet.layer1
+    )
+
+    pretrained.layer2 = resnet.layer2
+    pretrained.layer3 = resnet.layer3
+    pretrained.layer4 = resnet.layer4
+
+    return pretrained
+
+
+def _make_pretrained_resnext101_wsl(use_pretrained):
+    resnet = torch.hub.load("facebookresearch/WSL-Images", "resnext101_32x8d_wsl")
+    return _make_resnet_backbone(resnet)
+
+
+
+class Interpolate(nn.Module):
+    """Interpolation module.
+    """
+
+    def __init__(self, scale_factor, mode, align_corners=False):
+        """Init.
+
+        Args:
+            scale_factor (float): scaling
+            mode (str): interpolation mode
+        """
+        super(Interpolate, self).__init__()
+
+        self.interp = nn.functional.interpolate
+        self.scale_factor = scale_factor
+        self.mode = mode
+        self.align_corners = align_corners
+
+    def forward(self, x):
+        """Forward pass.
+
+        Args:
+            x (tensor): input
+
+        Returns:
+            tensor: interpolated data
+        """
+
+        x = self.interp(
+            x, scale_factor=self.scale_factor, mode=self.mode, align_corners=self.align_corners
+        )
+
+        return x
+
+
+class ResidualConvUnit(nn.Module):
+    """Residual convolution module.
+    """
+
+    def __init__(self, features):
+        """Init.
+
+        Args:
+            features (int): number of features
+        """
+        super().__init__()
+
+        self.conv1 = nn.Conv2d(
+            features, features, kernel_size=3, stride=1, padding=1, bias=True
+        )
+
+        self.conv2 = nn.Conv2d(
+            features, features, kernel_size=3, stride=1, padding=1, bias=True
+        )
+
+        self.relu = nn.ReLU(inplace=True)
+
+    def forward(self, x):
+        """Forward pass.
+
+        Args:
+            x (tensor): input
+
+        Returns:
+            tensor: output
+        """
+        out = self.relu(x)
+        out = self.conv1(out)
+        out = self.relu(out)
+        out = self.conv2(out)
+
+        return out + x
+
+
+class FeatureFusionBlock(nn.Module):
+    """Feature fusion block.
+    """
+
+    def __init__(self, features):
+        """Init.
+
+        Args:
+            features (int): number of features
+        """
+        super(FeatureFusionBlock, self).__init__()
+
+        self.resConfUnit1 = ResidualConvUnit(features)
+        self.resConfUnit2 = ResidualConvUnit(features)
+
+    def forward(self, *xs):
+        """Forward pass.
+
+        Returns:
+            tensor: output
+        """
+        output = xs[0]
+
+        if len(xs) == 2:
+            output += self.resConfUnit1(xs[1])
+
+        output = self.resConfUnit2(output)
+
+        output = nn.functional.interpolate(
+            output, scale_factor=2, mode="bilinear", align_corners=True
+        )
+
+        return output
+
+
+
+
+class ResidualConvUnit_custom(nn.Module):
+    """Residual convolution module.
+    """
+
+    def __init__(self, features, activation, bn):
+        """Init.
+
+        Args:
+            features (int): number of features
+        """
+        super().__init__()
+
+        self.bn = bn
+
+        self.groups=1
+
+        self.conv1 = nn.Conv2d(
+            features, features, kernel_size=3, stride=1, padding=1, bias=True, groups=self.groups
+        )
+        
+        self.conv2 = nn.Conv2d(
+            features, features, kernel_size=3, stride=1, padding=1, bias=True, groups=self.groups
+        )
+
+        if self.bn==True:
+            self.bn1 = nn.BatchNorm2d(features)
+            self.bn2 = nn.BatchNorm2d(features)
+
+        self.activation = activation
+
+        self.skip_add = nn.quantized.FloatFunctional()
+
+    def forward(self, x):
+        """Forward pass.
+
+        Args:
+            x (tensor): input
+
+        Returns:
+            tensor: output
+        """
+        
+        out = self.activation(x)
+        out = self.conv1(out)
+        if self.bn==True:
+            out = self.bn1(out)
+       
+        out = self.activation(out)
+        out = self.conv2(out)
+        if self.bn==True:
+            out = self.bn2(out)
+
+        if self.groups > 1:
+            out = self.conv_merge(out)
+
+        return self.skip_add.add(out, x)
+
+        # return out + x
+
+
+class FeatureFusionBlock_custom(nn.Module):
+    """Feature fusion block.
+    """
+
+    def __init__(self, features, activation, deconv=False, bn=False, expand=False, align_corners=True):
+        """Init.
+
+        Args:
+            features (int): number of features
+        """
+        super(FeatureFusionBlock_custom, self).__init__()
+
+        self.deconv = deconv
+        self.align_corners = align_corners
+
+        self.groups=1
+
+        self.expand = expand
+        out_features = features
+        if self.expand==True:
+            out_features = features//2
+        
+        self.out_conv = nn.Conv2d(features, out_features, kernel_size=1, stride=1, padding=0, bias=True, groups=1)
+
+        self.resConfUnit1 = ResidualConvUnit_custom(features, activation, bn)
+        self.resConfUnit2 = ResidualConvUnit_custom(features, activation, bn)
+        
+        self.skip_add = nn.quantized.FloatFunctional()
+
+    def forward(self, *xs):
+        """Forward pass.
+
+        Returns:
+            tensor: output
+        """
+        output = xs[0]
+
+        if len(xs) == 2:
+            res = self.resConfUnit1(xs[1])
+            output = self.skip_add.add(output, res)
+            # output += res
+
+        output = self.resConfUnit2(output)
+
+        output = nn.functional.interpolate(
+            output, scale_factor=2, mode="bilinear", align_corners=self.align_corners
+        )
+
+        output = self.out_conv(output)
+
+        return output
+

lavis/common/annotator/midas/midas/dpt_depth.py

@@ -0,0 +1,109 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from .base_model import BaseModel
+from .blocks import (
+    FeatureFusionBlock,
+    FeatureFusionBlock_custom,
+    Interpolate,
+    _make_encoder,
+    forward_vit,
+)
+
+
+def _make_fusion_block(features, use_bn):
+    return FeatureFusionBlock_custom(
+        features,
+        nn.ReLU(False),
+        deconv=False,
+        bn=use_bn,
+        expand=False,
+        align_corners=True,
+    )
+
+
+class DPT(BaseModel):
+    def __init__(
+        self,
+        head,
+        features=256,
+        backbone="vitb_rn50_384",
+        readout="project",
+        channels_last=False,
+        use_bn=False,
+    ):
+
+        super(DPT, self).__init__()
+
+        self.channels_last = channels_last
+
+        hooks = {
+            "vitb_rn50_384": [0, 1, 8, 11],
+            "vitb16_384": [2, 5, 8, 11],
+            "vitl16_384": [5, 11, 17, 23],
+        }
+
+        # Instantiate backbone and reassemble blocks
+        self.pretrained, self.scratch = _make_encoder(
+            backbone,
+            features,
+            False, # Set to true of you want to train from scratch, uses ImageNet weights
+            groups=1,
+            expand=False,
+            exportable=False,
+            hooks=hooks[backbone],
+            use_readout=readout,
+        )
+
+        self.scratch.refinenet1 = _make_fusion_block(features, use_bn)
+        self.scratch.refinenet2 = _make_fusion_block(features, use_bn)
+        self.scratch.refinenet3 = _make_fusion_block(features, use_bn)
+        self.scratch.refinenet4 = _make_fusion_block(features, use_bn)
+
+        self.scratch.output_conv = head
+
+
+    def forward(self, x):
+        if self.channels_last == True:
+            x.contiguous(memory_format=torch.channels_last)
+
+        layer_1, layer_2, layer_3, layer_4 = forward_vit(self.pretrained, x)
+
+        layer_1_rn = self.scratch.layer1_rn(layer_1)
+        layer_2_rn = self.scratch.layer2_rn(layer_2)
+        layer_3_rn = self.scratch.layer3_rn(layer_3)
+        layer_4_rn = self.scratch.layer4_rn(layer_4)
+
+        path_4 = self.scratch.refinenet4(layer_4_rn)
+        path_3 = self.scratch.refinenet3(path_4, layer_3_rn)
+        path_2 = self.scratch.refinenet2(path_3, layer_2_rn)
+        path_1 = self.scratch.refinenet1(path_2, layer_1_rn)
+
+        out = self.scratch.output_conv(path_1)
+
+        return out
+
+
+class DPTDepthModel(DPT):
+    def __init__(self, path=None, non_negative=True, **kwargs):
+        features = kwargs["features"] if "features" in kwargs else 256
+
+        head = nn.Sequential(
+            nn.Conv2d(features, features // 2, kernel_size=3, stride=1, padding=1),
+            Interpolate(scale_factor=2, mode="bilinear", align_corners=True),
+            nn.Conv2d(features // 2, 32, kernel_size=3, stride=1, padding=1),
+            nn.ReLU(True),
+            nn.Conv2d(32, 1, kernel_size=1, stride=1, padding=0),
+            nn.ReLU(True) if non_negative else nn.Identity(),
+            nn.Identity(),
+        )
+
+        super().__init__(head, **kwargs)
+
+        if path is not None:
+           self.load(path)
+
+    def forward(self, x):
+        return super().forward(x).squeeze(dim=1)
+

lavis/common/annotator/midas/midas/midas_net.py

@@ -0,0 +1,76 @@
+"""MidashNet: Network for monocular depth estimation trained by mixing several datasets.
+This file contains code that is adapted from
+https://github.com/thomasjpfan/pytorch_refinenet/blob/master/pytorch_refinenet/refinenet/refinenet_4cascade.py
+"""
+import torch
+import torch.nn as nn
+
+from .base_model import BaseModel
+from .blocks import FeatureFusionBlock, Interpolate, _make_encoder
+
+
+class MidasNet(BaseModel):
+    """Network for monocular depth estimation.
+    """
+
+    def __init__(self, path=None, features=256, non_negative=True):
+        """Init.
+
+        Args:
+            path (str, optional): Path to saved model. Defaults to None.
+            features (int, optional): Number of features. Defaults to 256.
+            backbone (str, optional): Backbone network for encoder. Defaults to resnet50
+        """
+        print("Loading weights: ", path)
+
+        super(MidasNet, self).__init__()
+
+        use_pretrained = False if path is None else True
+
+        self.pretrained, self.scratch = _make_encoder(backbone="resnext101_wsl", features=features, use_pretrained=use_pretrained)
+
+        self.scratch.refinenet4 = FeatureFusionBlock(features)
+        self.scratch.refinenet3 = FeatureFusionBlock(features)
+        self.scratch.refinenet2 = FeatureFusionBlock(features)
+        self.scratch.refinenet1 = FeatureFusionBlock(features)
+
+        self.scratch.output_conv = nn.Sequential(
+            nn.Conv2d(features, 128, kernel_size=3, stride=1, padding=1),
+            Interpolate(scale_factor=2, mode="bilinear"),
+            nn.Conv2d(128, 32, kernel_size=3, stride=1, padding=1),
+            nn.ReLU(True),
+            nn.Conv2d(32, 1, kernel_size=1, stride=1, padding=0),
+            nn.ReLU(True) if non_negative else nn.Identity(),
+        )
+
+        if path:
+            self.load(path)
+
+    def forward(self, x):
+        """Forward pass.
+
+        Args:
+            x (tensor): input data (image)
+
+        Returns:
+            tensor: depth
+        """
+
+        layer_1 = self.pretrained.layer1(x)
+        layer_2 = self.pretrained.layer2(layer_1)
+        layer_3 = self.pretrained.layer3(layer_2)
+        layer_4 = self.pretrained.layer4(layer_3)
+
+        layer_1_rn = self.scratch.layer1_rn(layer_1)
+        layer_2_rn = self.scratch.layer2_rn(layer_2)
+        layer_3_rn = self.scratch.layer3_rn(layer_3)
+        layer_4_rn = self.scratch.layer4_rn(layer_4)
+
+        path_4 = self.scratch.refinenet4(layer_4_rn)
+        path_3 = self.scratch.refinenet3(path_4, layer_3_rn)
+        path_2 = self.scratch.refinenet2(path_3, layer_2_rn)
+        path_1 = self.scratch.refinenet1(path_2, layer_1_rn)
+
+        out = self.scratch.output_conv(path_1)
+
+        return torch.squeeze(out, dim=1)

lavis/common/annotator/midas/midas/midas_net_custom.py

@@ -0,0 +1,128 @@
+"""MidashNet: Network for monocular depth estimation trained by mixing several datasets.
+This file contains code that is adapted from
+https://github.com/thomasjpfan/pytorch_refinenet/blob/master/pytorch_refinenet/refinenet/refinenet_4cascade.py
+"""
+import torch
+import torch.nn as nn
+
+from .base_model import BaseModel
+from .blocks import FeatureFusionBlock, FeatureFusionBlock_custom, Interpolate, _make_encoder
+
+
+class MidasNet_small(BaseModel):
+    """Network for monocular depth estimation.
+    """
+
+    def __init__(self, path=None, features=64, backbone="efficientnet_lite3", non_negative=True, exportable=True, channels_last=False, align_corners=True,
+        blocks={'expand': True}):
+        """Init.
+
+        Args:
+            path (str, optional): Path to saved model. Defaults to None.
+            features (int, optional): Number of features. Defaults to 256.
+            backbone (str, optional): Backbone network for encoder. Defaults to resnet50
+        """
+        print("Loading weights: ", path)
+
+        super(MidasNet_small, self).__init__()
+
+        use_pretrained = False if path else True
+                
+        self.channels_last = channels_last
+        self.blocks = blocks
+        self.backbone = backbone
+
+        self.groups = 1
+
+        features1=features
+        features2=features
+        features3=features
+        features4=features
+        self.expand = False
+        if "expand" in self.blocks and self.blocks['expand'] == True:
+            self.expand = True
+            features1=features
+            features2=features*2
+            features3=features*4
+            features4=features*8
+
+        self.pretrained, self.scratch = _make_encoder(self.backbone, features, use_pretrained, groups=self.groups, expand=self.expand, exportable=exportable)
+  
+        self.scratch.activation = nn.ReLU(False)    
+
+        self.scratch.refinenet4 = FeatureFusionBlock_custom(features4, self.scratch.activation, deconv=False, bn=False, expand=self.expand, align_corners=align_corners)
+        self.scratch.refinenet3 = FeatureFusionBlock_custom(features3, self.scratch.activation, deconv=False, bn=False, expand=self.expand, align_corners=align_corners)
+        self.scratch.refinenet2 = FeatureFusionBlock_custom(features2, self.scratch.activation, deconv=False, bn=False, expand=self.expand, align_corners=align_corners)
+        self.scratch.refinenet1 = FeatureFusionBlock_custom(features1, self.scratch.activation, deconv=False, bn=False, align_corners=align_corners)
+
+        
+        self.scratch.output_conv = nn.Sequential(
+            nn.Conv2d(features, features//2, kernel_size=3, stride=1, padding=1, groups=self.groups),
+            Interpolate(scale_factor=2, mode="bilinear"),
+            nn.Conv2d(features//2, 32, kernel_size=3, stride=1, padding=1),
+            self.scratch.activation,
+            nn.Conv2d(32, 1, kernel_size=1, stride=1, padding=0),
+            nn.ReLU(True) if non_negative else nn.Identity(),
+            nn.Identity(),
+        )
+        
+        if path:
+            self.load(path)
+
+
+    def forward(self, x):
+        """Forward pass.
+
+        Args:
+            x (tensor): input data (image)
+
+        Returns:
+            tensor: depth
+        """
+        if self.channels_last==True:
+            print("self.channels_last = ", self.channels_last)
+            x.contiguous(memory_format=torch.channels_last)
+
+
+        layer_1 = self.pretrained.layer1(x)
+        layer_2 = self.pretrained.layer2(layer_1)
+        layer_3 = self.pretrained.layer3(layer_2)
+        layer_4 = self.pretrained.layer4(layer_3)
+        
+        layer_1_rn = self.scratch.layer1_rn(layer_1)
+        layer_2_rn = self.scratch.layer2_rn(layer_2)
+        layer_3_rn = self.scratch.layer3_rn(layer_3)
+        layer_4_rn = self.scratch.layer4_rn(layer_4)
+
+
+        path_4 = self.scratch.refinenet4(layer_4_rn)
+        path_3 = self.scratch.refinenet3(path_4, layer_3_rn)
+        path_2 = self.scratch.refinenet2(path_3, layer_2_rn)
+        path_1 = self.scratch.refinenet1(path_2, layer_1_rn)
+        
+        out = self.scratch.output_conv(path_1)
+
+        return torch.squeeze(out, dim=1)
+
+
+
+def fuse_model(m):
+    prev_previous_type = nn.Identity()
+    prev_previous_name = ''
+    previous_type = nn.Identity()
+    previous_name = ''
+    for name, module in m.named_modules():
+        if prev_previous_type == nn.Conv2d and previous_type == nn.BatchNorm2d and type(module) == nn.ReLU:
+            # print("FUSED ", prev_previous_name, previous_name, name)
+            torch.quantization.fuse_modules(m, [prev_previous_name, previous_name, name], inplace=True)
+        elif prev_previous_type == nn.Conv2d and previous_type == nn.BatchNorm2d:
+            # print("FUSED ", prev_previous_name, previous_name)
+            torch.quantization.fuse_modules(m, [prev_previous_name, previous_name], inplace=True)
+        # elif previous_type == nn.Conv2d and type(module) == nn.ReLU:
+        #    print("FUSED ", previous_name, name)
+        #    torch.quantization.fuse_modules(m, [previous_name, name], inplace=True)
+
+        prev_previous_type = previous_type
+        prev_previous_name = previous_name
+        previous_type = type(module)
+        previous_name = name

lavis/common/annotator/midas/midas/transforms.py

@@ -0,0 +1,234 @@
+import numpy as np
+import cv2
+import math
+
+
+def apply_min_size(sample, size, image_interpolation_method=cv2.INTER_AREA):
+    """Rezise the sample to ensure the given size. Keeps aspect ratio.
+
+    Args:
+        sample (dict): sample
+        size (tuple): image size
+
+    Returns:
+        tuple: new size
+    """
+    shape = list(sample["disparity"].shape)
+
+    if shape[0] >= size[0] and shape[1] >= size[1]:
+        return sample
+
+    scale = [0, 0]
+    scale[0] = size[0] / shape[0]
+    scale[1] = size[1] / shape[1]
+
+    scale = max(scale)
+
+    shape[0] = math.ceil(scale * shape[0])
+    shape[1] = math.ceil(scale * shape[1])
+
+    # resize
+    sample["image"] = cv2.resize(
+        sample["image"], tuple(shape[::-1]), interpolation=image_interpolation_method
+    )
+
+    sample["disparity"] = cv2.resize(
+        sample["disparity"], tuple(shape[::-1]), interpolation=cv2.INTER_NEAREST
+    )
+    sample["mask"] = cv2.resize(
+        sample["mask"].astype(np.float32),
+        tuple(shape[::-1]),
+        interpolation=cv2.INTER_NEAREST,
+    )
+    sample["mask"] = sample["mask"].astype(bool)
+
+    return tuple(shape)
+
+
+class Resize(object):
+    """Resize sample to given size (width, height).
+    """
+
+    def __init__(
+        self,
+        width,
+        height,
+        resize_target=True,
+        keep_aspect_ratio=False,
+        ensure_multiple_of=1,
+        resize_method="lower_bound",
+        image_interpolation_method=cv2.INTER_AREA,
+    ):
+        """Init.
+
+        Args:
+            width (int): desired output width
+            height (int): desired output height
+            resize_target (bool, optional):
+                True: Resize the full sample (image, mask, target).
+                False: Resize image only.
+                Defaults to True.
+            keep_aspect_ratio (bool, optional):
+                True: Keep the aspect ratio of the input sample.
+                Output sample might not have the given width and height, and
+                resize behaviour depends on the parameter 'resize_method'.
+                Defaults to False.
+            ensure_multiple_of (int, optional):
+                Output width and height is constrained to be multiple of this parameter.
+                Defaults to 1.
+            resize_method (str, optional):
+                "lower_bound": Output will be at least as large as the given size.
+                "upper_bound": Output will be at max as large as the given size. (Output size might be smaller than given size.)
+                "minimal": Scale as least as possible.  (Output size might be smaller than given size.)
+                Defaults to "lower_bound".
+        """
+        self.__width = width
+        self.__height = height
+
+        self.__resize_target = resize_target
+        self.__keep_aspect_ratio = keep_aspect_ratio
+        self.__multiple_of = ensure_multiple_of
+        self.__resize_method = resize_method
+        self.__image_interpolation_method = image_interpolation_method
+
+    def constrain_to_multiple_of(self, x, min_val=0, max_val=None):
+        y = (np.round(x / self.__multiple_of) * self.__multiple_of).astype(int)
+
+        if max_val is not None and y > max_val:
+            y = (np.floor(x / self.__multiple_of) * self.__multiple_of).astype(int)
+
+        if y < min_val:
+            y = (np.ceil(x / self.__multiple_of) * self.__multiple_of).astype(int)
+
+        return y
+
+    def get_size(self, width, height):
+        # determine new height and width
+        scale_height = self.__height / height
+        scale_width = self.__width / width
+
+        if self.__keep_aspect_ratio:
+            if self.__resize_method == "lower_bound":
+                # scale such that output size is lower bound
+                if scale_width > scale_height:
+                    # fit width
+                    scale_height = scale_width
+                else:
+                    # fit height
+                    scale_width = scale_height
+            elif self.__resize_method == "upper_bound":
+                # scale such that output size is upper bound
+                if scale_width < scale_height:
+                    # fit width
+                    scale_height = scale_width
+                else:
+                    # fit height
+                    scale_width = scale_height
+            elif self.__resize_method == "minimal":
+                # scale as least as possbile
+                if abs(1 - scale_width) < abs(1 - scale_height):
+                    # fit width
+                    scale_height = scale_width
+                else:
+                    # fit height
+                    scale_width = scale_height
+            else:
+                raise ValueError(
+                    f"resize_method {self.__resize_method} not implemented"
+                )
+
+        if self.__resize_method == "lower_bound":
+            new_height = self.constrain_to_multiple_of(
+                scale_height * height, min_val=self.__height
+            )
+            new_width = self.constrain_to_multiple_of(
+                scale_width * width, min_val=self.__width
+            )
+        elif self.__resize_method == "upper_bound":
+            new_height = self.constrain_to_multiple_of(
+                scale_height * height, max_val=self.__height
+            )
+            new_width = self.constrain_to_multiple_of(
+                scale_width * width, max_val=self.__width
+            )
+        elif self.__resize_method == "minimal":
+            new_height = self.constrain_to_multiple_of(scale_height * height)
+            new_width = self.constrain_to_multiple_of(scale_width * width)
+        else:
+            raise ValueError(f"resize_method {self.__resize_method} not implemented")
+
+        return (new_width, new_height)
+
+    def __call__(self, sample):
+        width, height = self.get_size(
+            sample["image"].shape[1], sample["image"].shape[0]
+        )
+
+        # resize sample
+        sample["image"] = cv2.resize(
+            sample["image"],
+            (width, height),
+            interpolation=self.__image_interpolation_method,
+        )
+
+        if self.__resize_target:
+            if "disparity" in sample:
+                sample["disparity"] = cv2.resize(
+                    sample["disparity"],
+                    (width, height),
+                    interpolation=cv2.INTER_NEAREST,
+                )
+
+            if "depth" in sample:
+                sample["depth"] = cv2.resize(
+                    sample["depth"], (width, height), interpolation=cv2.INTER_NEAREST
+                )
+
+            sample["mask"] = cv2.resize(
+                sample["mask"].astype(np.float32),
+                (width, height),
+                interpolation=cv2.INTER_NEAREST,
+            )
+            sample["mask"] = sample["mask"].astype(bool)
+
+        return sample
+
+
+class NormalizeImage(object):
+    """Normlize image by given mean and std.
+    """
+
+    def __init__(self, mean, std):
+        self.__mean = mean
+        self.__std = std
+
+    def __call__(self, sample):
+        sample["image"] = (sample["image"] - self.__mean) / self.__std
+
+        return sample
+
+
+class PrepareForNet(object):
+    """Prepare sample for usage as network input.
+    """
+
+    def __init__(self):
+        pass
+
+    def __call__(self, sample):
+        image = np.transpose(sample["image"], (2, 0, 1))
+        sample["image"] = np.ascontiguousarray(image).astype(np.float32)
+
+        if "mask" in sample:
+            sample["mask"] = sample["mask"].astype(np.float32)
+            sample["mask"] = np.ascontiguousarray(sample["mask"])
+
+        if "disparity" in sample:
+            disparity = sample["disparity"].astype(np.float32)
+            sample["disparity"] = np.ascontiguousarray(disparity)
+
+        if "depth" in sample:
+            depth = sample["depth"].astype(np.float32)
+            sample["depth"] = np.ascontiguousarray(depth)
+
+        return sample

lavis/common/annotator/midas/midas/vit.py

@@ -0,0 +1,491 @@
+import torch
+import torch.nn as nn
+import timm
+import types
+import math
+import torch.nn.functional as F
+
+
+class Slice(nn.Module):
+    def __init__(self, start_index=1):
+        super(Slice, self).__init__()
+        self.start_index = start_index
+
+    def forward(self, x):
+        return x[:, self.start_index :]
+
+
+class AddReadout(nn.Module):
+    def __init__(self, start_index=1):
+        super(AddReadout, self).__init__()
+        self.start_index = start_index
+
+    def forward(self, x):
+        if self.start_index == 2:
+            readout = (x[:, 0] + x[:, 1]) / 2
+        else:
+            readout = x[:, 0]
+        return x[:, self.start_index :] + readout.unsqueeze(1)
+
+
+class ProjectReadout(nn.Module):
+    def __init__(self, in_features, start_index=1):
+        super(ProjectReadout, self).__init__()
+        self.start_index = start_index
+
+        self.project = nn.Sequential(nn.Linear(2 * in_features, in_features), nn.GELU())
+
+    def forward(self, x):
+        readout = x[:, 0].unsqueeze(1).expand_as(x[:, self.start_index :])
+        features = torch.cat((x[:, self.start_index :], readout), -1)
+
+        return self.project(features)
+
+
+class Transpose(nn.Module):
+    def __init__(self, dim0, dim1):
+        super(Transpose, self).__init__()
+        self.dim0 = dim0
+        self.dim1 = dim1
+
+    def forward(self, x):
+        x = x.transpose(self.dim0, self.dim1)
+        return x
+
+
+def forward_vit(pretrained, x):
+    b, c, h, w = x.shape
+
+    glob = pretrained.model.forward_flex(x)
+
+    layer_1 = pretrained.activations["1"]
+    layer_2 = pretrained.activations["2"]
+    layer_3 = pretrained.activations["3"]
+    layer_4 = pretrained.activations["4"]
+
+    layer_1 = pretrained.act_postprocess1[0:2](layer_1)
+    layer_2 = pretrained.act_postprocess2[0:2](layer_2)
+    layer_3 = pretrained.act_postprocess3[0:2](layer_3)
+    layer_4 = pretrained.act_postprocess4[0:2](layer_4)
+
+    unflatten = nn.Sequential(
+        nn.Unflatten(
+            2,
+            torch.Size(
+                [
+                    h // pretrained.model.patch_size[1],
+                    w // pretrained.model.patch_size[0],
+                ]
+            ),
+        )
+    )
+
+    if layer_1.ndim == 3:
+        layer_1 = unflatten(layer_1)
+    if layer_2.ndim == 3:
+        layer_2 = unflatten(layer_2)
+    if layer_3.ndim == 3:
+        layer_3 = unflatten(layer_3)
+    if layer_4.ndim == 3:
+        layer_4 = unflatten(layer_4)
+
+    layer_1 = pretrained.act_postprocess1[3 : len(pretrained.act_postprocess1)](layer_1)
+    layer_2 = pretrained.act_postprocess2[3 : len(pretrained.act_postprocess2)](layer_2)
+    layer_3 = pretrained.act_postprocess3[3 : len(pretrained.act_postprocess3)](layer_3)
+    layer_4 = pretrained.act_postprocess4[3 : len(pretrained.act_postprocess4)](layer_4)
+
+    return layer_1, layer_2, layer_3, layer_4
+
+
+def _resize_pos_embed(self, posemb, gs_h, gs_w):
+    posemb_tok, posemb_grid = (
+        posemb[:, : self.start_index],
+        posemb[0, self.start_index :],
+    )
+
+    gs_old = int(math.sqrt(len(posemb_grid)))
+
+    posemb_grid = posemb_grid.reshape(1, gs_old, gs_old, -1).permute(0, 3, 1, 2)
+    posemb_grid = F.interpolate(posemb_grid, size=(gs_h, gs_w), mode="bilinear")
+    posemb_grid = posemb_grid.permute(0, 2, 3, 1).reshape(1, gs_h * gs_w, -1)
+
+    posemb = torch.cat([posemb_tok, posemb_grid], dim=1)
+
+    return posemb
+
+
+def forward_flex(self, x):
+    b, c, h, w = x.shape
+
+    pos_embed = self._resize_pos_embed(
+        self.pos_embed, h // self.patch_size[1], w // self.patch_size[0]
+    )
+
+    B = x.shape[0]
+
+    if hasattr(self.patch_embed, "backbone"):
+        x = self.patch_embed.backbone(x)
+        if isinstance(x, (list, tuple)):
+            x = x[-1]  # last feature if backbone outputs list/tuple of features
+
+    x = self.patch_embed.proj(x).flatten(2).transpose(1, 2)
+
+    if getattr(self, "dist_token", None) is not None:
+        cls_tokens = self.cls_token.expand(
+            B, -1, -1
+        )  # stole cls_tokens impl from Phil Wang, thanks
+        dist_token = self.dist_token.expand(B, -1, -1)
+        x = torch.cat((cls_tokens, dist_token, x), dim=1)
+    else:
+        cls_tokens = self.cls_token.expand(
+            B, -1, -1
+        )  # stole cls_tokens impl from Phil Wang, thanks
+        x = torch.cat((cls_tokens, x), dim=1)
+
+    x = x + pos_embed
+    x = self.pos_drop(x)
+
+    for blk in self.blocks:
+        x = blk(x)
+
+    x = self.norm(x)
+
+    return x
+
+
+activations = {}
+
+
+def get_activation(name):
+    def hook(model, input, output):
+        activations[name] = output
+
+    return hook
+
+
+def get_readout_oper(vit_features, features, use_readout, start_index=1):
+    if use_readout == "ignore":
+        readout_oper = [Slice(start_index)] * len(features)
+    elif use_readout == "add":
+        readout_oper = [AddReadout(start_index)] * len(features)
+    elif use_readout == "project":
+        readout_oper = [
+            ProjectReadout(vit_features, start_index) for out_feat in features
+        ]
+    else:
+        assert (
+            False
+        ), "wrong operation for readout token, use_readout can be 'ignore', 'add', or 'project'"
+
+    return readout_oper
+
+
+def _make_vit_b16_backbone(
+    model,
+    features=[96, 192, 384, 768],
+    size=[384, 384],
+    hooks=[2, 5, 8, 11],
+    vit_features=768,
+    use_readout="ignore",
+    start_index=1,
+):
+    pretrained = nn.Module()
+
+    pretrained.model = model
+    pretrained.model.blocks[hooks[0]].register_forward_hook(get_activation("1"))
+    pretrained.model.blocks[hooks[1]].register_forward_hook(get_activation("2"))
+    pretrained.model.blocks[hooks[2]].register_forward_hook(get_activation("3"))
+    pretrained.model.blocks[hooks[3]].register_forward_hook(get_activation("4"))
+
+    pretrained.activations = activations
+
+    readout_oper = get_readout_oper(vit_features, features, use_readout, start_index)
+
+    # 32, 48, 136, 384
+    pretrained.act_postprocess1 = nn.Sequential(
+        readout_oper[0],
+        Transpose(1, 2),
+        nn.Unflatten(2, torch.Size([size[0] // 16, size[1] // 16])),
+        nn.Conv2d(
+            in_channels=vit_features,
+            out_channels=features[0],
+            kernel_size=1,
+            stride=1,
+            padding=0,
+        ),
+        nn.ConvTranspose2d(
+            in_channels=features[0],
+            out_channels=features[0],
+            kernel_size=4,
+            stride=4,
+            padding=0,
+            bias=True,
+            dilation=1,
+            groups=1,
+        ),
+    )
+
+    pretrained.act_postprocess2 = nn.Sequential(
+        readout_oper[1],
+        Transpose(1, 2),
+        nn.Unflatten(2, torch.Size([size[0] // 16, size[1] // 16])),
+        nn.Conv2d(
+            in_channels=vit_features,
+            out_channels=features[1],
+            kernel_size=1,
+            stride=1,
+            padding=0,
+        ),
+        nn.ConvTranspose2d(
+            in_channels=features[1],
+            out_channels=features[1],
+            kernel_size=2,
+            stride=2,
+            padding=0,
+            bias=True,
+            dilation=1,
+            groups=1,
+        ),
+    )
+
+    pretrained.act_postprocess3 = nn.Sequential(
+        readout_oper[2],
+        Transpose(1, 2),
+        nn.Unflatten(2, torch.Size([size[0] // 16, size[1] // 16])),
+        nn.Conv2d(
+            in_channels=vit_features,
+            out_channels=features[2],
+            kernel_size=1,
+            stride=1,
+            padding=0,
+        ),
+    )
+
+    pretrained.act_postprocess4 = nn.Sequential(
+        readout_oper[3],
+        Transpose(1, 2),
+        nn.Unflatten(2, torch.Size([size[0] // 16, size[1] // 16])),
+        nn.Conv2d(
+            in_channels=vit_features,
+            out_channels=features[3],
+            kernel_size=1,
+            stride=1,
+            padding=0,
+        ),
+        nn.Conv2d(
+            in_channels=features[3],
+            out_channels=features[3],
+            kernel_size=3,
+            stride=2,
+            padding=1,
+        ),
+    )
+
+    pretrained.model.start_index = start_index
+    pretrained.model.patch_size = [16, 16]
+
+    # We inject this function into the VisionTransformer instances so that
+    # we can use it with interpolated position embeddings without modifying the library source.
+    pretrained.model.forward_flex = types.MethodType(forward_flex, pretrained.model)
+    pretrained.model._resize_pos_embed = types.MethodType(
+        _resize_pos_embed, pretrained.model
+    )
+
+    return pretrained
+
+
+def _make_pretrained_vitl16_384(pretrained, use_readout="ignore", hooks=None):
+    model = timm.create_model("vit_large_patch16_384", pretrained=pretrained)
+
+    hooks = [5, 11, 17, 23] if hooks == None else hooks
+    return _make_vit_b16_backbone(
+        model,
+        features=[256, 512, 1024, 1024],
+        hooks=hooks,
+        vit_features=1024,
+        use_readout=use_readout,
+    )
+
+
+def _make_pretrained_vitb16_384(pretrained, use_readout="ignore", hooks=None):
+    model = timm.create_model("vit_base_patch16_384", pretrained=pretrained)
+
+    hooks = [2, 5, 8, 11] if hooks == None else hooks
+    return _make_vit_b16_backbone(
+        model, features=[96, 192, 384, 768], hooks=hooks, use_readout=use_readout
+    )
+
+
+def _make_pretrained_deitb16_384(pretrained, use_readout="ignore", hooks=None):
+    model = timm.create_model("vit_deit_base_patch16_384", pretrained=pretrained)
+
+    hooks = [2, 5, 8, 11] if hooks == None else hooks
+    return _make_vit_b16_backbone(
+        model, features=[96, 192, 384, 768], hooks=hooks, use_readout=use_readout
+    )
+
+
+def _make_pretrained_deitb16_distil_384(pretrained, use_readout="ignore", hooks=None):
+    model = timm.create_model(
+        "vit_deit_base_distilled_patch16_384", pretrained=pretrained
+    )
+
+    hooks = [2, 5, 8, 11] if hooks == None else hooks
+    return _make_vit_b16_backbone(
+        model,
+        features=[96, 192, 384, 768],
+        hooks=hooks,
+        use_readout=use_readout,
+        start_index=2,
+    )
+
+
+def _make_vit_b_rn50_backbone(
+    model,
+    features=[256, 512, 768, 768],
+    size=[384, 384],
+    hooks=[0, 1, 8, 11],
+    vit_features=768,
+    use_vit_only=False,
+    use_readout="ignore",
+    start_index=1,
+):
+    pretrained = nn.Module()
+
+    pretrained.model = model
+
+    if use_vit_only == True:
+        pretrained.model.blocks[hooks[0]].register_forward_hook(get_activation("1"))
+        pretrained.model.blocks[hooks[1]].register_forward_hook(get_activation("2"))
+    else:
+        pretrained.model.patch_embed.backbone.stages[0].register_forward_hook(
+            get_activation("1")
+        )
+        pretrained.model.patch_embed.backbone.stages[1].register_forward_hook(
+            get_activation("2")
+        )
+
+    pretrained.model.blocks[hooks[2]].register_forward_hook(get_activation("3"))
+    pretrained.model.blocks[hooks[3]].register_forward_hook(get_activation("4"))
+
+    pretrained.activations = activations
+
+    readout_oper = get_readout_oper(vit_features, features, use_readout, start_index)
+
+    if use_vit_only == True:
+        pretrained.act_postprocess1 = nn.Sequential(
+            readout_oper[0],
+            Transpose(1, 2),
+            nn.Unflatten(2, torch.Size([size[0] // 16, size[1] // 16])),
+            nn.Conv2d(
+                in_channels=vit_features,
+                out_channels=features[0],
+                kernel_size=1,
+                stride=1,
+                padding=0,
+            ),
+            nn.ConvTranspose2d(
+                in_channels=features[0],
+                out_channels=features[0],
+                kernel_size=4,
+                stride=4,
+                padding=0,
+                bias=True,
+                dilation=1,
+                groups=1,
+            ),
+        )
+
+        pretrained.act_postprocess2 = nn.Sequential(
+            readout_oper[1],
+            Transpose(1, 2),
+            nn.Unflatten(2, torch.Size([size[0] // 16, size[1] // 16])),
+            nn.Conv2d(
+                in_channels=vit_features,
+                out_channels=features[1],
+                kernel_size=1,
+                stride=1,
+                padding=0,
+            ),
+            nn.ConvTranspose2d(
+                in_channels=features[1],
+                out_channels=features[1],
+                kernel_size=2,
+                stride=2,
+                padding=0,
+                bias=True,
+                dilation=1,
+                groups=1,
+            ),
+        )
+    else:
+        pretrained.act_postprocess1 = nn.Sequential(
+            nn.Identity(), nn.Identity(), nn.Identity()
+        )
+        pretrained.act_postprocess2 = nn.Sequential(
+            nn.Identity(), nn.Identity(), nn.Identity()
+        )
+
+    pretrained.act_postprocess3 = nn.Sequential(
+        readout_oper[2],
+        Transpose(1, 2),
+        nn.Unflatten(2, torch.Size([size[0] // 16, size[1] // 16])),
+        nn.Conv2d(
+            in_channels=vit_features,
+            out_channels=features[2],
+            kernel_size=1,
+            stride=1,
+            padding=0,
+        ),
+    )
+
+    pretrained.act_postprocess4 = nn.Sequential(
+        readout_oper[3],
+        Transpose(1, 2),
+        nn.Unflatten(2, torch.Size([size[0] // 16, size[1] // 16])),
+        nn.Conv2d(
+            in_channels=vit_features,
+            out_channels=features[3],
+            kernel_size=1,
+            stride=1,
+            padding=0,
+        ),
+        nn.Conv2d(
+            in_channels=features[3],
+            out_channels=features[3],
+            kernel_size=3,
+            stride=2,
+            padding=1,
+        ),
+    )
+
+    pretrained.model.start_index = start_index
+    pretrained.model.patch_size = [16, 16]
+
+    # We inject this function into the VisionTransformer instances so that
+    # we can use it with interpolated position embeddings without modifying the library source.
+    pretrained.model.forward_flex = types.MethodType(forward_flex, pretrained.model)
+
+    # We inject this function into the VisionTransformer instances so that
+    # we can use it with interpolated position embeddings without modifying the library source.
+    pretrained.model._resize_pos_embed = types.MethodType(
+        _resize_pos_embed, pretrained.model
+    )
+
+    return pretrained
+
+
+def _make_pretrained_vitb_rn50_384(
+    pretrained, use_readout="ignore", hooks=None, use_vit_only=False
+):
+    model = timm.create_model("vit_base_resnet50_384", pretrained=pretrained)
+
+    hooks = [0, 1, 8, 11] if hooks == None else hooks
+    return _make_vit_b_rn50_backbone(
+        model,
+        features=[256, 512, 768, 768],
+        size=[384, 384],
+        hooks=hooks,
+        use_vit_only=use_vit_only,
+        use_readout=use_readout,
+    )

lavis/common/annotator/midas/utils.py

@@ -0,0 +1,189 @@
+"""Utils for monoDepth."""
+import sys
+import re
+import numpy as np
+import cv2
+import torch
+
+
+def read_pfm(path):
+    """Read pfm file.
+
+    Args:
+        path (str): path to file
+
+    Returns:
+        tuple: (data, scale)
+    """
+    with open(path, "rb") as file:
+
+        color = None
+        width = None
+        height = None
+        scale = None
+        endian = None
+
+        header = file.readline().rstrip()
+        if header.decode("ascii") == "PF":
+            color = True
+        elif header.decode("ascii") == "Pf":
+            color = False
+        else:
+            raise Exception("Not a PFM file: " + path)
+
+        dim_match = re.match(r"^(\d+)\s(\d+)\s$", file.readline().decode("ascii"))
+        if dim_match:
+            width, height = list(map(int, dim_match.groups()))
+        else:
+            raise Exception("Malformed PFM header.")
+
+        scale = float(file.readline().decode("ascii").rstrip())
+        if scale < 0:
+            # little-endian
+            endian = "<"
+            scale = -scale
+        else:
+            # big-endian
+            endian = ">"
+
+        data = np.fromfile(file, endian + "f")
+        shape = (height, width, 3) if color else (height, width)
+
+        data = np.reshape(data, shape)
+        data = np.flipud(data)
+
+        return data, scale
+
+
+def write_pfm(path, image, scale=1):
+    """Write pfm file.
+
+    Args:
+        path (str): pathto file
+        image (array): data
+        scale (int, optional): Scale. Defaults to 1.
+    """
+
+    with open(path, "wb") as file:
+        color = None
+
+        if image.dtype.name != "float32":
+            raise Exception("Image dtype must be float32.")
+
+        image = np.flipud(image)
+
+        if len(image.shape) == 3 and image.shape[2] == 3:  # color image
+            color = True
+        elif (
+            len(image.shape) == 2 or len(image.shape) == 3 and image.shape[2] == 1
+        ):  # greyscale
+            color = False
+        else:
+            raise Exception("Image must have H x W x 3, H x W x 1 or H x W dimensions.")
+
+        file.write("PF\n" if color else "Pf\n".encode())
+        file.write("%d %d\n".encode() % (image.shape[1], image.shape[0]))
+
+        endian = image.dtype.byteorder
+
+        if endian == "<" or endian == "=" and sys.byteorder == "little":
+            scale = -scale
+
+        file.write("%f\n".encode() % scale)
+
+        image.tofile(file)
+
+
+def read_image(path):
+    """Read image and output RGB image (0-1).
+
+    Args:
+        path (str): path to file
+
+    Returns:
+        array: RGB image (0-1)
+    """
+    img = cv2.imread(path)
+
+    if img.ndim == 2:
+        img = cv2.cvtColor(img, cv2.COLOR_GRAY2BGR)
+
+    img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB) / 255.0
+
+    return img
+
+
+def resize_image(img):
+    """Resize image and make it fit for network.
+
+    Args:
+        img (array): image
+
+    Returns:
+        tensor: data ready for network
+    """
+    height_orig = img.shape[0]
+    width_orig = img.shape[1]
+
+    if width_orig > height_orig:
+        scale = width_orig / 384
+    else:
+        scale = height_orig / 384
+
+    height = (np.ceil(height_orig / scale / 32) * 32).astype(int)
+    width = (np.ceil(width_orig / scale / 32) * 32).astype(int)
+
+    img_resized = cv2.resize(img, (width, height), interpolation=cv2.INTER_AREA)
+
+    img_resized = (
+        torch.from_numpy(np.transpose(img_resized, (2, 0, 1))).contiguous().float()
+    )
+    img_resized = img_resized.unsqueeze(0)
+
+    return img_resized
+
+
+def resize_depth(depth, width, height):
+    """Resize depth map and bring to CPU (numpy).
+
+    Args:
+        depth (tensor): depth
+        width (int): image width
+        height (int): image height
+
+    Returns:
+        array: processed depth
+    """
+    depth = torch.squeeze(depth[0, :, :, :]).to("cpu")
+
+    depth_resized = cv2.resize(
+        depth.numpy(), (width, height), interpolation=cv2.INTER_CUBIC
+    )
+
+    return depth_resized
+
+def write_depth(path, depth, bits=1):
+    """Write depth map to pfm and png file.
+
+    Args:
+        path (str): filepath without extension
+        depth (array): depth
+    """
+    write_pfm(path + ".pfm", depth.astype(np.float32))
+
+    depth_min = depth.min()
+    depth_max = depth.max()
+
+    max_val = (2**(8*bits))-1
+
+    if depth_max - depth_min > np.finfo("float").eps:
+        out = max_val * (depth - depth_min) / (depth_max - depth_min)
+    else:
+        out = np.zeros(depth.shape, dtype=depth.type)
+
+    if bits == 1:
+        cv2.imwrite(path + ".png", out.astype("uint8"))
+    elif bits == 2:
+        cv2.imwrite(path + ".png", out.astype("uint16"))
+
+    return

lavis/common/annotator/mlsd/__init__.py

@@ -0,0 +1,39 @@
+import cv2
+import numpy as np
+import torch
+import os
+
+from einops import rearrange
+from .models.mbv2_mlsd_tiny import MobileV2_MLSD_Tiny
+from .models.mbv2_mlsd_large import MobileV2_MLSD_Large
+from .utils import pred_lines
+
+from annotator.util import annotator_ckpts_path
+
+
+remote_model_path = "https://huggingface.co/lllyasviel/ControlNet/resolve/main/annotator/ckpts/mlsd_large_512_fp32.pth"
+
+
+class MLSDdetector:
+    def __init__(self):
+        model_path = os.path.join(annotator_ckpts_path, "mlsd_large_512_fp32.pth")
+        if not os.path.exists(model_path):
+            from basicsr.utils.download_util import load_file_from_url
+            load_file_from_url(remote_model_path, model_dir=annotator_ckpts_path)
+        model = MobileV2_MLSD_Large()
+        model.load_state_dict(torch.load(model_path), strict=True)
+        self.model = model.cuda().eval()
+
+    def __call__(self, input_image, thr_v, thr_d):
+        assert input_image.ndim == 3
+        img = input_image
+        img_output = np.zeros_like(img)
+        try:
+            with torch.no_grad():
+                lines = pred_lines(img, self.model, [img.shape[0], img.shape[1]], thr_v, thr_d)
+                for line in lines:
+                    x_start, y_start, x_end, y_end = [int(val) for val in line]
+                    cv2.line(img_output, (x_start, y_start), (x_end, y_end), [255, 255, 255], 1)
+        except Exception as e:
+            pass
+        return img_output[:, :, 0]

lavis/common/annotator/mlsd/models/mbv2_mlsd_large.py

@@ -0,0 +1,292 @@
+import os
+import sys
+import torch
+import torch.nn as nn
+import torch.utils.model_zoo as model_zoo
+from  torch.nn import  functional as F
+
+
+class BlockTypeA(nn.Module):
+    def __init__(self, in_c1, in_c2, out_c1, out_c2, upscale = True):
+        super(BlockTypeA, self).__init__()
+        self.conv1 = nn.Sequential(
+            nn.Conv2d(in_c2, out_c2, kernel_size=1),
+            nn.BatchNorm2d(out_c2),
+            nn.ReLU(inplace=True)
+        )
+        self.conv2 = nn.Sequential(
+            nn.Conv2d(in_c1, out_c1, kernel_size=1),
+            nn.BatchNorm2d(out_c1),
+            nn.ReLU(inplace=True)
+        )
+        self.upscale = upscale
+
+    def forward(self, a, b):
+        b = self.conv1(b)
+        a = self.conv2(a)
+        if self.upscale:
+             b = F.interpolate(b, scale_factor=2.0, mode='bilinear', align_corners=True)
+        return torch.cat((a, b), dim=1)
+
+
+class BlockTypeB(nn.Module):
+    def __init__(self, in_c, out_c):
+        super(BlockTypeB, self).__init__()
+        self.conv1 = nn.Sequential(
+            nn.Conv2d(in_c, in_c,  kernel_size=3, padding=1),
+            nn.BatchNorm2d(in_c),
+            nn.ReLU()
+        )
+        self.conv2 = nn.Sequential(
+            nn.Conv2d(in_c, out_c, kernel_size=3, padding=1),
+            nn.BatchNorm2d(out_c),
+            nn.ReLU()
+        )
+
+    def forward(self, x):
+        x = self.conv1(x) + x
+        x = self.conv2(x)
+        return x
+
+class BlockTypeC(nn.Module):
+    def __init__(self, in_c, out_c):
+        super(BlockTypeC, self).__init__()
+        self.conv1 = nn.Sequential(
+            nn.Conv2d(in_c, in_c,  kernel_size=3, padding=5, dilation=5),
+            nn.BatchNorm2d(in_c),
+            nn.ReLU()
+        )
+        self.conv2 = nn.Sequential(
+            nn.Conv2d(in_c, in_c,  kernel_size=3, padding=1),
+            nn.BatchNorm2d(in_c),
+            nn.ReLU()
+        )
+        self.conv3 = nn.Conv2d(in_c, out_c, kernel_size=1)
+
+    def forward(self, x):
+        x = self.conv1(x)
+        x = self.conv2(x)
+        x = self.conv3(x)
+        return x
+
+def _make_divisible(v, divisor, min_value=None):
+    """
+    This function is taken from the original tf repo.
+    It ensures that all layers have a channel number that is divisible by 8
+    It can be seen here:
+    https://github.com/tensorflow/models/blob/master/research/slim/nets/mobilenet/mobilenet.py
+    :param v:
+    :param divisor:
+    :param min_value:
+    :return:
+    """
+    if min_value is None:
+        min_value = divisor
+    new_v = max(min_value, int(v + divisor / 2) // divisor * divisor)
+    # Make sure that round down does not go down by more than 10%.
+    if new_v < 0.9 * v:
+        new_v += divisor
+    return new_v
+
+
+class ConvBNReLU(nn.Sequential):
+    def __init__(self, in_planes, out_planes, kernel_size=3, stride=1, groups=1):
+        self.channel_pad = out_planes - in_planes
+        self.stride = stride
+        #padding = (kernel_size - 1) // 2
+
+        # TFLite uses slightly different padding than PyTorch
+        if stride == 2:
+            padding = 0
+        else:
+            padding = (kernel_size - 1) // 2
+
+        super(ConvBNReLU, self).__init__(
+            nn.Conv2d(in_planes, out_planes, kernel_size, stride, padding, groups=groups, bias=False),
+            nn.BatchNorm2d(out_planes),
+            nn.ReLU6(inplace=True)
+        )
+        self.max_pool = nn.MaxPool2d(kernel_size=stride, stride=stride)
+
+
+    def forward(self, x):
+        # TFLite uses  different padding
+        if self.stride == 2:
+            x = F.pad(x, (0, 1, 0, 1), "constant", 0)
+            #print(x.shape)
+
+        for module in self:
+            if not isinstance(module, nn.MaxPool2d):
+                x = module(x)
+        return x
+
+
+class InvertedResidual(nn.Module):
+    def __init__(self, inp, oup, stride, expand_ratio):
+        super(InvertedResidual, self).__init__()
+        self.stride = stride
+        assert stride in [1, 2]
+
+        hidden_dim = int(round(inp * expand_ratio))
+        self.use_res_connect = self.stride == 1 and inp == oup
+
+        layers = []
+        if expand_ratio != 1:
+            # pw
+            layers.append(ConvBNReLU(inp, hidden_dim, kernel_size=1))
+        layers.extend([
+            # dw
+            ConvBNReLU(hidden_dim, hidden_dim, stride=stride, groups=hidden_dim),
+            # pw-linear
+            nn.Conv2d(hidden_dim, oup, 1, 1, 0, bias=False),
+            nn.BatchNorm2d(oup),
+        ])
+        self.conv = nn.Sequential(*layers)
+
+    def forward(self, x):
+        if self.use_res_connect:
+            return x + self.conv(x)
+        else:
+            return self.conv(x)
+
+
+class MobileNetV2(nn.Module):
+    def __init__(self, pretrained=True):
+        """
+        MobileNet V2 main class
+        Args:
+            num_classes (int): Number of classes
+            width_mult (float): Width multiplier - adjusts number of channels in each layer by this amount
+            inverted_residual_setting: Network structure
+            round_nearest (int): Round the number of channels in each layer to be a multiple of this number
+            Set to 1 to turn off rounding
+            block: Module specifying inverted residual building block for mobilenet
+        """
+        super(MobileNetV2, self).__init__()
+
+        block = InvertedResidual
+        input_channel = 32
+        last_channel = 1280
+        width_mult = 1.0
+        round_nearest = 8
+
+        inverted_residual_setting = [
+            # t, c, n, s
+            [1, 16, 1, 1],
+            [6, 24, 2, 2],
+            [6, 32, 3, 2],
+            [6, 64, 4, 2],
+            [6, 96, 3, 1],
+            #[6, 160, 3, 2],
+            #[6, 320, 1, 1],
+        ]
+
+        # only check the first element, assuming user knows t,c,n,s are required
+        if len(inverted_residual_setting) == 0 or len(inverted_residual_setting[0]) != 4:
+            raise ValueError("inverted_residual_setting should be non-empty "
+                             "or a 4-element list, got {}".format(inverted_residual_setting))
+
+        # building first layer
+        input_channel = _make_divisible(input_channel * width_mult, round_nearest)
+        self.last_channel = _make_divisible(last_channel * max(1.0, width_mult), round_nearest)
+        features = [ConvBNReLU(4, input_channel, stride=2)]
+        # building inverted residual blocks
+        for t, c, n, s in inverted_residual_setting:
+            output_channel = _make_divisible(c * width_mult, round_nearest)
+            for i in range(n):
+                stride = s if i == 0 else 1
+                features.append(block(input_channel, output_channel, stride, expand_ratio=t))
+                input_channel = output_channel
+
+        self.features = nn.Sequential(*features)
+        self.fpn_selected = [1, 3, 6, 10, 13]
+        # weight initialization
+        for m in self.modules():
+            if isinstance(m, nn.Conv2d):
+                nn.init.kaiming_normal_(m.weight, mode='fan_out')
+                if m.bias is not None:
+                    nn.init.zeros_(m.bias)
+            elif isinstance(m, nn.BatchNorm2d):
+                nn.init.ones_(m.weight)
+                nn.init.zeros_(m.bias)
+            elif isinstance(m, nn.Linear):
+                nn.init.normal_(m.weight, 0, 0.01)
+                nn.init.zeros_(m.bias)
+        if pretrained:
+           self._load_pretrained_model()
+
+    def _forward_impl(self, x):
+        # This exists since TorchScript doesn't support inheritance, so the superclass method
+        # (this one) needs to have a name other than `forward` that can be accessed in a subclass
+        fpn_features = []
+        for i, f in enumerate(self.features):
+            if i > self.fpn_selected[-1]:
+                break
+            x = f(x)
+            if i in self.fpn_selected:
+                fpn_features.append(x)
+
+        c1, c2, c3, c4, c5 = fpn_features
+        return c1, c2, c3, c4, c5
+
+
+    def forward(self, x):
+        return self._forward_impl(x)
+
+    def _load_pretrained_model(self):
+        pretrain_dict = model_zoo.load_url('https://download.pytorch.org/models/mobilenet_v2-b0353104.pth')
+        model_dict = {}
+        state_dict = self.state_dict()
+        for k, v in pretrain_dict.items():
+            if k in state_dict:
+                model_dict[k] = v
+        state_dict.update(model_dict)
+        self.load_state_dict(state_dict)
+
+
+class MobileV2_MLSD_Large(nn.Module):
+    def __init__(self):
+        super(MobileV2_MLSD_Large, self).__init__()
+
+        self.backbone = MobileNetV2(pretrained=False)
+        ## A, B
+        self.block15 = BlockTypeA(in_c1= 64, in_c2= 96,
+                                  out_c1= 64, out_c2=64,
+                                  upscale=False)
+        self.block16 = BlockTypeB(128, 64)
+
+        ## A, B
+        self.block17 = BlockTypeA(in_c1 = 32,  in_c2 = 64,
+                                  out_c1= 64,  out_c2= 64)
+        self.block18 = BlockTypeB(128, 64)
+
+        ## A, B
+        self.block19 = BlockTypeA(in_c1=24, in_c2=64,
+                                  out_c1=64, out_c2=64)
+        self.block20 = BlockTypeB(128, 64)
+
+        ## A, B, C
+        self.block21 = BlockTypeA(in_c1=16, in_c2=64,
+                                  out_c1=64, out_c2=64)
+        self.block22 = BlockTypeB(128, 64)
+
+        self.block23 = BlockTypeC(64, 16)
+
+    def forward(self, x):
+        c1, c2, c3, c4, c5 = self.backbone(x)
+
+        x = self.block15(c4, c5)
+        x = self.block16(x)
+
+        x = self.block17(c3, x)
+        x = self.block18(x)
+
+        x = self.block19(c2, x)
+        x = self.block20(x)
+
+        x = self.block21(c1, x)
+        x = self.block22(x)
+        x = self.block23(x)
+        x = x[:, 7:, :, :]
+
+        return x

lavis/common/annotator/mlsd/models/mbv2_mlsd_tiny.py

@@ -0,0 +1,275 @@
+import os
+import sys
+import torch
+import torch.nn as nn
+import torch.utils.model_zoo as model_zoo
+from  torch.nn import  functional as F
+
+
+class BlockTypeA(nn.Module):
+    def __init__(self, in_c1, in_c2, out_c1, out_c2, upscale = True):
+        super(BlockTypeA, self).__init__()
+        self.conv1 = nn.Sequential(
+            nn.Conv2d(in_c2, out_c2, kernel_size=1),
+            nn.BatchNorm2d(out_c2),
+            nn.ReLU(inplace=True)
+        )
+        self.conv2 = nn.Sequential(
+            nn.Conv2d(in_c1, out_c1, kernel_size=1),
+            nn.BatchNorm2d(out_c1),
+            nn.ReLU(inplace=True)
+        )
+        self.upscale = upscale
+
+    def forward(self, a, b):
+        b = self.conv1(b)
+        a = self.conv2(a)
+        b = F.interpolate(b, scale_factor=2.0, mode='bilinear', align_corners=True)
+        return torch.cat((a, b), dim=1)
+
+
+class BlockTypeB(nn.Module):
+    def __init__(self, in_c, out_c):
+        super(BlockTypeB, self).__init__()
+        self.conv1 = nn.Sequential(
+            nn.Conv2d(in_c, in_c,  kernel_size=3, padding=1),
+            nn.BatchNorm2d(in_c),
+            nn.ReLU()
+        )
+        self.conv2 = nn.Sequential(
+            nn.Conv2d(in_c, out_c, kernel_size=3, padding=1),
+            nn.BatchNorm2d(out_c),
+            nn.ReLU()
+        )
+
+    def forward(self, x):
+        x = self.conv1(x) + x
+        x = self.conv2(x)
+        return x
+
+class BlockTypeC(nn.Module):
+    def __init__(self, in_c, out_c):
+        super(BlockTypeC, self).__init__()
+        self.conv1 = nn.Sequential(
+            nn.Conv2d(in_c, in_c,  kernel_size=3, padding=5, dilation=5),
+            nn.BatchNorm2d(in_c),
+            nn.ReLU()
+        )
+        self.conv2 = nn.Sequential(
+            nn.Conv2d(in_c, in_c,  kernel_size=3, padding=1),
+            nn.BatchNorm2d(in_c),
+            nn.ReLU()
+        )
+        self.conv3 = nn.Conv2d(in_c, out_c, kernel_size=1)
+
+    def forward(self, x):
+        x = self.conv1(x)
+        x = self.conv2(x)
+        x = self.conv3(x)
+        return x
+
+def _make_divisible(v, divisor, min_value=None):
+    """
+    This function is taken from the original tf repo.
+    It ensures that all layers have a channel number that is divisible by 8
+    It can be seen here:
+    https://github.com/tensorflow/models/blob/master/research/slim/nets/mobilenet/mobilenet.py
+    :param v:
+    :param divisor:
+    :param min_value:
+    :return:
+    """
+    if min_value is None:
+        min_value = divisor
+    new_v = max(min_value, int(v + divisor / 2) // divisor * divisor)
+    # Make sure that round down does not go down by more than 10%.
+    if new_v < 0.9 * v:
+        new_v += divisor
+    return new_v
+
+
+class ConvBNReLU(nn.Sequential):
+    def __init__(self, in_planes, out_planes, kernel_size=3, stride=1, groups=1):
+        self.channel_pad = out_planes - in_planes
+        self.stride = stride
+        #padding = (kernel_size - 1) // 2
+
+        # TFLite uses slightly different padding than PyTorch
+        if stride == 2:
+            padding = 0
+        else:
+            padding = (kernel_size - 1) // 2
+
+        super(ConvBNReLU, self).__init__(
+            nn.Conv2d(in_planes, out_planes, kernel_size, stride, padding, groups=groups, bias=False),
+            nn.BatchNorm2d(out_planes),
+            nn.ReLU6(inplace=True)
+        )
+        self.max_pool = nn.MaxPool2d(kernel_size=stride, stride=stride)
+
+
+    def forward(self, x):
+        # TFLite uses  different padding
+        if self.stride == 2:
+            x = F.pad(x, (0, 1, 0, 1), "constant", 0)
+            #print(x.shape)
+
+        for module in self:
+            if not isinstance(module, nn.MaxPool2d):
+                x = module(x)
+        return x
+
+
+class InvertedResidual(nn.Module):
+    def __init__(self, inp, oup, stride, expand_ratio):
+        super(InvertedResidual, self).__init__()
+        self.stride = stride
+        assert stride in [1, 2]
+
+        hidden_dim = int(round(inp * expand_ratio))
+        self.use_res_connect = self.stride == 1 and inp == oup
+
+        layers = []
+        if expand_ratio != 1:
+            # pw
+            layers.append(ConvBNReLU(inp, hidden_dim, kernel_size=1))
+        layers.extend([
+            # dw
+            ConvBNReLU(hidden_dim, hidden_dim, stride=stride, groups=hidden_dim),
+            # pw-linear
+            nn.Conv2d(hidden_dim, oup, 1, 1, 0, bias=False),
+            nn.BatchNorm2d(oup),
+        ])
+        self.conv = nn.Sequential(*layers)
+
+    def forward(self, x):
+        if self.use_res_connect:
+            return x + self.conv(x)
+        else:
+            return self.conv(x)
+
+
+class MobileNetV2(nn.Module):
+    def __init__(self, pretrained=True):
+        """
+        MobileNet V2 main class
+        Args:
+            num_classes (int): Number of classes
+            width_mult (float): Width multiplier - adjusts number of channels in each layer by this amount
+            inverted_residual_setting: Network structure
+            round_nearest (int): Round the number of channels in each layer to be a multiple of this number
+            Set to 1 to turn off rounding
+            block: Module specifying inverted residual building block for mobilenet
+        """
+        super(MobileNetV2, self).__init__()
+
+        block = InvertedResidual
+        input_channel = 32
+        last_channel = 1280
+        width_mult = 1.0
+        round_nearest = 8
+
+        inverted_residual_setting = [
+            # t, c, n, s
+            [1, 16, 1, 1],
+            [6, 24, 2, 2],
+            [6, 32, 3, 2],
+            [6, 64, 4, 2],
+            #[6, 96, 3, 1],
+            #[6, 160, 3, 2],
+            #[6, 320, 1, 1],
+        ]
+
+        # only check the first element, assuming user knows t,c,n,s are required
+        if len(inverted_residual_setting) == 0 or len(inverted_residual_setting[0]) != 4:
+            raise ValueError("inverted_residual_setting should be non-empty "
+                             "or a 4-element list, got {}".format(inverted_residual_setting))
+
+        # building first layer
+        input_channel = _make_divisible(input_channel * width_mult, round_nearest)
+        self.last_channel = _make_divisible(last_channel * max(1.0, width_mult), round_nearest)
+        features = [ConvBNReLU(4, input_channel, stride=2)]
+        # building inverted residual blocks
+        for t, c, n, s in inverted_residual_setting:
+            output_channel = _make_divisible(c * width_mult, round_nearest)
+            for i in range(n):
+                stride = s if i == 0 else 1
+                features.append(block(input_channel, output_channel, stride, expand_ratio=t))
+                input_channel = output_channel
+        self.features = nn.Sequential(*features)
+
+        self.fpn_selected = [3, 6, 10]
+        # weight initialization
+        for m in self.modules():
+            if isinstance(m, nn.Conv2d):
+                nn.init.kaiming_normal_(m.weight, mode='fan_out')
+                if m.bias is not None:
+                    nn.init.zeros_(m.bias)
+            elif isinstance(m, nn.BatchNorm2d):
+                nn.init.ones_(m.weight)
+                nn.init.zeros_(m.bias)
+            elif isinstance(m, nn.Linear):
+                nn.init.normal_(m.weight, 0, 0.01)
+                nn.init.zeros_(m.bias)
+
+        #if pretrained:
+        #    self._load_pretrained_model()
+
+    def _forward_impl(self, x):
+        # This exists since TorchScript doesn't support inheritance, so the superclass method
+        # (this one) needs to have a name other than `forward` that can be accessed in a subclass
+        fpn_features = []
+        for i, f in enumerate(self.features):
+            if i > self.fpn_selected[-1]:
+                break
+            x = f(x)
+            if i in self.fpn_selected:
+                fpn_features.append(x)
+
+        c2, c3, c4 = fpn_features
+        return c2, c3, c4
+
+
+    def forward(self, x):
+        return self._forward_impl(x)
+
+    def _load_pretrained_model(self):
+        pretrain_dict = model_zoo.load_url('https://download.pytorch.org/models/mobilenet_v2-b0353104.pth')
+        model_dict = {}
+        state_dict = self.state_dict()
+        for k, v in pretrain_dict.items():
+            if k in state_dict:
+                model_dict[k] = v
+        state_dict.update(model_dict)
+        self.load_state_dict(state_dict)
+
+
+class MobileV2_MLSD_Tiny(nn.Module):
+    def __init__(self):
+        super(MobileV2_MLSD_Tiny, self).__init__()
+
+        self.backbone = MobileNetV2(pretrained=True)
+
+        self.block12 = BlockTypeA(in_c1= 32, in_c2= 64,
+                                  out_c1= 64, out_c2=64)
+        self.block13 = BlockTypeB(128, 64)
+
+        self.block14 = BlockTypeA(in_c1 = 24,  in_c2 = 64,
+                                  out_c1= 32,  out_c2= 32)
+        self.block15 = BlockTypeB(64, 64)
+
+        self.block16 = BlockTypeC(64, 16)
+
+    def forward(self, x):
+        c2, c3, c4 = self.backbone(x)
+
+        x = self.block12(c3, c4)
+        x = self.block13(x)
+        x = self.block14(c2, x)
+        x = self.block15(x)
+        x = self.block16(x)
+        x = x[:, 7:, :, :]
+        #print(x.shape)
+        x = F.interpolate(x, scale_factor=2.0, mode='bilinear', align_corners=True)
+
+        return x

lavis/common/annotator/mlsd/utils.py

@@ -0,0 +1,580 @@
+'''
+modified by  lihaoweicv
+pytorch version
+'''
+
+'''
+M-LSD
+Copyright 2021-present NAVER Corp.
+Apache License v2.0
+'''
+
+import os
+import numpy as np
+import cv2
+import torch
+from  torch.nn import  functional as F
+
+
+def deccode_output_score_and_ptss(tpMap, topk_n = 200, ksize = 5):
+    '''
+    tpMap:
+    center: tpMap[1, 0, :, :]
+    displacement: tpMap[1, 1:5, :, :]
+    '''
+    b, c, h, w = tpMap.shape
+    assert  b==1, 'only support bsize==1'
+    displacement = tpMap[:, 1:5, :, :][0]
+    center = tpMap[:, 0, :, :]
+    heat = torch.sigmoid(center)
+    hmax = F.max_pool2d( heat, (ksize, ksize), stride=1, padding=(ksize-1)//2)
+    keep = (hmax == heat).float()
+    heat = heat * keep
+    heat = heat.reshape(-1, )
+
+    scores, indices = torch.topk(heat, topk_n, dim=-1, largest=True)
+    yy = torch.floor_divide(indices, w).unsqueeze(-1)
+    xx = torch.fmod(indices, w).unsqueeze(-1)
+    ptss = torch.cat((yy, xx),dim=-1)
+
+    ptss   = ptss.detach().cpu().numpy()
+    scores = scores.detach().cpu().numpy()
+    displacement = displacement.detach().cpu().numpy()
+    displacement = displacement.transpose((1,2,0))
+    return  ptss, scores, displacement
+
+
+def pred_lines(image, model,
+               input_shape=[512, 512],
+               score_thr=0.10,
+               dist_thr=20.0):
+    h, w, _ = image.shape
+    h_ratio, w_ratio = [h / input_shape[0], w / input_shape[1]]
+
+    resized_image = np.concatenate([cv2.resize(image, (input_shape[1], input_shape[0]), interpolation=cv2.INTER_AREA),
+                                    np.ones([input_shape[0], input_shape[1], 1])], axis=-1)
+
+    resized_image = resized_image.transpose((2,0,1))
+    batch_image = np.expand_dims(resized_image, axis=0).astype('float32')
+    batch_image = (batch_image / 127.5) - 1.0
+
+    batch_image = torch.from_numpy(batch_image).float().cuda()
+    outputs = model(batch_image)
+    pts, pts_score, vmap = deccode_output_score_and_ptss(outputs, 200, 3)
+    start = vmap[:, :, :2]
+    end = vmap[:, :, 2:]
+    dist_map = np.sqrt(np.sum((start - end) ** 2, axis=-1))
+
+    segments_list = []
+    for center, score in zip(pts, pts_score):
+        y, x = center
+        distance = dist_map[y, x]
+        if score > score_thr and distance > dist_thr:
+            disp_x_start, disp_y_start, disp_x_end, disp_y_end = vmap[y, x, :]
+            x_start = x + disp_x_start
+            y_start = y + disp_y_start
+            x_end = x + disp_x_end
+            y_end = y + disp_y_end
+            segments_list.append([x_start, y_start, x_end, y_end])
+
+    lines = 2 * np.array(segments_list)  # 256 > 512
+    lines[:, 0] = lines[:, 0] * w_ratio
+    lines[:, 1] = lines[:, 1] * h_ratio
+    lines[:, 2] = lines[:, 2] * w_ratio
+    lines[:, 3] = lines[:, 3] * h_ratio
+
+    return lines
+
+
+def pred_squares(image,
+                 model,
+                 input_shape=[512, 512],
+                 params={'score': 0.06,
+                         'outside_ratio': 0.28,
+                         'inside_ratio': 0.45,
+                         'w_overlap': 0.0,
+                         'w_degree': 1.95,
+                         'w_length': 0.0,
+                         'w_area': 1.86,
+                         'w_center': 0.14}):
+    '''
+    shape = [height, width]
+    '''
+    h, w, _ = image.shape
+    original_shape = [h, w]
+
+    resized_image = np.concatenate([cv2.resize(image, (input_shape[0], input_shape[1]), interpolation=cv2.INTER_AREA),
+                                    np.ones([input_shape[0], input_shape[1], 1])], axis=-1)
+    resized_image = resized_image.transpose((2, 0, 1))
+    batch_image = np.expand_dims(resized_image, axis=0).astype('float32')
+    batch_image = (batch_image / 127.5) - 1.0
+
+    batch_image = torch.from_numpy(batch_image).float().cuda()
+    outputs = model(batch_image)
+
+    pts, pts_score, vmap = deccode_output_score_and_ptss(outputs, 200, 3)
+    start = vmap[:, :, :2]  # (x, y)
+    end = vmap[:, :, 2:]  # (x, y)
+    dist_map = np.sqrt(np.sum((start - end) ** 2, axis=-1))
+
+    junc_list = []
+    segments_list = []
+    for junc, score in zip(pts, pts_score):
+        y, x = junc
+        distance = dist_map[y, x]
+        if score > params['score'] and distance > 20.0:
+            junc_list.append([x, y])
+            disp_x_start, disp_y_start, disp_x_end, disp_y_end = vmap[y, x, :]
+            d_arrow = 1.0
+            x_start = x + d_arrow * disp_x_start
+            y_start = y + d_arrow * disp_y_start
+            x_end = x + d_arrow * disp_x_end
+            y_end = y + d_arrow * disp_y_end
+            segments_list.append([x_start, y_start, x_end, y_end])
+
+    segments = np.array(segments_list)
+
+    ####### post processing for squares
+    # 1. get unique lines
+    point = np.array([[0, 0]])
+    point = point[0]
+    start = segments[:, :2]
+    end = segments[:, 2:]
+    diff = start - end
+    a = diff[:, 1]
+    b = -diff[:, 0]
+    c = a * start[:, 0] + b * start[:, 1]
+
+    d = np.abs(a * point[0] + b * point[1] - c) / np.sqrt(a ** 2 + b ** 2 + 1e-10)
+    theta = np.arctan2(diff[:, 0], diff[:, 1]) * 180 / np.pi
+    theta[theta < 0.0] += 180
+    hough = np.concatenate([d[:, None], theta[:, None]], axis=-1)
+
+    d_quant = 1
+    theta_quant = 2
+    hough[:, 0] //= d_quant
+    hough[:, 1] //= theta_quant
+    _, indices, counts = np.unique(hough, axis=0, return_index=True, return_counts=True)
+
+    acc_map = np.zeros([512 // d_quant + 1, 360 // theta_quant + 1], dtype='float32')
+    idx_map = np.zeros([512 // d_quant + 1, 360 // theta_quant + 1], dtype='int32') - 1
+    yx_indices = hough[indices, :].astype('int32')
+    acc_map[yx_indices[:, 0], yx_indices[:, 1]] = counts
+    idx_map[yx_indices[:, 0], yx_indices[:, 1]] = indices
+
+    acc_map_np = acc_map
+    # acc_map = acc_map[None, :, :, None]
+    #
+    # ### fast suppression using tensorflow op
+    # acc_map = tf.constant(acc_map, dtype=tf.float32)
+    # max_acc_map = tf.keras.layers.MaxPool2D(pool_size=(5, 5), strides=1, padding='same')(acc_map)
+    # acc_map = acc_map * tf.cast(tf.math.equal(acc_map, max_acc_map), tf.float32)
+    # flatten_acc_map = tf.reshape(acc_map, [1, -1])
+    # topk_values, topk_indices = tf.math.top_k(flatten_acc_map, k=len(pts))
+    # _, h, w, _ = acc_map.shape
+    # y = tf.expand_dims(topk_indices // w, axis=-1)
+    # x = tf.expand_dims(topk_indices % w, axis=-1)
+    # yx = tf.concat([y, x], axis=-1)
+
+    ### fast suppression using pytorch op
+    acc_map = torch.from_numpy(acc_map_np).unsqueeze(0).unsqueeze(0)
+    _,_, h, w = acc_map.shape
+    max_acc_map = F.max_pool2d(acc_map,kernel_size=5, stride=1, padding=2)
+    acc_map = acc_map * ( (acc_map == max_acc_map).float() )
+    flatten_acc_map = acc_map.reshape([-1, ])
+
+    scores, indices = torch.topk(flatten_acc_map, len(pts), dim=-1, largest=True)
+    yy = torch.div(indices, w, rounding_mode='floor').unsqueeze(-1)
+    xx = torch.fmod(indices, w).unsqueeze(-1)
+    yx = torch.cat((yy, xx), dim=-1)
+
+    yx = yx.detach().cpu().numpy()
+
+    topk_values = scores.detach().cpu().numpy()
+    indices = idx_map[yx[:, 0], yx[:, 1]]
+    basis = 5 // 2
+
+    merged_segments = []
+    for yx_pt, max_indice, value in zip(yx, indices, topk_values):
+        y, x = yx_pt
+        if max_indice == -1 or value == 0:
+            continue
+        segment_list = []
+        for y_offset in range(-basis, basis + 1):
+            for x_offset in range(-basis, basis + 1):
+                indice = idx_map[y + y_offset, x + x_offset]
+                cnt = int(acc_map_np[y + y_offset, x + x_offset])
+                if indice != -1:
+                    segment_list.append(segments[indice])
+                if cnt > 1:
+                    check_cnt = 1
+                    current_hough = hough[indice]
+                    for new_indice, new_hough in enumerate(hough):
+                        if (current_hough == new_hough).all() and indice != new_indice:
+                            segment_list.append(segments[new_indice])
+                            check_cnt += 1
+                        if check_cnt == cnt:
+                            break
+        group_segments = np.array(segment_list).reshape([-1, 2])
+        sorted_group_segments = np.sort(group_segments, axis=0)
+        x_min, y_min = sorted_group_segments[0, :]
+        x_max, y_max = sorted_group_segments[-1, :]
+
+        deg = theta[max_indice]
+        if deg >= 90:
+            merged_segments.append([x_min, y_max, x_max, y_min])
+        else:
+            merged_segments.append([x_min, y_min, x_max, y_max])
+
+    # 2. get intersections
+    new_segments = np.array(merged_segments)  # (x1, y1, x2, y2)
+    start = new_segments[:, :2]  # (x1, y1)
+    end = new_segments[:, 2:]  # (x2, y2)
+    new_centers = (start + end) / 2.0
+    diff = start - end
+    dist_segments = np.sqrt(np.sum(diff ** 2, axis=-1))
+
+    # ax + by = c
+    a = diff[:, 1]
+    b = -diff[:, 0]
+    c = a * start[:, 0] + b * start[:, 1]
+    pre_det = a[:, None] * b[None, :]
+    det = pre_det - np.transpose(pre_det)
+
+    pre_inter_y = a[:, None] * c[None, :]
+    inter_y = (pre_inter_y - np.transpose(pre_inter_y)) / (det + 1e-10)
+    pre_inter_x = c[:, None] * b[None, :]
+    inter_x = (pre_inter_x - np.transpose(pre_inter_x)) / (det + 1e-10)
+    inter_pts = np.concatenate([inter_x[:, :, None], inter_y[:, :, None]], axis=-1).astype('int32')
+
+    # 3. get corner information
+    # 3.1 get distance
+    '''
+    dist_segments:
+        | dist(0), dist(1), dist(2), ...|
+    dist_inter_to_segment1:
+        | dist(inter,0), dist(inter,0), dist(inter,0), ... |
+        | dist(inter,1), dist(inter,1), dist(inter,1), ... |
+        ...
+    dist_inter_to_semgnet2:
+        | dist(inter,0), dist(inter,1), dist(inter,2), ... |
+        | dist(inter,0), dist(inter,1), dist(inter,2), ... |
+        ...
+    '''
+
+    dist_inter_to_segment1_start = np.sqrt(
+        np.sum(((inter_pts - start[:, None, :]) ** 2), axis=-1, keepdims=True))  # [n_batch, n_batch, 1]
+    dist_inter_to_segment1_end = np.sqrt(
+        np.sum(((inter_pts - end[:, None, :]) ** 2), axis=-1, keepdims=True))  # [n_batch, n_batch, 1]
+    dist_inter_to_segment2_start = np.sqrt(
+        np.sum(((inter_pts - start[None, :, :]) ** 2), axis=-1, keepdims=True))  # [n_batch, n_batch, 1]
+    dist_inter_to_segment2_end = np.sqrt(
+        np.sum(((inter_pts - end[None, :, :]) ** 2), axis=-1, keepdims=True))  # [n_batch, n_batch, 1]
+
+    # sort ascending
+    dist_inter_to_segment1 = np.sort(
+        np.concatenate([dist_inter_to_segment1_start, dist_inter_to_segment1_end], axis=-1),
+        axis=-1)  # [n_batch, n_batch, 2]
+    dist_inter_to_segment2 = np.sort(
+        np.concatenate([dist_inter_to_segment2_start, dist_inter_to_segment2_end], axis=-1),
+        axis=-1)  # [n_batch, n_batch, 2]
+
+    # 3.2 get degree
+    inter_to_start = new_centers[:, None, :] - inter_pts
+    deg_inter_to_start = np.arctan2(inter_to_start[:, :, 1], inter_to_start[:, :, 0]) * 180 / np.pi
+    deg_inter_to_start[deg_inter_to_start < 0.0] += 360
+    inter_to_end = new_centers[None, :, :] - inter_pts
+    deg_inter_to_end = np.arctan2(inter_to_end[:, :, 1], inter_to_end[:, :, 0]) * 180 / np.pi
+    deg_inter_to_end[deg_inter_to_end < 0.0] += 360
+
+    '''
+    B -- G
+    |    |
+    C -- R
+    B : blue / G: green / C: cyan / R: red
+
+    0 -- 1
+    |    |
+    3 -- 2
+    '''
+    # rename variables
+    deg1_map, deg2_map = deg_inter_to_start, deg_inter_to_end
+    # sort deg ascending
+    deg_sort = np.sort(np.concatenate([deg1_map[:, :, None], deg2_map[:, :, None]], axis=-1), axis=-1)
+
+    deg_diff_map = np.abs(deg1_map - deg2_map)
+    # we only consider the smallest degree of intersect
+    deg_diff_map[deg_diff_map > 180] = 360 - deg_diff_map[deg_diff_map > 180]
+
+    # define available degree range
+    deg_range = [60, 120]
+
+    corner_dict = {corner_info: [] for corner_info in range(4)}
+    inter_points = []
+    for i in range(inter_pts.shape[0]):
+        for j in range(i + 1, inter_pts.shape[1]):
+            # i, j > line index, always i < j
+            x, y = inter_pts[i, j, :]
+            deg1, deg2 = deg_sort[i, j, :]
+            deg_diff = deg_diff_map[i, j]
+
+            check_degree = deg_diff > deg_range[0] and deg_diff < deg_range[1]
+
+            outside_ratio = params['outside_ratio']  # over ratio >>> drop it!
+            inside_ratio = params['inside_ratio']  # over ratio >>> drop it!
+            check_distance = ((dist_inter_to_segment1[i, j, 1] >= dist_segments[i] and \
+                               dist_inter_to_segment1[i, j, 0] <= dist_segments[i] * outside_ratio) or \
+                              (dist_inter_to_segment1[i, j, 1] <= dist_segments[i] and \
+                               dist_inter_to_segment1[i, j, 0] <= dist_segments[i] * inside_ratio)) and \
+                             ((dist_inter_to_segment2[i, j, 1] >= dist_segments[j] and \
+                               dist_inter_to_segment2[i, j, 0] <= dist_segments[j] * outside_ratio) or \
+                              (dist_inter_to_segment2[i, j, 1] <= dist_segments[j] and \
+                               dist_inter_to_segment2[i, j, 0] <= dist_segments[j] * inside_ratio))
+
+            if check_degree and check_distance:
+                corner_info = None
+
+                if (deg1 >= 0 and deg1 <= 45 and deg2 >= 45 and deg2 <= 120) or \
+                        (deg2 >= 315 and deg1 >= 45 and deg1 <= 120):
+                    corner_info, color_info = 0, 'blue'
+                elif (deg1 >= 45 and deg1 <= 125 and deg2 >= 125 and deg2 <= 225):
+                    corner_info, color_info = 1, 'green'
+                elif (deg1 >= 125 and deg1 <= 225 and deg2 >= 225 and deg2 <= 315):
+                    corner_info, color_info = 2, 'black'
+                elif (deg1 >= 0 and deg1 <= 45 and deg2 >= 225 and deg2 <= 315) or \
+                        (deg2 >= 315 and deg1 >= 225 and deg1 <= 315):
+                    corner_info, color_info = 3, 'cyan'
+                else:
+                    corner_info, color_info = 4, 'red'  # we don't use it
+                    continue
+
+                corner_dict[corner_info].append([x, y, i, j])
+                inter_points.append([x, y])
+
+    square_list = []
+    connect_list = []
+    segments_list = []
+    for corner0 in corner_dict[0]:
+        for corner1 in corner_dict[1]:
+            connect01 = False
+            for corner0_line in corner0[2:]:
+                if corner0_line in corner1[2:]:
+                    connect01 = True
+                    break
+            if connect01:
+                for corner2 in corner_dict[2]:
+                    connect12 = False
+                    for corner1_line in corner1[2:]:
+                        if corner1_line in corner2[2:]:
+                            connect12 = True
+                            break
+                    if connect12:
+                        for corner3 in corner_dict[3]:
+                            connect23 = False
+                            for corner2_line in corner2[2:]:
+                                if corner2_line in corner3[2:]:
+                                    connect23 = True
+                                    break
+                            if connect23:
+                                for corner3_line in corner3[2:]:
+                                    if corner3_line in corner0[2:]:
+                                        # SQUARE!!!
+                                        '''
+                                        0 -- 1
+                                        |    |
+                                        3 -- 2
+                                        square_list:
+                                            order: 0 > 1 > 2 > 3
+                                            | x0, y0, x1, y1, x2, y2, x3, y3 |
+                                            | x0, y0, x1, y1, x2, y2, x3, y3 |
+                                            ...
+                                        connect_list:
+                                            order: 01 > 12 > 23 > 30
+                                            | line_idx01, line_idx12, line_idx23, line_idx30 |
+                                            | line_idx01, line_idx12, line_idx23, line_idx30 |
+                                            ...
+                                        segments_list:
+                                            order: 0 > 1 > 2 > 3
+                                            | line_idx0_i, line_idx0_j, line_idx1_i, line_idx1_j, line_idx2_i, line_idx2_j, line_idx3_i, line_idx3_j |
+                                            | line_idx0_i, line_idx0_j, line_idx1_i, line_idx1_j, line_idx2_i, line_idx2_j, line_idx3_i, line_idx3_j |
+                                            ...
+                                        '''
+                                        square_list.append(corner0[:2] + corner1[:2] + corner2[:2] + corner3[:2])
+                                        connect_list.append([corner0_line, corner1_line, corner2_line, corner3_line])
+                                        segments_list.append(corner0[2:] + corner1[2:] + corner2[2:] + corner3[2:])
+
+    def check_outside_inside(segments_info, connect_idx):
+        # return 'outside or inside', min distance, cover_param, peri_param
+        if connect_idx == segments_info[0]:
+            check_dist_mat = dist_inter_to_segment1
+        else:
+            check_dist_mat = dist_inter_to_segment2
+
+        i, j = segments_info
+        min_dist, max_dist = check_dist_mat[i, j, :]
+        connect_dist = dist_segments[connect_idx]
+        if max_dist > connect_dist:
+            return 'outside', min_dist, 0, 1
+        else:
+            return 'inside', min_dist, -1, -1
+
+    top_square = None
+
+    try:
+        map_size = input_shape[0] / 2
+        squares = np.array(square_list).reshape([-1, 4, 2])
+        score_array = []
+        connect_array = np.array(connect_list)
+        segments_array = np.array(segments_list).reshape([-1, 4, 2])
+
+        # get degree of corners:
+        squares_rollup = np.roll(squares, 1, axis=1)
+        squares_rolldown = np.roll(squares, -1, axis=1)
+        vec1 = squares_rollup - squares
+        normalized_vec1 = vec1 / (np.linalg.norm(vec1, axis=-1, keepdims=True) + 1e-10)
+        vec2 = squares_rolldown - squares
+        normalized_vec2 = vec2 / (np.linalg.norm(vec2, axis=-1, keepdims=True) + 1e-10)
+        inner_products = np.sum(normalized_vec1 * normalized_vec2, axis=-1)  # [n_squares, 4]
+        squares_degree = np.arccos(inner_products) * 180 / np.pi  # [n_squares, 4]
+
+        # get square score
+        overlap_scores = []
+        degree_scores = []
+        length_scores = []
+
+        for connects, segments, square, degree in zip(connect_array, segments_array, squares, squares_degree):
+            '''
+            0 -- 1
+            |    |
+            3 -- 2
+
+            # segments: [4, 2]
+            # connects: [4]
+            '''
+
+            ###################################### OVERLAP SCORES
+            cover = 0
+            perimeter = 0
+            # check 0 > 1 > 2 > 3
+            square_length = []
+
+            for start_idx in range(4):
+                end_idx = (start_idx + 1) % 4
+
+                connect_idx = connects[start_idx]  # segment idx of segment01
+                start_segments = segments[start_idx]
+                end_segments = segments[end_idx]
+
+                start_point = square[start_idx]
+                end_point = square[end_idx]
+
+                # check whether outside or inside
+                start_position, start_min, start_cover_param, start_peri_param = check_outside_inside(start_segments,
+                                                                                                      connect_idx)
+                end_position, end_min, end_cover_param, end_peri_param = check_outside_inside(end_segments, connect_idx)
+
+                cover += dist_segments[connect_idx] + start_cover_param * start_min + end_cover_param * end_min
+                perimeter += dist_segments[connect_idx] + start_peri_param * start_min + end_peri_param * end_min
+
+                square_length.append(
+                    dist_segments[connect_idx] + start_peri_param * start_min + end_peri_param * end_min)
+
+            overlap_scores.append(cover / perimeter)
+            ######################################
+            ###################################### DEGREE SCORES
+            '''
+            deg0 vs deg2
+            deg1 vs deg3
+            '''
+            deg0, deg1, deg2, deg3 = degree
+            deg_ratio1 = deg0 / deg2
+            if deg_ratio1 > 1.0:
+                deg_ratio1 = 1 / deg_ratio1
+            deg_ratio2 = deg1 / deg3
+            if deg_ratio2 > 1.0:
+                deg_ratio2 = 1 / deg_ratio2
+            degree_scores.append((deg_ratio1 + deg_ratio2) / 2)
+            ######################################
+            ###################################### LENGTH SCORES
+            '''
+            len0 vs len2
+            len1 vs len3
+            '''
+            len0, len1, len2, len3 = square_length
+            len_ratio1 = len0 / len2 if len2 > len0 else len2 / len0
+            len_ratio2 = len1 / len3 if len3 > len1 else len3 / len1
+            length_scores.append((len_ratio1 + len_ratio2) / 2)
+
+            ######################################
+
+        overlap_scores = np.array(overlap_scores)
+        overlap_scores /= np.max(overlap_scores)
+
+        degree_scores = np.array(degree_scores)
+        # degree_scores /= np.max(degree_scores)
+
+        length_scores = np.array(length_scores)
+
+        ###################################### AREA SCORES
+        area_scores = np.reshape(squares, [-1, 4, 2])
+        area_x = area_scores[:, :, 0]
+        area_y = area_scores[:, :, 1]
+        correction = area_x[:, -1] * area_y[:, 0] - area_y[:, -1] * area_x[:, 0]
+        area_scores = np.sum(area_x[:, :-1] * area_y[:, 1:], axis=-1) - np.sum(area_y[:, :-1] * area_x[:, 1:], axis=-1)
+        area_scores = 0.5 * np.abs(area_scores + correction)
+        area_scores /= (map_size * map_size)  # np.max(area_scores)
+        ######################################
+
+        ###################################### CENTER SCORES
+        centers = np.array([[256 // 2, 256 // 2]], dtype='float32')  # [1, 2]
+        # squares: [n, 4, 2]
+        square_centers = np.mean(squares, axis=1)  # [n, 2]
+        center2center = np.sqrt(np.sum((centers - square_centers) ** 2))
+        center_scores = center2center / (map_size / np.sqrt(2.0))
+
+        '''
+        score_w = [overlap, degree, area, center, length]
+        '''
+        score_w = [0.0, 1.0, 10.0, 0.5, 1.0]
+        score_array = params['w_overlap'] * overlap_scores \
+                      + params['w_degree'] * degree_scores \
+                      + params['w_area'] * area_scores \
+                      - params['w_center'] * center_scores \
+                      + params['w_length'] * length_scores
+
+        best_square = []
+
+        sorted_idx = np.argsort(score_array)[::-1]
+        score_array = score_array[sorted_idx]
+        squares = squares[sorted_idx]
+
+    except Exception as e:
+        pass
+
+    '''return list
+    merged_lines, squares, scores
+    '''
+
+    try:
+        new_segments[:, 0] = new_segments[:, 0] * 2 / input_shape[1] * original_shape[1]
+        new_segments[:, 1] = new_segments[:, 1] * 2 / input_shape[0] * original_shape[0]
+        new_segments[:, 2] = new_segments[:, 2] * 2 / input_shape[1] * original_shape[1]
+        new_segments[:, 3] = new_segments[:, 3] * 2 / input_shape[0] * original_shape[0]
+    except:
+        new_segments = []
+
+    try:
+        squares[:, :, 0] = squares[:, :, 0] * 2 / input_shape[1] * original_shape[1]
+        squares[:, :, 1] = squares[:, :, 1] * 2 / input_shape[0] * original_shape[0]
+    except:
+        squares = []
+        score_array = []
+
+    try:
+        inter_points = np.array(inter_points)
+        inter_points[:, 0] = inter_points[:, 0] * 2 / input_shape[1] * original_shape[1]
+        inter_points[:, 1] = inter_points[:, 1] * 2 / input_shape[0] * original_shape[0]
+    except:
+        inter_points = []
+
+    return new_segments, squares, score_array, inter_points

lavis/common/annotator/openpose/__init__.py

@@ -0,0 +1,44 @@
+import os
+os.environ["KMP_DUPLICATE_LIB_OK"]="TRUE"
+
+import torch
+import numpy as np
+from . import util
+from .body import Body
+from .hand import Hand
+from annotator.util import annotator_ckpts_path
+
+
+body_model_path = "https://huggingface.co/lllyasviel/ControlNet/resolve/main/annotator/ckpts/body_pose_model.pth"
+hand_model_path = "https://huggingface.co/lllyasviel/ControlNet/resolve/main/annotator/ckpts/hand_pose_model.pth"
+
+
+class OpenposeDetector:
+    def __init__(self):
+        body_modelpath = os.path.join(annotator_ckpts_path, "body_pose_model.pth")
+        hand_modelpath = os.path.join(annotator_ckpts_path, "hand_pose_model.pth")
+
+        if not os.path.exists(hand_modelpath):
+            from basicsr.utils.download_util import load_file_from_url
+            load_file_from_url(body_model_path, model_dir=annotator_ckpts_path)
+            load_file_from_url(hand_model_path, model_dir=annotator_ckpts_path)
+
+        self.body_estimation = Body(body_modelpath)
+        self.hand_estimation = Hand(hand_modelpath)
+
+    def __call__(self, oriImg, hand=False):
+        oriImg = oriImg[:, :, ::-1].copy()
+        with torch.no_grad():
+            candidate, subset = self.body_estimation(oriImg)
+            canvas = np.zeros_like(oriImg)
+            canvas = util.draw_bodypose(canvas, candidate, subset)
+            if hand:
+                hands_list = util.handDetect(candidate, subset, oriImg)
+                all_hand_peaks = []
+                for x, y, w, is_left in hands_list:
+                    peaks = self.hand_estimation(oriImg[y:y+w, x:x+w, :])
+                    peaks[:, 0] = np.where(peaks[:, 0] == 0, peaks[:, 0], peaks[:, 0] + x)
+                    peaks[:, 1] = np.where(peaks[:, 1] == 0, peaks[:, 1], peaks[:, 1] + y)
+                    all_hand_peaks.append(peaks)
+                canvas = util.draw_handpose(canvas, all_hand_peaks)
+            return canvas, dict(candidate=candidate.tolist(), subset=subset.tolist())

lavis/common/annotator/openpose/body.py

@@ -0,0 +1,219 @@
+import cv2
+import numpy as np
+import math
+import time
+from scipy.ndimage.filters import gaussian_filter
+import matplotlib.pyplot as plt
+import matplotlib
+import torch
+from torchvision import transforms
+
+from . import util
+from .model import bodypose_model
+
+class Body(object):
+    def __init__(self, model_path):
+        self.model = bodypose_model()
+        if torch.cuda.is_available():
+            self.model = self.model.cuda()
+            print('cuda')
+        model_dict = util.transfer(self.model, torch.load(model_path))
+        self.model.load_state_dict(model_dict)
+        self.model.eval()
+
+    def __call__(self, oriImg):
+        # scale_search = [0.5, 1.0, 1.5, 2.0]
+        scale_search = [0.5]
+        boxsize = 368
+        stride = 8
+        padValue = 128
+        thre1 = 0.1
+        thre2 = 0.05
+        multiplier = [x * boxsize / oriImg.shape[0] for x in scale_search]
+        heatmap_avg = np.zeros((oriImg.shape[0], oriImg.shape[1], 19))
+        paf_avg = np.zeros((oriImg.shape[0], oriImg.shape[1], 38))
+
+        for m in range(len(multiplier)):
+            scale = multiplier[m]
+            imageToTest = cv2.resize(oriImg, (0, 0), fx=scale, fy=scale, interpolation=cv2.INTER_CUBIC)
+            imageToTest_padded, pad = util.padRightDownCorner(imageToTest, stride, padValue)
+            im = np.transpose(np.float32(imageToTest_padded[:, :, :, np.newaxis]), (3, 2, 0, 1)) / 256 - 0.5
+            im = np.ascontiguousarray(im)
+
+            data = torch.from_numpy(im).float()
+            if torch.cuda.is_available():
+                data = data.cuda()
+            # data = data.permute([2, 0, 1]).unsqueeze(0).float()
+            with torch.no_grad():
+                Mconv7_stage6_L1, Mconv7_stage6_L2 = self.model(data)
+            Mconv7_stage6_L1 = Mconv7_stage6_L1.cpu().numpy()
+            Mconv7_stage6_L2 = Mconv7_stage6_L2.cpu().numpy()
+
+            # extract outputs, resize, and remove padding
+            # heatmap = np.transpose(np.squeeze(net.blobs[output_blobs.keys()[1]].data), (1, 2, 0))  # output 1 is heatmaps
+            heatmap = np.transpose(np.squeeze(Mconv7_stage6_L2), (1, 2, 0))  # output 1 is heatmaps
+            heatmap = cv2.resize(heatmap, (0, 0), fx=stride, fy=stride, interpolation=cv2.INTER_CUBIC)
+            heatmap = heatmap[:imageToTest_padded.shape[0] - pad[2], :imageToTest_padded.shape[1] - pad[3], :]
+            heatmap = cv2.resize(heatmap, (oriImg.shape[1], oriImg.shape[0]), interpolation=cv2.INTER_CUBIC)
+
+            # paf = np.transpose(np.squeeze(net.blobs[output_blobs.keys()[0]].data), (1, 2, 0))  # output 0 is PAFs
+            paf = np.transpose(np.squeeze(Mconv7_stage6_L1), (1, 2, 0))  # output 0 is PAFs
+            paf = cv2.resize(paf, (0, 0), fx=stride, fy=stride, interpolation=cv2.INTER_CUBIC)
+            paf = paf[:imageToTest_padded.shape[0] - pad[2], :imageToTest_padded.shape[1] - pad[3], :]
+            paf = cv2.resize(paf, (oriImg.shape[1], oriImg.shape[0]), interpolation=cv2.INTER_CUBIC)
+
+            heatmap_avg += heatmap_avg + heatmap / len(multiplier)
+            paf_avg += + paf / len(multiplier)
+
+        all_peaks = []
+        peak_counter = 0
+
+        for part in range(18):
+            map_ori = heatmap_avg[:, :, part]
+            one_heatmap = gaussian_filter(map_ori, sigma=3)
+
+            map_left = np.zeros(one_heatmap.shape)
+            map_left[1:, :] = one_heatmap[:-1, :]
+            map_right = np.zeros(one_heatmap.shape)
+            map_right[:-1, :] = one_heatmap[1:, :]
+            map_up = np.zeros(one_heatmap.shape)
+            map_up[:, 1:] = one_heatmap[:, :-1]
+            map_down = np.zeros(one_heatmap.shape)
+            map_down[:, :-1] = one_heatmap[:, 1:]
+
+            peaks_binary = np.logical_and.reduce(
+                (one_heatmap >= map_left, one_heatmap >= map_right, one_heatmap >= map_up, one_heatmap >= map_down, one_heatmap > thre1))
+            peaks = list(zip(np.nonzero(peaks_binary)[1], np.nonzero(peaks_binary)[0]))  # note reverse
+            peaks_with_score = [x + (map_ori[x[1], x[0]],) for x in peaks]
+            peak_id = range(peak_counter, peak_counter + len(peaks))
+            peaks_with_score_and_id = [peaks_with_score[i] + (peak_id[i],) for i in range(len(peak_id))]
+
+            all_peaks.append(peaks_with_score_and_id)
+            peak_counter += len(peaks)
+
+        # find connection in the specified sequence, center 29 is in the position 15
+        limbSeq = [[2, 3], [2, 6], [3, 4], [4, 5], [6, 7], [7, 8], [2, 9], [9, 10], \
+                   [10, 11], [2, 12], [12, 13], [13, 14], [2, 1], [1, 15], [15, 17], \
+                   [1, 16], [16, 18], [3, 17], [6, 18]]
+        # the middle joints heatmap correpondence
+        mapIdx = [[31, 32], [39, 40], [33, 34], [35, 36], [41, 42], [43, 44], [19, 20], [21, 22], \
+                  [23, 24], [25, 26], [27, 28], [29, 30], [47, 48], [49, 50], [53, 54], [51, 52], \
+                  [55, 56], [37, 38], [45, 46]]
+
+        connection_all = []
+        special_k = []
+        mid_num = 10
+
+        for k in range(len(mapIdx)):
+            score_mid = paf_avg[:, :, [x - 19 for x in mapIdx[k]]]
+            candA = all_peaks[limbSeq[k][0] - 1]
+            candB = all_peaks[limbSeq[k][1] - 1]
+            nA = len(candA)
+            nB = len(candB)
+            indexA, indexB = limbSeq[k]
+            if (nA != 0 and nB != 0):
+                connection_candidate = []
+                for i in range(nA):
+                    for j in range(nB):
+                        vec = np.subtract(candB[j][:2], candA[i][:2])
+                        norm = math.sqrt(vec[0] * vec[0] + vec[1] * vec[1])
+                        norm = max(0.001, norm)
+                        vec = np.divide(vec, norm)
+
+                        startend = list(zip(np.linspace(candA[i][0], candB[j][0], num=mid_num), \
+                                            np.linspace(candA[i][1], candB[j][1], num=mid_num)))
+
+                        vec_x = np.array([score_mid[int(round(startend[I][1])), int(round(startend[I][0])), 0] \
+                                          for I in range(len(startend))])
+                        vec_y = np.array([score_mid[int(round(startend[I][1])), int(round(startend[I][0])), 1] \
+                                          for I in range(len(startend))])
+
+                        score_midpts = np.multiply(vec_x, vec[0]) + np.multiply(vec_y, vec[1])
+                        score_with_dist_prior = sum(score_midpts) / len(score_midpts) + min(
+                            0.5 * oriImg.shape[0] / norm - 1, 0)
+                        criterion1 = len(np.nonzero(score_midpts > thre2)[0]) > 0.8 * len(score_midpts)
+                        criterion2 = score_with_dist_prior > 0
+                        if criterion1 and criterion2:
+                            connection_candidate.append(
+                                [i, j, score_with_dist_prior, score_with_dist_prior + candA[i][2] + candB[j][2]])
+
+                connection_candidate = sorted(connection_candidate, key=lambda x: x[2], reverse=True)
+                connection = np.zeros((0, 5))
+                for c in range(len(connection_candidate)):
+                    i, j, s = connection_candidate[c][0:3]
+                    if (i not in connection[:, 3] and j not in connection[:, 4]):
+                        connection = np.vstack([connection, [candA[i][3], candB[j][3], s, i, j]])
+                        if (len(connection) >= min(nA, nB)):
+                            break
+
+                connection_all.append(connection)
+            else:
+                special_k.append(k)
+                connection_all.append([])
+
+        # last number in each row is the total parts number of that person
+        # the second last number in each row is the score of the overall configuration
+        subset = -1 * np.ones((0, 20))
+        candidate = np.array([item for sublist in all_peaks for item in sublist])
+
+        for k in range(len(mapIdx)):
+            if k not in special_k:
+                partAs = connection_all[k][:, 0]
+                partBs = connection_all[k][:, 1]
+                indexA, indexB = np.array(limbSeq[k]) - 1
+
+                for i in range(len(connection_all[k])):  # = 1:size(temp,1)
+                    found = 0
+                    subset_idx = [-1, -1]
+                    for j in range(len(subset)):  # 1:size(subset,1):
+                        if subset[j][indexA] == partAs[i] or subset[j][indexB] == partBs[i]:
+                            subset_idx[found] = j
+                            found += 1
+
+                    if found == 1:
+                        j = subset_idx[0]
+                        if subset[j][indexB] != partBs[i]:
+                            subset[j][indexB] = partBs[i]
+                            subset[j][-1] += 1
+                            subset[j][-2] += candidate[partBs[i].astype(int), 2] + connection_all[k][i][2]
+                    elif found == 2:  # if found 2 and disjoint, merge them
+                        j1, j2 = subset_idx
+                        membership = ((subset[j1] >= 0).astype(int) + (subset[j2] >= 0).astype(int))[:-2]
+                        if len(np.nonzero(membership == 2)[0]) == 0:  # merge
+                            subset[j1][:-2] += (subset[j2][:-2] + 1)
+                            subset[j1][-2:] += subset[j2][-2:]
+                            subset[j1][-2] += connection_all[k][i][2]
+                            subset = np.delete(subset, j2, 0)
+                        else:  # as like found == 1
+                            subset[j1][indexB] = partBs[i]
+                            subset[j1][-1] += 1
+                            subset[j1][-2] += candidate[partBs[i].astype(int), 2] + connection_all[k][i][2]
+
+                    # if find no partA in the subset, create a new subset
+                    elif not found and k < 17:
+                        row = -1 * np.ones(20)
+                        row[indexA] = partAs[i]
+                        row[indexB] = partBs[i]
+                        row[-1] = 2
+                        row[-2] = sum(candidate[connection_all[k][i, :2].astype(int), 2]) + connection_all[k][i][2]
+                        subset = np.vstack([subset, row])
+        # delete some rows of subset which has few parts occur
+        deleteIdx = []
+        for i in range(len(subset)):
+            if subset[i][-1] < 4 or subset[i][-2] / subset[i][-1] < 0.4:
+                deleteIdx.append(i)
+        subset = np.delete(subset, deleteIdx, axis=0)
+
+        # subset: n*20 array, 0-17 is the index in candidate, 18 is the total score, 19 is the total parts
+        # candidate: x, y, score, id
+        return candidate, subset
+
+if __name__ == "__main__":
+    body_estimation = Body('../model/body_pose_model.pth')
+
+    test_image = '../images/ski.jpg'
+    oriImg = cv2.imread(test_image)  # B,G,R order
+    candidate, subset = body_estimation(oriImg)
+    canvas = util.draw_bodypose(oriImg, candidate, subset)
+    plt.imshow(canvas[:, :, [2, 1, 0]])
+    plt.show()

lavis/common/annotator/openpose/hand.py

@@ -0,0 +1,86 @@
+import cv2
+import json
+import numpy as np
+import math
+import time
+from scipy.ndimage.filters import gaussian_filter
+import matplotlib.pyplot as plt
+import matplotlib
+import torch
+from skimage.measure import label
+
+from .model import handpose_model
+from . import util
+
+class Hand(object):
+    def __init__(self, model_path):
+        self.model = handpose_model()
+        if torch.cuda.is_available():
+            self.model = self.model.cuda()
+            print('cuda')
+        model_dict = util.transfer(self.model, torch.load(model_path))
+        self.model.load_state_dict(model_dict)
+        self.model.eval()
+
+    def __call__(self, oriImg):
+        scale_search = [0.5, 1.0, 1.5, 2.0]
+        # scale_search = [0.5]
+        boxsize = 368
+        stride = 8
+        padValue = 128
+        thre = 0.05
+        multiplier = [x * boxsize / oriImg.shape[0] for x in scale_search]
+        heatmap_avg = np.zeros((oriImg.shape[0], oriImg.shape[1], 22))
+        # paf_avg = np.zeros((oriImg.shape[0], oriImg.shape[1], 38))
+
+        for m in range(len(multiplier)):
+            scale = multiplier[m]
+            imageToTest = cv2.resize(oriImg, (0, 0), fx=scale, fy=scale, interpolation=cv2.INTER_CUBIC)
+            imageToTest_padded, pad = util.padRightDownCorner(imageToTest, stride, padValue)
+            im = np.transpose(np.float32(imageToTest_padded[:, :, :, np.newaxis]), (3, 2, 0, 1)) / 256 - 0.5
+            im = np.ascontiguousarray(im)
+
+            data = torch.from_numpy(im).float()
+            if torch.cuda.is_available():
+                data = data.cuda()
+            # data = data.permute([2, 0, 1]).unsqueeze(0).float()
+            with torch.no_grad():
+                output = self.model(data).cpu().numpy()
+                # output = self.model(data).numpy()q
+
+            # extract outputs, resize, and remove padding
+            heatmap = np.transpose(np.squeeze(output), (1, 2, 0))  # output 1 is heatmaps
+            heatmap = cv2.resize(heatmap, (0, 0), fx=stride, fy=stride, interpolation=cv2.INTER_CUBIC)
+            heatmap = heatmap[:imageToTest_padded.shape[0] - pad[2], :imageToTest_padded.shape[1] - pad[3], :]
+            heatmap = cv2.resize(heatmap, (oriImg.shape[1], oriImg.shape[0]), interpolation=cv2.INTER_CUBIC)
+
+            heatmap_avg += heatmap / len(multiplier)
+
+        all_peaks = []
+        for part in range(21):
+            map_ori = heatmap_avg[:, :, part]
+            one_heatmap = gaussian_filter(map_ori, sigma=3)
+            binary = np.ascontiguousarray(one_heatmap > thre, dtype=np.uint8)
+            # 全部小于阈值
+            if np.sum(binary) == 0:
+                all_peaks.append([0, 0])
+                continue
+            label_img, label_numbers = label(binary, return_num=True, connectivity=binary.ndim)
+            max_index = np.argmax([np.sum(map_ori[label_img == i]) for i in range(1, label_numbers + 1)]) + 1
+            label_img[label_img != max_index] = 0
+            map_ori[label_img == 0] = 0
+
+            y, x = util.npmax(map_ori)
+            all_peaks.append([x, y])
+        return np.array(all_peaks)
+
+if __name__ == "__main__":
+    hand_estimation = Hand('../model/hand_pose_model.pth')
+
+    # test_image = '../images/hand.jpg'
+    test_image = '../images/hand.jpg'
+    oriImg = cv2.imread(test_image)  # B,G,R order
+    peaks = hand_estimation(oriImg)
+    canvas = util.draw_handpose(oriImg, peaks, True)
+    cv2.imshow('', canvas)
+    cv2.waitKey(0)

lavis/common/annotator/openpose/model.py

@@ -0,0 +1,219 @@
+import torch
+from collections import OrderedDict
+
+import torch
+import torch.nn as nn
+
+def make_layers(block, no_relu_layers):
+    layers = []
+    for layer_name, v in block.items():
+        if 'pool' in layer_name:
+            layer = nn.MaxPool2d(kernel_size=v[0], stride=v[1],
+                                    padding=v[2])
+            layers.append((layer_name, layer))
+        else:
+            conv2d = nn.Conv2d(in_channels=v[0], out_channels=v[1],
+                               kernel_size=v[2], stride=v[3],
+                               padding=v[4])
+            layers.append((layer_name, conv2d))
+            if layer_name not in no_relu_layers:
+                layers.append(('relu_'+layer_name, nn.ReLU(inplace=True)))
+
+    return nn.Sequential(OrderedDict(layers))
+
+class bodypose_model(nn.Module):
+    def __init__(self):
+        super(bodypose_model, self).__init__()
+
+        # these layers have no relu layer
+        no_relu_layers = ['conv5_5_CPM_L1', 'conv5_5_CPM_L2', 'Mconv7_stage2_L1',\
+                          'Mconv7_stage2_L2', 'Mconv7_stage3_L1', 'Mconv7_stage3_L2',\
+                          'Mconv7_stage4_L1', 'Mconv7_stage4_L2', 'Mconv7_stage5_L1',\
+                          'Mconv7_stage5_L2', 'Mconv7_stage6_L1', 'Mconv7_stage6_L1']
+        blocks = {}
+        block0 = OrderedDict([
+                      ('conv1_1', [3, 64, 3, 1, 1]),
+                      ('conv1_2', [64, 64, 3, 1, 1]),
+                      ('pool1_stage1', [2, 2, 0]),
+                      ('conv2_1', [64, 128, 3, 1, 1]),
+                      ('conv2_2', [128, 128, 3, 1, 1]),
+                      ('pool2_stage1', [2, 2, 0]),
+                      ('conv3_1', [128, 256, 3, 1, 1]),
+                      ('conv3_2', [256, 256, 3, 1, 1]),
+                      ('conv3_3', [256, 256, 3, 1, 1]),
+                      ('conv3_4', [256, 256, 3, 1, 1]),
+                      ('pool3_stage1', [2, 2, 0]),
+                      ('conv4_1', [256, 512, 3, 1, 1]),
+                      ('conv4_2', [512, 512, 3, 1, 1]),
+                      ('conv4_3_CPM', [512, 256, 3, 1, 1]),
+                      ('conv4_4_CPM', [256, 128, 3, 1, 1])
+                  ])
+
+
+        # Stage 1
+        block1_1 = OrderedDict([
+                        ('conv5_1_CPM_L1', [128, 128, 3, 1, 1]),
+                        ('conv5_2_CPM_L1', [128, 128, 3, 1, 1]),
+                        ('conv5_3_CPM_L1', [128, 128, 3, 1, 1]),
+                        ('conv5_4_CPM_L1', [128, 512, 1, 1, 0]),
+                        ('conv5_5_CPM_L1', [512, 38, 1, 1, 0])
+                    ])
+
+        block1_2 = OrderedDict([
+                        ('conv5_1_CPM_L2', [128, 128, 3, 1, 1]),
+                        ('conv5_2_CPM_L2', [128, 128, 3, 1, 1]),
+                        ('conv5_3_CPM_L2', [128, 128, 3, 1, 1]),
+                        ('conv5_4_CPM_L2', [128, 512, 1, 1, 0]),
+                        ('conv5_5_CPM_L2', [512, 19, 1, 1, 0])
+                    ])
+        blocks['block1_1'] = block1_1
+        blocks['block1_2'] = block1_2
+
+        self.model0 = make_layers(block0, no_relu_layers)
+
+        # Stages 2 - 6
+        for i in range(2, 7):
+            blocks['block%d_1' % i] = OrderedDict([
+                    ('Mconv1_stage%d_L1' % i, [185, 128, 7, 1, 3]),
+                    ('Mconv2_stage%d_L1' % i, [128, 128, 7, 1, 3]),
+                    ('Mconv3_stage%d_L1' % i, [128, 128, 7, 1, 3]),
+                    ('Mconv4_stage%d_L1' % i, [128, 128, 7, 1, 3]),
+                    ('Mconv5_stage%d_L1' % i, [128, 128, 7, 1, 3]),
+                    ('Mconv6_stage%d_L1' % i, [128, 128, 1, 1, 0]),
+                    ('Mconv7_stage%d_L1' % i, [128, 38, 1, 1, 0])
+                ])
+
+            blocks['block%d_2' % i] = OrderedDict([
+                    ('Mconv1_stage%d_L2' % i, [185, 128, 7, 1, 3]),
+                    ('Mconv2_stage%d_L2' % i, [128, 128, 7, 1, 3]),
+                    ('Mconv3_stage%d_L2' % i, [128, 128, 7, 1, 3]),
+                    ('Mconv4_stage%d_L2' % i, [128, 128, 7, 1, 3]),
+                    ('Mconv5_stage%d_L2' % i, [128, 128, 7, 1, 3]),
+                    ('Mconv6_stage%d_L2' % i, [128, 128, 1, 1, 0]),
+                    ('Mconv7_stage%d_L2' % i, [128, 19, 1, 1, 0])
+                ])
+
+        for k in blocks.keys():
+            blocks[k] = make_layers(blocks[k], no_relu_layers)
+
+        self.model1_1 = blocks['block1_1']
+        self.model2_1 = blocks['block2_1']
+        self.model3_1 = blocks['block3_1']
+        self.model4_1 = blocks['block4_1']
+        self.model5_1 = blocks['block5_1']
+        self.model6_1 = blocks['block6_1']
+
+        self.model1_2 = blocks['block1_2']
+        self.model2_2 = blocks['block2_2']
+        self.model3_2 = blocks['block3_2']
+        self.model4_2 = blocks['block4_2']
+        self.model5_2 = blocks['block5_2']
+        self.model6_2 = blocks['block6_2']
+
+
+    def forward(self, x):
+
+        out1 = self.model0(x)
+
+        out1_1 = self.model1_1(out1)
+        out1_2 = self.model1_2(out1)
+        out2 = torch.cat([out1_1, out1_2, out1], 1)
+
+        out2_1 = self.model2_1(out2)
+        out2_2 = self.model2_2(out2)
+        out3 = torch.cat([out2_1, out2_2, out1], 1)
+
+        out3_1 = self.model3_1(out3)
+        out3_2 = self.model3_2(out3)
+        out4 = torch.cat([out3_1, out3_2, out1], 1)
+
+        out4_1 = self.model4_1(out4)
+        out4_2 = self.model4_2(out4)
+        out5 = torch.cat([out4_1, out4_2, out1], 1)
+
+        out5_1 = self.model5_1(out5)
+        out5_2 = self.model5_2(out5)
+        out6 = torch.cat([out5_1, out5_2, out1], 1)
+
+        out6_1 = self.model6_1(out6)
+        out6_2 = self.model6_2(out6)
+
+        return out6_1, out6_2
+
+class handpose_model(nn.Module):
+    def __init__(self):
+        super(handpose_model, self).__init__()
+
+        # these layers have no relu layer
+        no_relu_layers = ['conv6_2_CPM', 'Mconv7_stage2', 'Mconv7_stage3',\
+                          'Mconv7_stage4', 'Mconv7_stage5', 'Mconv7_stage6']
+        # stage 1
+        block1_0 = OrderedDict([
+                ('conv1_1', [3, 64, 3, 1, 1]),
+                ('conv1_2', [64, 64, 3, 1, 1]),
+                ('pool1_stage1', [2, 2, 0]),
+                ('conv2_1', [64, 128, 3, 1, 1]),
+                ('conv2_2', [128, 128, 3, 1, 1]),
+                ('pool2_stage1', [2, 2, 0]),
+                ('conv3_1', [128, 256, 3, 1, 1]),
+                ('conv3_2', [256, 256, 3, 1, 1]),
+                ('conv3_3', [256, 256, 3, 1, 1]),
+                ('conv3_4', [256, 256, 3, 1, 1]),
+                ('pool3_stage1', [2, 2, 0]),
+                ('conv4_1', [256, 512, 3, 1, 1]),
+                ('conv4_2', [512, 512, 3, 1, 1]),
+                ('conv4_3', [512, 512, 3, 1, 1]),
+                ('conv4_4', [512, 512, 3, 1, 1]),
+                ('conv5_1', [512, 512, 3, 1, 1]),
+                ('conv5_2', [512, 512, 3, 1, 1]),
+                ('conv5_3_CPM', [512, 128, 3, 1, 1])
+            ])
+
+        block1_1 = OrderedDict([
+            ('conv6_1_CPM', [128, 512, 1, 1, 0]),
+            ('conv6_2_CPM', [512, 22, 1, 1, 0])
+        ])
+
+        blocks = {}
+        blocks['block1_0'] = block1_0
+        blocks['block1_1'] = block1_1
+
+        # stage 2-6
+        for i in range(2, 7):
+            blocks['block%d' % i] = OrderedDict([
+                    ('Mconv1_stage%d' % i, [150, 128, 7, 1, 3]),
+                    ('Mconv2_stage%d' % i, [128, 128, 7, 1, 3]),
+                    ('Mconv3_stage%d' % i, [128, 128, 7, 1, 3]),
+                    ('Mconv4_stage%d' % i, [128, 128, 7, 1, 3]),
+                    ('Mconv5_stage%d' % i, [128, 128, 7, 1, 3]),
+                    ('Mconv6_stage%d' % i, [128, 128, 1, 1, 0]),
+                    ('Mconv7_stage%d' % i, [128, 22, 1, 1, 0])
+                ])
+
+        for k in blocks.keys():
+            blocks[k] = make_layers(blocks[k], no_relu_layers)
+
+        self.model1_0 = blocks['block1_0']
+        self.model1_1 = blocks['block1_1']
+        self.model2 = blocks['block2']
+        self.model3 = blocks['block3']
+        self.model4 = blocks['block4']
+        self.model5 = blocks['block5']
+        self.model6 = blocks['block6']
+
+    def forward(self, x):
+        out1_0 = self.model1_0(x)
+        out1_1 = self.model1_1(out1_0)
+        concat_stage2 = torch.cat([out1_1, out1_0], 1)
+        out_stage2 = self.model2(concat_stage2)
+        concat_stage3 = torch.cat([out_stage2, out1_0], 1)
+        out_stage3 = self.model3(concat_stage3)
+        concat_stage4 = torch.cat([out_stage3, out1_0], 1)
+        out_stage4 = self.model4(concat_stage4)
+        concat_stage5 = torch.cat([out_stage4, out1_0], 1)
+        out_stage5 = self.model5(concat_stage5)
+        concat_stage6 = torch.cat([out_stage5, out1_0], 1)
+        out_stage6 = self.model6(concat_stage6)
+        return out_stage6
+
+

lavis/common/annotator/openpose/util.py

@@ -0,0 +1,164 @@
+import math
+import numpy as np
+import matplotlib
+import cv2
+
+
+def padRightDownCorner(img, stride, padValue):
+    h = img.shape[0]
+    w = img.shape[1]
+
+    pad = 4 * [None]
+    pad[0] = 0 # up
+    pad[1] = 0 # left
+    pad[2] = 0 if (h % stride == 0) else stride - (h % stride) # down
+    pad[3] = 0 if (w % stride == 0) else stride - (w % stride) # right
+
+    img_padded = img
+    pad_up = np.tile(img_padded[0:1, :, :]*0 + padValue, (pad[0], 1, 1))
+    img_padded = np.concatenate((pad_up, img_padded), axis=0)
+    pad_left = np.tile(img_padded[:, 0:1, :]*0 + padValue, (1, pad[1], 1))
+    img_padded = np.concatenate((pad_left, img_padded), axis=1)
+    pad_down = np.tile(img_padded[-2:-1, :, :]*0 + padValue, (pad[2], 1, 1))
+    img_padded = np.concatenate((img_padded, pad_down), axis=0)
+    pad_right = np.tile(img_padded[:, -2:-1, :]*0 + padValue, (1, pad[3], 1))
+    img_padded = np.concatenate((img_padded, pad_right), axis=1)
+
+    return img_padded, pad
+
+# transfer caffe model to pytorch which will match the layer name
+def transfer(model, model_weights):
+    transfered_model_weights = {}
+    for weights_name in model.state_dict().keys():
+        transfered_model_weights[weights_name] = model_weights['.'.join(weights_name.split('.')[1:])]
+    return transfered_model_weights
+
+# draw the body keypoint and lims
+def draw_bodypose(canvas, candidate, subset):
+    stickwidth = 4
+    limbSeq = [[2, 3], [2, 6], [3, 4], [4, 5], [6, 7], [7, 8], [2, 9], [9, 10], \
+               [10, 11], [2, 12], [12, 13], [13, 14], [2, 1], [1, 15], [15, 17], \
+               [1, 16], [16, 18], [3, 17], [6, 18]]
+
+    colors = [[255, 0, 0], [255, 85, 0], [255, 170, 0], [255, 255, 0], [170, 255, 0], [85, 255, 0], [0, 255, 0], \
+              [0, 255, 85], [0, 255, 170], [0, 255, 255], [0, 170, 255], [0, 85, 255], [0, 0, 255], [85, 0, 255], \
+              [170, 0, 255], [255, 0, 255], [255, 0, 170], [255, 0, 85]]
+    for i in range(18):
+        for n in range(len(subset)):
+            index = int(subset[n][i])
+            if index == -1:
+                continue
+            x, y = candidate[index][0:2]
+            cv2.circle(canvas, (int(x), int(y)), 4, colors[i], thickness=-1)
+    for i in range(17):
+        for n in range(len(subset)):
+            index = subset[n][np.array(limbSeq[i]) - 1]
+            if -1 in index:
+                continue
+            cur_canvas = canvas.copy()
+            Y = candidate[index.astype(int), 0]
+            X = candidate[index.astype(int), 1]
+            mX = np.mean(X)
+            mY = np.mean(Y)
+            length = ((X[0] - X[1]) ** 2 + (Y[0] - Y[1]) ** 2) ** 0.5
+            angle = math.degrees(math.atan2(X[0] - X[1], Y[0] - Y[1]))
+            polygon = cv2.ellipse2Poly((int(mY), int(mX)), (int(length / 2), stickwidth), int(angle), 0, 360, 1)
+            cv2.fillConvexPoly(cur_canvas, polygon, colors[i])
+            canvas = cv2.addWeighted(canvas, 0.4, cur_canvas, 0.6, 0)
+    # plt.imsave("preview.jpg", canvas[:, :, [2, 1, 0]])
+    # plt.imshow(canvas[:, :, [2, 1, 0]])
+    return canvas
+
+
+# image drawed by opencv is not good.
+def draw_handpose(canvas, all_hand_peaks, show_number=False):
+    edges = [[0, 1], [1, 2], [2, 3], [3, 4], [0, 5], [5, 6], [6, 7], [7, 8], [0, 9], [9, 10], \
+             [10, 11], [11, 12], [0, 13], [13, 14], [14, 15], [15, 16], [0, 17], [17, 18], [18, 19], [19, 20]]
+
+    for peaks in all_hand_peaks:
+        for ie, e in enumerate(edges):
+            if np.sum(np.all(peaks[e], axis=1)==0)==0:
+                x1, y1 = peaks[e[0]]
+                x2, y2 = peaks[e[1]]
+                cv2.line(canvas, (x1, y1), (x2, y2), matplotlib.colors.hsv_to_rgb([ie/float(len(edges)), 1.0, 1.0])*255, thickness=2)
+
+        for i, keyponit in enumerate(peaks):
+            x, y = keyponit
+            cv2.circle(canvas, (x, y), 4, (0, 0, 255), thickness=-1)
+            if show_number:
+                cv2.putText(canvas, str(i), (x, y), cv2.FONT_HERSHEY_SIMPLEX, 0.3, (0, 0, 0), lineType=cv2.LINE_AA)
+    return canvas
+
+# detect hand according to body pose keypoints
+# please refer to https://github.com/CMU-Perceptual-Computing-Lab/openpose/blob/master/src/openpose/hand/handDetector.cpp
+def handDetect(candidate, subset, oriImg):
+    # right hand: wrist 4, elbow 3, shoulder 2
+    # left hand: wrist 7, elbow 6, shoulder 5
+    ratioWristElbow = 0.33
+    detect_result = []
+    image_height, image_width = oriImg.shape[0:2]
+    for person in subset.astype(int):
+        # if any of three not detected
+        has_left = np.sum(person[[5, 6, 7]] == -1) == 0
+        has_right = np.sum(person[[2, 3, 4]] == -1) == 0
+        if not (has_left or has_right):
+            continue
+        hands = []
+        #left hand
+        if has_left:
+            left_shoulder_index, left_elbow_index, left_wrist_index = person[[5, 6, 7]]
+            x1, y1 = candidate[left_shoulder_index][:2]
+            x2, y2 = candidate[left_elbow_index][:2]
+            x3, y3 = candidate[left_wrist_index][:2]
+            hands.append([x1, y1, x2, y2, x3, y3, True])
+        # right hand
+        if has_right:
+            right_shoulder_index, right_elbow_index, right_wrist_index = person[[2, 3, 4]]
+            x1, y1 = candidate[right_shoulder_index][:2]
+            x2, y2 = candidate[right_elbow_index][:2]
+            x3, y3 = candidate[right_wrist_index][:2]
+            hands.append([x1, y1, x2, y2, x3, y3, False])
+
+        for x1, y1, x2, y2, x3, y3, is_left in hands:
+            # pos_hand = pos_wrist + ratio * (pos_wrist - pos_elbox) = (1 + ratio) * pos_wrist - ratio * pos_elbox
+            # handRectangle.x = posePtr[wrist*3] + ratioWristElbow * (posePtr[wrist*3] - posePtr[elbow*3]);
+            # handRectangle.y = posePtr[wrist*3+1] + ratioWristElbow * (posePtr[wrist*3+1] - posePtr[elbow*3+1]);
+            # const auto distanceWristElbow = getDistance(poseKeypoints, person, wrist, elbow);
+            # const auto distanceElbowShoulder = getDistance(poseKeypoints, person, elbow, shoulder);
+            # handRectangle.width = 1.5f * fastMax(distanceWristElbow, 0.9f * distanceElbowShoulder);
+            x = x3 + ratioWristElbow * (x3 - x2)
+            y = y3 + ratioWristElbow * (y3 - y2)
+            distanceWristElbow = math.sqrt((x3 - x2) ** 2 + (y3 - y2) ** 2)
+            distanceElbowShoulder = math.sqrt((x2 - x1) ** 2 + (y2 - y1) ** 2)
+            width = 1.5 * max(distanceWristElbow, 0.9 * distanceElbowShoulder)
+            # x-y refers to the center --> offset to topLeft point
+            # handRectangle.x -= handRectangle.width / 2.f;
+            # handRectangle.y -= handRectangle.height / 2.f;
+            x -= width / 2
+            y -= width / 2  # width = height
+            # overflow the image
+            if x < 0: x = 0
+            if y < 0: y = 0
+            width1 = width
+            width2 = width
+            if x + width > image_width: width1 = image_width - x
+            if y + width > image_height: width2 = image_height - y
+            width = min(width1, width2)
+            # the max hand box value is 20 pixels
+            if width >= 20:
+                detect_result.append([int(x), int(y), int(width), is_left])
+
+    '''
+    return value: [[x, y, w, True if left hand else False]].
+    width=height since the network require squared input.
+    x, y is the coordinate of top left 
+    '''
+    return detect_result
+
+# get max index of 2d array
+def npmax(array):
+    arrayindex = array.argmax(1)
+    arrayvalue = array.max(1)
+    i = arrayvalue.argmax()
+    j = arrayindex[i]
+    return i, j

lavis/common/annotator/uniformer/__init__.py

@@ -0,0 +1,23 @@
+import os
+
+from annotator.uniformer.mmseg.apis import init_segmentor, inference_segmentor, show_result_pyplot
+from annotator.uniformer.mmseg.core.evaluation import get_palette
+from annotator.util import annotator_ckpts_path
+
+
+checkpoint_file = "https://huggingface.co/lllyasviel/ControlNet/resolve/main/annotator/ckpts/upernet_global_small.pth"
+
+
+class UniformerDetector:
+    def __init__(self):
+        modelpath = os.path.join(annotator_ckpts_path, "upernet_global_small.pth")
+        if not os.path.exists(modelpath):
+            from basicsr.utils.download_util import load_file_from_url
+            load_file_from_url(checkpoint_file, model_dir=annotator_ckpts_path)
+        config_file = os.path.join(os.path.dirname(annotator_ckpts_path), "uniformer", "exp", "upernet_global_small", "config.py")
+        self.model = init_segmentor(config_file, modelpath).cuda()
+
+    def __call__(self, img):
+        result = inference_segmentor(self.model, img)
+        res_img = show_result_pyplot(self.model, img, result, get_palette('ade'), opacity=1)
+        return res_img

lavis/common/annotator/uniformer/configs/_base_/datasets/ade20k.py

@@ -0,0 +1,54 @@
+# dataset settings
+dataset_type = 'ADE20KDataset'
+data_root = 'data/ade/ADEChallengeData2016'
+img_norm_cfg = dict(
+    mean=[123.675, 116.28, 103.53], std=[58.395, 57.12, 57.375], to_rgb=True)
+crop_size = (512, 512)
+train_pipeline = [
+    dict(type='LoadImageFromFile'),
+    dict(type='LoadAnnotations', reduce_zero_label=True),
+    dict(type='Resize', img_scale=(2048, 512), ratio_range=(0.5, 2.0)),
+    dict(type='RandomCrop', crop_size=crop_size, cat_max_ratio=0.75),
+    dict(type='RandomFlip', prob=0.5),
+    dict(type='PhotoMetricDistortion'),
+    dict(type='Normalize', **img_norm_cfg),
+    dict(type='Pad', size=crop_size, pad_val=0, seg_pad_val=255),
+    dict(type='DefaultFormatBundle'),
+    dict(type='Collect', keys=['img', 'gt_semantic_seg']),
+]
+test_pipeline = [
+    dict(type='LoadImageFromFile'),
+    dict(
+        type='MultiScaleFlipAug',
+        img_scale=(2048, 512),
+        # img_ratios=[0.5, 0.75, 1.0, 1.25, 1.5, 1.75],
+        flip=False,
+        transforms=[
+            dict(type='Resize', keep_ratio=True),
+            dict(type='RandomFlip'),
+            dict(type='Normalize', **img_norm_cfg),
+            dict(type='ImageToTensor', keys=['img']),
+            dict(type='Collect', keys=['img']),
+        ])
+]
+data = dict(
+    samples_per_gpu=4,
+    workers_per_gpu=4,
+    train=dict(
+        type=dataset_type,
+        data_root=data_root,
+        img_dir='images/training',
+        ann_dir='annotations/training',
+        pipeline=train_pipeline),
+    val=dict(
+        type=dataset_type,
+        data_root=data_root,
+        img_dir='images/validation',
+        ann_dir='annotations/validation',
+        pipeline=test_pipeline),
+    test=dict(
+        type=dataset_type,
+        data_root=data_root,
+        img_dir='images/validation',
+        ann_dir='annotations/validation',
+        pipeline=test_pipeline))

lavis/common/annotator/uniformer/configs/_base_/datasets/chase_db1.py

@@ -0,0 +1,59 @@
+# dataset settings
+dataset_type = 'ChaseDB1Dataset'
+data_root = 'data/CHASE_DB1'
+img_norm_cfg = dict(
+    mean=[123.675, 116.28, 103.53], std=[58.395, 57.12, 57.375], to_rgb=True)
+img_scale = (960, 999)
+crop_size = (128, 128)
+train_pipeline = [
+    dict(type='LoadImageFromFile'),
+    dict(type='LoadAnnotations'),
+    dict(type='Resize', img_scale=img_scale, ratio_range=(0.5, 2.0)),
+    dict(type='RandomCrop', crop_size=crop_size, cat_max_ratio=0.75),
+    dict(type='RandomFlip', prob=0.5),
+    dict(type='PhotoMetricDistortion'),
+    dict(type='Normalize', **img_norm_cfg),
+    dict(type='Pad', size=crop_size, pad_val=0, seg_pad_val=255),
+    dict(type='DefaultFormatBundle'),
+    dict(type='Collect', keys=['img', 'gt_semantic_seg'])
+]
+test_pipeline = [
+    dict(type='LoadImageFromFile'),
+    dict(
+        type='MultiScaleFlipAug',
+        img_scale=img_scale,
+        # img_ratios=[0.5, 0.75, 1.0, 1.25, 1.5, 1.75, 2.0],
+        flip=False,
+        transforms=[
+            dict(type='Resize', keep_ratio=True),
+            dict(type='RandomFlip'),
+            dict(type='Normalize', **img_norm_cfg),
+            dict(type='ImageToTensor', keys=['img']),
+            dict(type='Collect', keys=['img'])
+        ])
+]
+
+data = dict(
+    samples_per_gpu=4,
+    workers_per_gpu=4,
+    train=dict(
+        type='RepeatDataset',
+        times=40000,
+        dataset=dict(
+            type=dataset_type,
+            data_root=data_root,
+            img_dir='images/training',
+            ann_dir='annotations/training',
+            pipeline=train_pipeline)),
+    val=dict(
+        type=dataset_type,
+        data_root=data_root,
+        img_dir='images/validation',
+        ann_dir='annotations/validation',
+        pipeline=test_pipeline),
+    test=dict(
+        type=dataset_type,
+        data_root=data_root,
+        img_dir='images/validation',
+        ann_dir='annotations/validation',
+        pipeline=test_pipeline))

lavis/common/annotator/uniformer/configs/_base_/datasets/cityscapes.py

@@ -0,0 +1,54 @@
+# dataset settings
+dataset_type = 'CityscapesDataset'
+data_root = 'data/cityscapes/'
+img_norm_cfg = dict(
+    mean=[123.675, 116.28, 103.53], std=[58.395, 57.12, 57.375], to_rgb=True)
+crop_size = (512, 1024)
+train_pipeline = [
+    dict(type='LoadImageFromFile'),
+    dict(type='LoadAnnotations'),
+    dict(type='Resize', img_scale=(2048, 1024), ratio_range=(0.5, 2.0)),
+    dict(type='RandomCrop', crop_size=crop_size, cat_max_ratio=0.75),
+    dict(type='RandomFlip', prob=0.5),
+    dict(type='PhotoMetricDistortion'),
+    dict(type='Normalize', **img_norm_cfg),
+    dict(type='Pad', size=crop_size, pad_val=0, seg_pad_val=255),
+    dict(type='DefaultFormatBundle'),
+    dict(type='Collect', keys=['img', 'gt_semantic_seg']),
+]
+test_pipeline = [
+    dict(type='LoadImageFromFile'),
+    dict(
+        type='MultiScaleFlipAug',
+        img_scale=(2048, 1024),
+        # img_ratios=[0.5, 0.75, 1.0, 1.25, 1.5, 1.75],
+        flip=False,
+        transforms=[
+            dict(type='Resize', keep_ratio=True),
+            dict(type='RandomFlip'),
+            dict(type='Normalize', **img_norm_cfg),
+            dict(type='ImageToTensor', keys=['img']),
+            dict(type='Collect', keys=['img']),
+        ])
+]
+data = dict(
+    samples_per_gpu=2,
+    workers_per_gpu=2,
+    train=dict(
+        type=dataset_type,
+        data_root=data_root,
+        img_dir='leftImg8bit/train',
+        ann_dir='gtFine/train',
+        pipeline=train_pipeline),
+    val=dict(
+        type=dataset_type,
+        data_root=data_root,
+        img_dir='leftImg8bit/val',
+        ann_dir='gtFine/val',
+        pipeline=test_pipeline),
+    test=dict(
+        type=dataset_type,
+        data_root=data_root,
+        img_dir='leftImg8bit/val',
+        ann_dir='gtFine/val',
+        pipeline=test_pipeline))

lavis/common/annotator/uniformer/configs/_base_/datasets/cityscapes_769x769.py

@@ -0,0 +1,35 @@
+_base_ = './cityscapes.py'
+img_norm_cfg = dict(
+    mean=[123.675, 116.28, 103.53], std=[58.395, 57.12, 57.375], to_rgb=True)
+crop_size = (769, 769)
+train_pipeline = [
+    dict(type='LoadImageFromFile'),
+    dict(type='LoadAnnotations'),
+    dict(type='Resize', img_scale=(2049, 1025), ratio_range=(0.5, 2.0)),
+    dict(type='RandomCrop', crop_size=crop_size, cat_max_ratio=0.75),
+    dict(type='RandomFlip', prob=0.5),
+    dict(type='PhotoMetricDistortion'),
+    dict(type='Normalize', **img_norm_cfg),
+    dict(type='Pad', size=crop_size, pad_val=0, seg_pad_val=255),
+    dict(type='DefaultFormatBundle'),
+    dict(type='Collect', keys=['img', 'gt_semantic_seg']),
+]
+test_pipeline = [
+    dict(type='LoadImageFromFile'),
+    dict(
+        type='MultiScaleFlipAug',
+        img_scale=(2049, 1025),
+        # img_ratios=[0.5, 0.75, 1.0, 1.25, 1.5, 1.75],
+        flip=False,
+        transforms=[
+            dict(type='Resize', keep_ratio=True),
+            dict(type='RandomFlip'),
+            dict(type='Normalize', **img_norm_cfg),
+            dict(type='ImageToTensor', keys=['img']),
+            dict(type='Collect', keys=['img']),
+        ])
+]
+data = dict(
+    train=dict(pipeline=train_pipeline),
+    val=dict(pipeline=test_pipeline),
+    test=dict(pipeline=test_pipeline))

lavis/common/annotator/uniformer/configs/_base_/datasets/drive.py

@@ -0,0 +1,59 @@
+# dataset settings
+dataset_type = 'DRIVEDataset'
+data_root = 'data/DRIVE'
+img_norm_cfg = dict(
+    mean=[123.675, 116.28, 103.53], std=[58.395, 57.12, 57.375], to_rgb=True)
+img_scale = (584, 565)
+crop_size = (64, 64)
+train_pipeline = [
+    dict(type='LoadImageFromFile'),
+    dict(type='LoadAnnotations'),
+    dict(type='Resize', img_scale=img_scale, ratio_range=(0.5, 2.0)),
+    dict(type='RandomCrop', crop_size=crop_size, cat_max_ratio=0.75),
+    dict(type='RandomFlip', prob=0.5),
+    dict(type='PhotoMetricDistortion'),
+    dict(type='Normalize', **img_norm_cfg),
+    dict(type='Pad', size=crop_size, pad_val=0, seg_pad_val=255),
+    dict(type='DefaultFormatBundle'),
+    dict(type='Collect', keys=['img', 'gt_semantic_seg'])
+]
+test_pipeline = [
+    dict(type='LoadImageFromFile'),
+    dict(
+        type='MultiScaleFlipAug',
+        img_scale=img_scale,
+        # img_ratios=[0.5, 0.75, 1.0, 1.25, 1.5, 1.75, 2.0],
+        flip=False,
+        transforms=[
+            dict(type='Resize', keep_ratio=True),
+            dict(type='RandomFlip'),
+            dict(type='Normalize', **img_norm_cfg),
+            dict(type='ImageToTensor', keys=['img']),
+            dict(type='Collect', keys=['img'])
+        ])
+]
+
+data = dict(
+    samples_per_gpu=4,
+    workers_per_gpu=4,
+    train=dict(
+        type='RepeatDataset',
+        times=40000,
+        dataset=dict(
+            type=dataset_type,
+            data_root=data_root,
+            img_dir='images/training',
+            ann_dir='annotations/training',
+            pipeline=train_pipeline)),
+    val=dict(
+        type=dataset_type,
+        data_root=data_root,
+        img_dir='images/validation',
+        ann_dir='annotations/validation',
+        pipeline=test_pipeline),
+    test=dict(
+        type=dataset_type,
+        data_root=data_root,
+        img_dir='images/validation',
+        ann_dir='annotations/validation',
+        pipeline=test_pipeline))

lavis/common/annotator/uniformer/configs/_base_/datasets/hrf.py

@@ -0,0 +1,59 @@
+# dataset settings
+dataset_type = 'HRFDataset'
+data_root = 'data/HRF'
+img_norm_cfg = dict(
+    mean=[123.675, 116.28, 103.53], std=[58.395, 57.12, 57.375], to_rgb=True)
+img_scale = (2336, 3504)
+crop_size = (256, 256)
+train_pipeline = [
+    dict(type='LoadImageFromFile'),
+    dict(type='LoadAnnotations'),
+    dict(type='Resize', img_scale=img_scale, ratio_range=(0.5, 2.0)),
+    dict(type='RandomCrop', crop_size=crop_size, cat_max_ratio=0.75),
+    dict(type='RandomFlip', prob=0.5),
+    dict(type='PhotoMetricDistortion'),
+    dict(type='Normalize', **img_norm_cfg),
+    dict(type='Pad', size=crop_size, pad_val=0, seg_pad_val=255),
+    dict(type='DefaultFormatBundle'),
+    dict(type='Collect', keys=['img', 'gt_semantic_seg'])
+]
+test_pipeline = [
+    dict(type='LoadImageFromFile'),
+    dict(
+        type='MultiScaleFlipAug',
+        img_scale=img_scale,
+        # img_ratios=[0.5, 0.75, 1.0, 1.25, 1.5, 1.75, 2.0],
+        flip=False,
+        transforms=[
+            dict(type='Resize', keep_ratio=True),
+            dict(type='RandomFlip'),
+            dict(type='Normalize', **img_norm_cfg),
+            dict(type='ImageToTensor', keys=['img']),
+            dict(type='Collect', keys=['img'])
+        ])
+]
+
+data = dict(
+    samples_per_gpu=4,
+    workers_per_gpu=4,
+    train=dict(
+        type='RepeatDataset',
+        times=40000,
+        dataset=dict(
+            type=dataset_type,
+            data_root=data_root,
+            img_dir='images/training',
+            ann_dir='annotations/training',
+            pipeline=train_pipeline)),
+    val=dict(
+        type=dataset_type,
+        data_root=data_root,
+        img_dir='images/validation',
+        ann_dir='annotations/validation',
+        pipeline=test_pipeline),
+    test=dict(
+        type=dataset_type,
+        data_root=data_root,
+        img_dir='images/validation',
+        ann_dir='annotations/validation',
+        pipeline=test_pipeline))

lavis/common/annotator/uniformer/configs/_base_/datasets/pascal_context.py

@@ -0,0 +1,60 @@
+# dataset settings
+dataset_type = 'PascalContextDataset'
+data_root = 'data/VOCdevkit/VOC2010/'
+img_norm_cfg = dict(
+    mean=[123.675, 116.28, 103.53], std=[58.395, 57.12, 57.375], to_rgb=True)
+
+img_scale = (520, 520)
+crop_size = (480, 480)
+
+train_pipeline = [
+    dict(type='LoadImageFromFile'),
+    dict(type='LoadAnnotations'),
+    dict(type='Resize', img_scale=img_scale, ratio_range=(0.5, 2.0)),
+    dict(type='RandomCrop', crop_size=crop_size, cat_max_ratio=0.75),
+    dict(type='RandomFlip', prob=0.5),
+    dict(type='PhotoMetricDistortion'),
+    dict(type='Normalize', **img_norm_cfg),
+    dict(type='Pad', size=crop_size, pad_val=0, seg_pad_val=255),
+    dict(type='DefaultFormatBundle'),
+    dict(type='Collect', keys=['img', 'gt_semantic_seg']),
+]
+test_pipeline = [
+    dict(type='LoadImageFromFile'),
+    dict(
+        type='MultiScaleFlipAug',
+        img_scale=img_scale,
+        # img_ratios=[0.5, 0.75, 1.0, 1.25, 1.5, 1.75],
+        flip=False,
+        transforms=[
+            dict(type='Resize', keep_ratio=True),
+            dict(type='RandomFlip'),
+            dict(type='Normalize', **img_norm_cfg),
+            dict(type='ImageToTensor', keys=['img']),
+            dict(type='Collect', keys=['img']),
+        ])
+]
+data = dict(
+    samples_per_gpu=4,
+    workers_per_gpu=4,
+    train=dict(
+        type=dataset_type,
+        data_root=data_root,
+        img_dir='JPEGImages',
+        ann_dir='SegmentationClassContext',
+        split='ImageSets/SegmentationContext/train.txt',
+        pipeline=train_pipeline),
+    val=dict(
+        type=dataset_type,
+        data_root=data_root,
+        img_dir='JPEGImages',
+        ann_dir='SegmentationClassContext',
+        split='ImageSets/SegmentationContext/val.txt',
+        pipeline=test_pipeline),
+    test=dict(
+        type=dataset_type,
+        data_root=data_root,
+        img_dir='JPEGImages',
+        ann_dir='SegmentationClassContext',
+        split='ImageSets/SegmentationContext/val.txt',
+        pipeline=test_pipeline))

lavis/common/annotator/uniformer/configs/_base_/datasets/pascal_context_59.py

@@ -0,0 +1,60 @@
+# dataset settings
+dataset_type = 'PascalContextDataset59'
+data_root = 'data/VOCdevkit/VOC2010/'
+img_norm_cfg = dict(
+    mean=[123.675, 116.28, 103.53], std=[58.395, 57.12, 57.375], to_rgb=True)
+
+img_scale = (520, 520)
+crop_size = (480, 480)
+
+train_pipeline = [
+    dict(type='LoadImageFromFile'),
+    dict(type='LoadAnnotations', reduce_zero_label=True),
+    dict(type='Resize', img_scale=img_scale, ratio_range=(0.5, 2.0)),
+    dict(type='RandomCrop', crop_size=crop_size, cat_max_ratio=0.75),
+    dict(type='RandomFlip', prob=0.5),
+    dict(type='PhotoMetricDistortion'),
+    dict(type='Normalize', **img_norm_cfg),
+    dict(type='Pad', size=crop_size, pad_val=0, seg_pad_val=255),
+    dict(type='DefaultFormatBundle'),
+    dict(type='Collect', keys=['img', 'gt_semantic_seg']),
+]
+test_pipeline = [
+    dict(type='LoadImageFromFile'),
+    dict(
+        type='MultiScaleFlipAug',
+        img_scale=img_scale,
+        # img_ratios=[0.5, 0.75, 1.0, 1.25, 1.5, 1.75],
+        flip=False,
+        transforms=[
+            dict(type='Resize', keep_ratio=True),
+            dict(type='RandomFlip'),
+            dict(type='Normalize', **img_norm_cfg),
+            dict(type='ImageToTensor', keys=['img']),
+            dict(type='Collect', keys=['img']),
+        ])
+]
+data = dict(
+    samples_per_gpu=4,
+    workers_per_gpu=4,
+    train=dict(
+        type=dataset_type,
+        data_root=data_root,
+        img_dir='JPEGImages',
+        ann_dir='SegmentationClassContext',
+        split='ImageSets/SegmentationContext/train.txt',
+        pipeline=train_pipeline),
+    val=dict(
+        type=dataset_type,
+        data_root=data_root,
+        img_dir='JPEGImages',
+        ann_dir='SegmentationClassContext',
+        split='ImageSets/SegmentationContext/val.txt',
+        pipeline=test_pipeline),
+    test=dict(
+        type=dataset_type,
+        data_root=data_root,
+        img_dir='JPEGImages',
+        ann_dir='SegmentationClassContext',
+        split='ImageSets/SegmentationContext/val.txt',
+        pipeline=test_pipeline))

lavis/common/annotator/uniformer/configs/_base_/datasets/pascal_voc12.py

@@ -0,0 +1,57 @@
+# dataset settings
+dataset_type = 'PascalVOCDataset'
+data_root = 'data/VOCdevkit/VOC2012'
+img_norm_cfg = dict(
+    mean=[123.675, 116.28, 103.53], std=[58.395, 57.12, 57.375], to_rgb=True)
+crop_size = (512, 512)
+train_pipeline = [
+    dict(type='LoadImageFromFile'),
+    dict(type='LoadAnnotations'),
+    dict(type='Resize', img_scale=(2048, 512), ratio_range=(0.5, 2.0)),
+    dict(type='RandomCrop', crop_size=crop_size, cat_max_ratio=0.75),
+    dict(type='RandomFlip', prob=0.5),
+    dict(type='PhotoMetricDistortion'),
+    dict(type='Normalize', **img_norm_cfg),
+    dict(type='Pad', size=crop_size, pad_val=0, seg_pad_val=255),
+    dict(type='DefaultFormatBundle'),
+    dict(type='Collect', keys=['img', 'gt_semantic_seg']),
+]
+test_pipeline = [
+    dict(type='LoadImageFromFile'),
+    dict(
+        type='MultiScaleFlipAug',
+        img_scale=(2048, 512),
+        # img_ratios=[0.5, 0.75, 1.0, 1.25, 1.5, 1.75],
+        flip=False,
+        transforms=[
+            dict(type='Resize', keep_ratio=True),
+            dict(type='RandomFlip'),
+            dict(type='Normalize', **img_norm_cfg),
+            dict(type='ImageToTensor', keys=['img']),
+            dict(type='Collect', keys=['img']),
+        ])
+]
+data = dict(
+    samples_per_gpu=4,
+    workers_per_gpu=4,
+    train=dict(
+        type=dataset_type,
+        data_root=data_root,
+        img_dir='JPEGImages',
+        ann_dir='SegmentationClass',
+        split='ImageSets/Segmentation/train.txt',
+        pipeline=train_pipeline),
+    val=dict(
+        type=dataset_type,
+        data_root=data_root,
+        img_dir='JPEGImages',
+        ann_dir='SegmentationClass',
+        split='ImageSets/Segmentation/val.txt',
+        pipeline=test_pipeline),
+    test=dict(
+        type=dataset_type,
+        data_root=data_root,
+        img_dir='JPEGImages',
+        ann_dir='SegmentationClass',
+        split='ImageSets/Segmentation/val.txt',
+        pipeline=test_pipeline))

lavis/common/annotator/uniformer/configs/_base_/datasets/pascal_voc12_aug.py

@@ -0,0 +1,9 @@
+_base_ = './pascal_voc12.py'
+# dataset settings
+data = dict(
+    train=dict(
+        ann_dir=['SegmentationClass', 'SegmentationClassAug'],
+        split=[
+            'ImageSets/Segmentation/train.txt',
+            'ImageSets/Segmentation/aug.txt'
+        ]))

lavis/common/annotator/uniformer/configs/_base_/datasets/stare.py

@@ -0,0 +1,59 @@
+# dataset settings
+dataset_type = 'STAREDataset'
+data_root = 'data/STARE'
+img_norm_cfg = dict(
+    mean=[123.675, 116.28, 103.53], std=[58.395, 57.12, 57.375], to_rgb=True)
+img_scale = (605, 700)
+crop_size = (128, 128)
+train_pipeline = [
+    dict(type='LoadImageFromFile'),
+    dict(type='LoadAnnotations'),
+    dict(type='Resize', img_scale=img_scale, ratio_range=(0.5, 2.0)),
+    dict(type='RandomCrop', crop_size=crop_size, cat_max_ratio=0.75),
+    dict(type='RandomFlip', prob=0.5),
+    dict(type='PhotoMetricDistortion'),
+    dict(type='Normalize', **img_norm_cfg),
+    dict(type='Pad', size=crop_size, pad_val=0, seg_pad_val=255),
+    dict(type='DefaultFormatBundle'),
+    dict(type='Collect', keys=['img', 'gt_semantic_seg'])
+]
+test_pipeline = [
+    dict(type='LoadImageFromFile'),
+    dict(
+        type='MultiScaleFlipAug',
+        img_scale=img_scale,
+        # img_ratios=[0.5, 0.75, 1.0, 1.25, 1.5, 1.75, 2.0],
+        flip=False,
+        transforms=[
+            dict(type='Resize', keep_ratio=True),
+            dict(type='RandomFlip'),
+            dict(type='Normalize', **img_norm_cfg),
+            dict(type='ImageToTensor', keys=['img']),
+            dict(type='Collect', keys=['img'])
+        ])
+]
+
+data = dict(
+    samples_per_gpu=4,
+    workers_per_gpu=4,
+    train=dict(
+        type='RepeatDataset',
+        times=40000,
+        dataset=dict(
+            type=dataset_type,
+            data_root=data_root,
+            img_dir='images/training',
+            ann_dir='annotations/training',
+            pipeline=train_pipeline)),
+    val=dict(
+        type=dataset_type,
+        data_root=data_root,
+        img_dir='images/validation',
+        ann_dir='annotations/validation',
+        pipeline=test_pipeline),
+    test=dict(
+        type=dataset_type,
+        data_root=data_root,
+        img_dir='images/validation',
+        ann_dir='annotations/validation',
+        pipeline=test_pipeline))

lavis/common/annotator/uniformer/configs/_base_/default_runtime.py

@@ -0,0 +1,14 @@
+# yapf:disable
+log_config = dict(
+    interval=50,
+    hooks=[
+        dict(type='TextLoggerHook', by_epoch=False),
+        # dict(type='TensorboardLoggerHook')
+    ])
+# yapf:enable
+dist_params = dict(backend='nccl')
+log_level = 'INFO'
+load_from = None
+resume_from = None
+workflow = [('train', 1)]
+cudnn_benchmark = True

lavis/common/annotator/uniformer/configs/_base_/models/ann_r50-d8.py

@@ -0,0 +1,46 @@
+# model settings
+norm_cfg = dict(type='SyncBN', requires_grad=True)
+model = dict(
+    type='EncoderDecoder',
+    pretrained='open-mmlab://resnet50_v1c',
+    backbone=dict(
+        type='ResNetV1c',
+        depth=50,
+        num_stages=4,
+        out_indices=(0, 1, 2, 3),
+        dilations=(1, 1, 2, 4),
+        strides=(1, 2, 1, 1),
+        norm_cfg=norm_cfg,
+        norm_eval=False,
+        style='pytorch',
+        contract_dilation=True),
+    decode_head=dict(
+        type='ANNHead',
+        in_channels=[1024, 2048],
+        in_index=[2, 3],
+        channels=512,
+        project_channels=256,
+        query_scales=(1, ),
+        key_pool_scales=(1, 3, 6, 8),
+        dropout_ratio=0.1,
+        num_classes=19,
+        norm_cfg=norm_cfg,
+        align_corners=False,
+        loss_decode=dict(
+            type='CrossEntropyLoss', use_sigmoid=False, loss_weight=1.0)),
+    auxiliary_head=dict(
+        type='FCNHead',
+        in_channels=1024,
+        in_index=2,
+        channels=256,
+        num_convs=1,
+        concat_input=False,
+        dropout_ratio=0.1,
+        num_classes=19,
+        norm_cfg=norm_cfg,
+        align_corners=False,
+        loss_decode=dict(
+            type='CrossEntropyLoss', use_sigmoid=False, loss_weight=0.4)),
+    # model training and testing settings
+    train_cfg=dict(),
+    test_cfg=dict(mode='whole'))

lavis/common/annotator/uniformer/configs/_base_/models/apcnet_r50-d8.py

@@ -0,0 +1,44 @@
+# model settings
+norm_cfg = dict(type='SyncBN', requires_grad=True)
+model = dict(
+    type='EncoderDecoder',
+    pretrained='open-mmlab://resnet50_v1c',
+    backbone=dict(
+        type='ResNetV1c',
+        depth=50,
+        num_stages=4,
+        out_indices=(0, 1, 2, 3),
+        dilations=(1, 1, 2, 4),
+        strides=(1, 2, 1, 1),
+        norm_cfg=norm_cfg,
+        norm_eval=False,
+        style='pytorch',
+        contract_dilation=True),
+    decode_head=dict(
+        type='APCHead',
+        in_channels=2048,
+        in_index=3,
+        channels=512,
+        pool_scales=(1, 2, 3, 6),
+        dropout_ratio=0.1,
+        num_classes=19,
+        norm_cfg=dict(type='SyncBN', requires_grad=True),
+        align_corners=False,
+        loss_decode=dict(
+            type='CrossEntropyLoss', use_sigmoid=False, loss_weight=1.0)),
+    auxiliary_head=dict(
+        type='FCNHead',
+        in_channels=1024,
+        in_index=2,
+        channels=256,
+        num_convs=1,
+        concat_input=False,
+        dropout_ratio=0.1,
+        num_classes=19,
+        norm_cfg=norm_cfg,
+        align_corners=False,
+        loss_decode=dict(
+            type='CrossEntropyLoss', use_sigmoid=False, loss_weight=0.4)),
+    # model training and testing settings
+    train_cfg=dict(),
+    test_cfg=dict(mode='whole'))

lavis/common/annotator/uniformer/configs/_base_/models/ccnet_r50-d8.py

@@ -0,0 +1,44 @@
+# model settings
+norm_cfg = dict(type='SyncBN', requires_grad=True)
+model = dict(
+    type='EncoderDecoder',
+    pretrained='open-mmlab://resnet50_v1c',
+    backbone=dict(
+        type='ResNetV1c',
+        depth=50,
+        num_stages=4,
+        out_indices=(0, 1, 2, 3),
+        dilations=(1, 1, 2, 4),
+        strides=(1, 2, 1, 1),
+        norm_cfg=norm_cfg,
+        norm_eval=False,
+        style='pytorch',
+        contract_dilation=True),
+    decode_head=dict(
+        type='CCHead',
+        in_channels=2048,
+        in_index=3,
+        channels=512,
+        recurrence=2,
+        dropout_ratio=0.1,
+        num_classes=19,
+        norm_cfg=norm_cfg,
+        align_corners=False,
+        loss_decode=dict(
+            type='CrossEntropyLoss', use_sigmoid=False, loss_weight=1.0)),
+    auxiliary_head=dict(
+        type='FCNHead',
+        in_channels=1024,
+        in_index=2,
+        channels=256,
+        num_convs=1,
+        concat_input=False,
+        dropout_ratio=0.1,
+        num_classes=19,
+        norm_cfg=norm_cfg,
+        align_corners=False,
+        loss_decode=dict(
+            type='CrossEntropyLoss', use_sigmoid=False, loss_weight=0.4)),
+    # model training and testing settings
+    train_cfg=dict(),
+    test_cfg=dict(mode='whole'))

lavis/common/annotator/uniformer/configs/_base_/models/cgnet.py

@@ -0,0 +1,35 @@
+# model settings
+norm_cfg = dict(type='SyncBN', eps=1e-03, requires_grad=True)
+model = dict(
+    type='EncoderDecoder',
+    backbone=dict(
+        type='CGNet',
+        norm_cfg=norm_cfg,
+        in_channels=3,
+        num_channels=(32, 64, 128),
+        num_blocks=(3, 21),
+        dilations=(2, 4),
+        reductions=(8, 16)),
+    decode_head=dict(
+        type='FCNHead',
+        in_channels=256,
+        in_index=2,
+        channels=256,
+        num_convs=0,
+        concat_input=False,
+        dropout_ratio=0,
+        num_classes=19,
+        norm_cfg=norm_cfg,
+        loss_decode=dict(
+            type='CrossEntropyLoss',
+            use_sigmoid=False,
+            loss_weight=1.0,
+            class_weight=[
+                2.5959933, 6.7415504, 3.5354059, 9.8663225, 9.690899, 9.369352,
+                10.289121, 9.953208, 4.3097677, 9.490387, 7.674431, 9.396905,
+                10.347791, 6.3927646, 10.226669, 10.241062, 10.280587,
+                10.396974, 10.055647
+            ])),
+    # model training and testing settings
+    train_cfg=dict(sampler=None),
+    test_cfg=dict(mode='whole'))

lavis/common/annotator/uniformer/configs/_base_/models/danet_r50-d8.py

@@ -0,0 +1,44 @@
+# model settings
+norm_cfg = dict(type='SyncBN', requires_grad=True)
+model = dict(
+    type='EncoderDecoder',
+    pretrained='open-mmlab://resnet50_v1c',
+    backbone=dict(
+        type='ResNetV1c',
+        depth=50,
+        num_stages=4,
+        out_indices=(0, 1, 2, 3),
+        dilations=(1, 1, 2, 4),
+        strides=(1, 2, 1, 1),
+        norm_cfg=norm_cfg,
+        norm_eval=False,
+        style='pytorch',
+        contract_dilation=True),
+    decode_head=dict(
+        type='DAHead',
+        in_channels=2048,
+        in_index=3,
+        channels=512,
+        pam_channels=64,
+        dropout_ratio=0.1,
+        num_classes=19,
+        norm_cfg=norm_cfg,
+        align_corners=False,
+        loss_decode=dict(
+            type='CrossEntropyLoss', use_sigmoid=False, loss_weight=1.0)),
+    auxiliary_head=dict(
+        type='FCNHead',
+        in_channels=1024,
+        in_index=2,
+        channels=256,
+        num_convs=1,
+        concat_input=False,
+        dropout_ratio=0.1,
+        num_classes=19,
+        norm_cfg=norm_cfg,
+        align_corners=False,
+        loss_decode=dict(
+            type='CrossEntropyLoss', use_sigmoid=False, loss_weight=0.4)),
+    # model training and testing settings
+    train_cfg=dict(),
+    test_cfg=dict(mode='whole'))

lavis/common/annotator/uniformer/configs/_base_/models/deeplabv3_r50-d8.py

@@ -0,0 +1,44 @@
+# model settings
+norm_cfg = dict(type='SyncBN', requires_grad=True)
+model = dict(
+    type='EncoderDecoder',
+    pretrained='open-mmlab://resnet50_v1c',
+    backbone=dict(
+        type='ResNetV1c',
+        depth=50,
+        num_stages=4,
+        out_indices=(0, 1, 2, 3),
+        dilations=(1, 1, 2, 4),
+        strides=(1, 2, 1, 1),
+        norm_cfg=norm_cfg,
+        norm_eval=False,
+        style='pytorch',
+        contract_dilation=True),
+    decode_head=dict(
+        type='ASPPHead',
+        in_channels=2048,
+        in_index=3,
+        channels=512,
+        dilations=(1, 12, 24, 36),
+        dropout_ratio=0.1,
+        num_classes=19,
+        norm_cfg=norm_cfg,
+        align_corners=False,
+        loss_decode=dict(
+            type='CrossEntropyLoss', use_sigmoid=False, loss_weight=1.0)),
+    auxiliary_head=dict(
+        type='FCNHead',
+        in_channels=1024,
+        in_index=2,
+        channels=256,
+        num_convs=1,
+        concat_input=False,
+        dropout_ratio=0.1,
+        num_classes=19,
+        norm_cfg=norm_cfg,
+        align_corners=False,
+        loss_decode=dict(
+            type='CrossEntropyLoss', use_sigmoid=False, loss_weight=0.4)),
+    # model training and testing settings
+    train_cfg=dict(),
+    test_cfg=dict(mode='whole'))

lavis/common/annotator/uniformer/configs/_base_/models/deeplabv3_unet_s5-d16.py

@@ -0,0 +1,50 @@
+# model settings
+norm_cfg = dict(type='SyncBN', requires_grad=True)
+model = dict(
+    type='EncoderDecoder',
+    pretrained=None,
+    backbone=dict(
+        type='UNet',
+        in_channels=3,
+        base_channels=64,
+        num_stages=5,
+        strides=(1, 1, 1, 1, 1),
+        enc_num_convs=(2, 2, 2, 2, 2),
+        dec_num_convs=(2, 2, 2, 2),
+        downsamples=(True, True, True, True),
+        enc_dilations=(1, 1, 1, 1, 1),
+        dec_dilations=(1, 1, 1, 1),
+        with_cp=False,
+        conv_cfg=None,
+        norm_cfg=norm_cfg,
+        act_cfg=dict(type='ReLU'),
+        upsample_cfg=dict(type='InterpConv'),
+        norm_eval=False),
+    decode_head=dict(
+        type='ASPPHead',
+        in_channels=64,
+        in_index=4,
+        channels=16,
+        dilations=(1, 12, 24, 36),
+        dropout_ratio=0.1,
+        num_classes=2,
+        norm_cfg=norm_cfg,
+        align_corners=False,
+        loss_decode=dict(
+            type='CrossEntropyLoss', use_sigmoid=False, loss_weight=1.0)),
+    auxiliary_head=dict(
+        type='FCNHead',
+        in_channels=128,
+        in_index=3,
+        channels=64,
+        num_convs=1,
+        concat_input=False,
+        dropout_ratio=0.1,
+        num_classes=2,
+        norm_cfg=norm_cfg,
+        align_corners=False,
+        loss_decode=dict(
+            type='CrossEntropyLoss', use_sigmoid=False, loss_weight=0.4)),
+    # model training and testing settings
+    train_cfg=dict(),
+    test_cfg=dict(mode='slide', crop_size=256, stride=170))

lavis/common/annotator/uniformer/configs/_base_/models/deeplabv3plus_r50-d8.py

@@ -0,0 +1,46 @@
+# model settings
+norm_cfg = dict(type='SyncBN', requires_grad=True)
+model = dict(
+    type='EncoderDecoder',
+    pretrained='open-mmlab://resnet50_v1c',
+    backbone=dict(
+        type='ResNetV1c',
+        depth=50,
+        num_stages=4,
+        out_indices=(0, 1, 2, 3),
+        dilations=(1, 1, 2, 4),
+        strides=(1, 2, 1, 1),
+        norm_cfg=norm_cfg,
+        norm_eval=False,
+        style='pytorch',
+        contract_dilation=True),
+    decode_head=dict(
+        type='DepthwiseSeparableASPPHead',
+        in_channels=2048,
+        in_index=3,
+        channels=512,
+        dilations=(1, 12, 24, 36),
+        c1_in_channels=256,
+        c1_channels=48,
+        dropout_ratio=0.1,
+        num_classes=19,
+        norm_cfg=norm_cfg,
+        align_corners=False,
+        loss_decode=dict(
+            type='CrossEntropyLoss', use_sigmoid=False, loss_weight=1.0)),
+    auxiliary_head=dict(
+        type='FCNHead',
+        in_channels=1024,
+        in_index=2,
+        channels=256,
+        num_convs=1,
+        concat_input=False,
+        dropout_ratio=0.1,
+        num_classes=19,
+        norm_cfg=norm_cfg,
+        align_corners=False,
+        loss_decode=dict(
+            type='CrossEntropyLoss', use_sigmoid=False, loss_weight=0.4)),
+    # model training and testing settings
+    train_cfg=dict(),
+    test_cfg=dict(mode='whole'))

lavis/common/annotator/uniformer/configs/_base_/models/dmnet_r50-d8.py

@@ -0,0 +1,44 @@
+# model settings
+norm_cfg = dict(type='SyncBN', requires_grad=True)
+model = dict(
+    type='EncoderDecoder',
+    pretrained='open-mmlab://resnet50_v1c',
+    backbone=dict(
+        type='ResNetV1c',
+        depth=50,
+        num_stages=4,
+        out_indices=(0, 1, 2, 3),
+        dilations=(1, 1, 2, 4),
+        strides=(1, 2, 1, 1),
+        norm_cfg=norm_cfg,
+        norm_eval=False,
+        style='pytorch',
+        contract_dilation=True),
+    decode_head=dict(
+        type='DMHead',
+        in_channels=2048,
+        in_index=3,
+        channels=512,
+        filter_sizes=(1, 3, 5, 7),
+        dropout_ratio=0.1,
+        num_classes=19,
+        norm_cfg=dict(type='SyncBN', requires_grad=True),
+        align_corners=False,
+        loss_decode=dict(
+            type='CrossEntropyLoss', use_sigmoid=False, loss_weight=1.0)),
+    auxiliary_head=dict(
+        type='FCNHead',
+        in_channels=1024,
+        in_index=2,
+        channels=256,
+        num_convs=1,
+        concat_input=False,
+        dropout_ratio=0.1,
+        num_classes=19,
+        norm_cfg=norm_cfg,
+        align_corners=False,
+        loss_decode=dict(
+            type='CrossEntropyLoss', use_sigmoid=False, loss_weight=0.4)),
+    # model training and testing settings
+    train_cfg=dict(),
+    test_cfg=dict(mode='whole'))

lavis/common/annotator/uniformer/configs/_base_/models/dnl_r50-d8.py

@@ -0,0 +1,46 @@
+# model settings
+norm_cfg = dict(type='SyncBN', requires_grad=True)
+model = dict(
+    type='EncoderDecoder',
+    pretrained='open-mmlab://resnet50_v1c',
+    backbone=dict(
+        type='ResNetV1c',
+        depth=50,
+        num_stages=4,
+        out_indices=(0, 1, 2, 3),
+        dilations=(1, 1, 2, 4),
+        strides=(1, 2, 1, 1),
+        norm_cfg=norm_cfg,
+        norm_eval=False,
+        style='pytorch',
+        contract_dilation=True),
+    decode_head=dict(
+        type='DNLHead',
+        in_channels=2048,
+        in_index=3,
+        channels=512,
+        dropout_ratio=0.1,
+        reduction=2,
+        use_scale=True,
+        mode='embedded_gaussian',
+        num_classes=19,
+        norm_cfg=norm_cfg,
+        align_corners=False,
+        loss_decode=dict(
+            type='CrossEntropyLoss', use_sigmoid=False, loss_weight=1.0)),
+    auxiliary_head=dict(
+        type='FCNHead',
+        in_channels=1024,
+        in_index=2,
+        channels=256,
+        num_convs=1,
+        concat_input=False,
+        dropout_ratio=0.1,
+        num_classes=19,
+        norm_cfg=norm_cfg,
+        align_corners=False,
+        loss_decode=dict(
+            type='CrossEntropyLoss', use_sigmoid=False, loss_weight=0.4)),
+    # model training and testing settings
+    train_cfg=dict(),
+    test_cfg=dict(mode='whole'))

lavis/common/annotator/uniformer/configs/_base_/models/emanet_r50-d8.py

@@ -0,0 +1,47 @@
+# model settings
+norm_cfg = dict(type='SyncBN', requires_grad=True)
+model = dict(
+    type='EncoderDecoder',
+    pretrained='open-mmlab://resnet50_v1c',
+    backbone=dict(
+        type='ResNetV1c',
+        depth=50,
+        num_stages=4,
+        out_indices=(0, 1, 2, 3),
+        dilations=(1, 1, 2, 4),
+        strides=(1, 2, 1, 1),
+        norm_cfg=norm_cfg,
+        norm_eval=False,
+        style='pytorch',
+        contract_dilation=True),
+    decode_head=dict(
+        type='EMAHead',
+        in_channels=2048,
+        in_index=3,
+        channels=256,
+        ema_channels=512,
+        num_bases=64,
+        num_stages=3,
+        momentum=0.1,
+        dropout_ratio=0.1,
+        num_classes=19,
+        norm_cfg=norm_cfg,
+        align_corners=False,
+        loss_decode=dict(
+            type='CrossEntropyLoss', use_sigmoid=False, loss_weight=1.0)),
+    auxiliary_head=dict(
+        type='FCNHead',
+        in_channels=1024,
+        in_index=2,
+        channels=256,
+        num_convs=1,
+        concat_input=False,
+        dropout_ratio=0.1,
+        num_classes=19,
+        norm_cfg=norm_cfg,
+        align_corners=False,
+        loss_decode=dict(
+            type='CrossEntropyLoss', use_sigmoid=False, loss_weight=0.4)),
+    # model training and testing settings
+    train_cfg=dict(),
+    test_cfg=dict(mode='whole'))

lavis/common/annotator/uniformer/configs/_base_/models/encnet_r50-d8.py

@@ -0,0 +1,48 @@
+# model settings
+norm_cfg = dict(type='SyncBN', requires_grad=True)
+model = dict(
+    type='EncoderDecoder',
+    pretrained='open-mmlab://resnet50_v1c',
+    backbone=dict(
+        type='ResNetV1c',
+        depth=50,
+        num_stages=4,
+        out_indices=(0, 1, 2, 3),
+        dilations=(1, 1, 2, 4),
+        strides=(1, 2, 1, 1),
+        norm_cfg=norm_cfg,
+        norm_eval=False,
+        style='pytorch',
+        contract_dilation=True),
+    decode_head=dict(
+        type='EncHead',
+        in_channels=[512, 1024, 2048],
+        in_index=(1, 2, 3),
+        channels=512,
+        num_codes=32,
+        use_se_loss=True,
+        add_lateral=False,
+        dropout_ratio=0.1,
+        num_classes=19,
+        norm_cfg=norm_cfg,
+        align_corners=False,
+        loss_decode=dict(
+            type='CrossEntropyLoss', use_sigmoid=False, loss_weight=1.0),
+        loss_se_decode=dict(
+            type='CrossEntropyLoss', use_sigmoid=True, loss_weight=0.2)),
+    auxiliary_head=dict(
+        type='FCNHead',
+        in_channels=1024,
+        in_index=2,
+        channels=256,
+        num_convs=1,
+        concat_input=False,
+        dropout_ratio=0.1,
+        num_classes=19,
+        norm_cfg=norm_cfg,
+        align_corners=False,
+        loss_decode=dict(
+            type='CrossEntropyLoss', use_sigmoid=False, loss_weight=0.4)),
+    # model training and testing settings
+    train_cfg=dict(),
+    test_cfg=dict(mode='whole'))

lavis/common/annotator/uniformer/configs/_base_/models/fast_scnn.py

@@ -0,0 +1,57 @@
+# model settings
+norm_cfg = dict(type='SyncBN', requires_grad=True, momentum=0.01)
+model = dict(
+    type='EncoderDecoder',
+    backbone=dict(
+        type='FastSCNN',
+        downsample_dw_channels=(32, 48),
+        global_in_channels=64,
+        global_block_channels=(64, 96, 128),
+        global_block_strides=(2, 2, 1),
+        global_out_channels=128,
+        higher_in_channels=64,
+        lower_in_channels=128,
+        fusion_out_channels=128,
+        out_indices=(0, 1, 2),
+        norm_cfg=norm_cfg,
+        align_corners=False),
+    decode_head=dict(
+        type='DepthwiseSeparableFCNHead',
+        in_channels=128,
+        channels=128,
+        concat_input=False,
+        num_classes=19,
+        in_index=-1,
+        norm_cfg=norm_cfg,
+        align_corners=False,
+        loss_decode=dict(
+            type='CrossEntropyLoss', use_sigmoid=True, loss_weight=0.4)),
+    auxiliary_head=[
+        dict(
+            type='FCNHead',
+            in_channels=128,
+            channels=32,
+            num_convs=1,
+            num_classes=19,
+            in_index=-2,
+            norm_cfg=norm_cfg,
+            concat_input=False,
+            align_corners=False,
+            loss_decode=dict(
+                type='CrossEntropyLoss', use_sigmoid=True, loss_weight=0.4)),
+        dict(
+            type='FCNHead',
+            in_channels=64,
+            channels=32,
+            num_convs=1,
+            num_classes=19,
+            in_index=-3,
+            norm_cfg=norm_cfg,
+            concat_input=False,
+            align_corners=False,
+            loss_decode=dict(
+                type='CrossEntropyLoss', use_sigmoid=True, loss_weight=0.4)),
+    ],
+    # model training and testing settings
+    train_cfg=dict(),
+    test_cfg=dict(mode='whole'))