Spaces:
Running
on
Zero
Running
on
Zero
SunderAli17
commited on
Create module/diffusers_vae/autoencoder_kl.py
Browse files
module/diffusers_vae/autoencoder_kl.py
ADDED
@@ -0,0 +1,467 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Copyright 2023 The HuggingFace Team. All rights reserved.
|
2 |
+
#
|
3 |
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
4 |
+
# you may not use this file except in compliance with the License.
|
5 |
+
# You may obtain a copy of the License at
|
6 |
+
#
|
7 |
+
# http://www.apache.org/licenses/LICENSE-2.0
|
8 |
+
#
|
9 |
+
# Unless required by applicable law or agreed to in writing, software
|
10 |
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
11 |
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
12 |
+
# See the License for the specific language governing permissions and
|
13 |
+
# limitations under the License.
|
14 |
+
from typing import Dict, Optional, Tuple, Union
|
15 |
+
|
16 |
+
import torch
|
17 |
+
import torch.nn as nn
|
18 |
+
|
19 |
+
from diffusers.configuration_utils import ConfigMixin, register_to_config
|
20 |
+
from diffusers.loaders import FromOriginalVAEMixin
|
21 |
+
from diffusers.utils.accelerate_utils import apply_forward_hook
|
22 |
+
from diffusers.models.attention_processor import (
|
23 |
+
ADDED_KV_ATTENTION_PROCESSORS,
|
24 |
+
CROSS_ATTENTION_PROCESSORS,
|
25 |
+
Attention,
|
26 |
+
AttentionProcessor,
|
27 |
+
AttnAddedKVProcessor,
|
28 |
+
AttnProcessor,
|
29 |
+
)
|
30 |
+
from diffusers.models.modeling_outputs import AutoencoderKLOutput
|
31 |
+
from diffusers.models.modeling_utils import ModelMixin
|
32 |
+
from .vae import Decoder, DecoderOutput, DiagonalGaussianDistribution, Encoder
|
33 |
+
|
34 |
+
|
35 |
+
class AutoencoderKL(ModelMixin, ConfigMixin, FromOriginalVAEMixin):
|
36 |
+
r"""
|
37 |
+
A VAE model with KL loss for encoding images into latents and decoding latent representations into images.
|
38 |
+
This model inherits from [`ModelMixin`]. Check the superclass documentation for it's generic methods implemented
|
39 |
+
for all models (such as downloading or saving).
|
40 |
+
Parameters:
|
41 |
+
in_channels (int, *optional*, defaults to 3): Number of channels in the input image.
|
42 |
+
out_channels (int, *optional*, defaults to 3): Number of channels in the output.
|
43 |
+
down_block_types (`Tuple[str]`, *optional*, defaults to `("DownEncoderBlock2D",)`):
|
44 |
+
Tuple of downsample block types.
|
45 |
+
up_block_types (`Tuple[str]`, *optional*, defaults to `("UpDecoderBlock2D",)`):
|
46 |
+
Tuple of upsample block types.
|
47 |
+
block_out_channels (`Tuple[int]`, *optional*, defaults to `(64,)`):
|
48 |
+
Tuple of block output channels.
|
49 |
+
act_fn (`str`, *optional*, defaults to `"silu"`): The activation function to use.
|
50 |
+
latent_channels (`int`, *optional*, defaults to 4): Number of channels in the latent space.
|
51 |
+
sample_size (`int`, *optional*, defaults to `32`): Sample input size.
|
52 |
+
scaling_factor (`float`, *optional*, defaults to 0.18215):
|
53 |
+
The component-wise standard deviation of the trained latent space computed using the first batch of the
|
54 |
+
training set. This is used to scale the latent space to have unit variance when training the diffusion
|
55 |
+
model. The latents are scaled with the formula `z = z * scaling_factor` before being passed to the
|
56 |
+
diffusion model. When decoding, the latents are scaled back to the original scale with the formula: `z = 1
|
57 |
+
/ scaling_factor * z`. For more details, refer to sections 4.3.2 and D.1 of the [High-Resolution Image
|
58 |
+
Synthesis with Latent Diffusion Models](https://arxiv.org/abs/2112.10752) paper.
|
59 |
+
force_upcast (`bool`, *optional*, default to `True`):
|
60 |
+
If enabled it will force the VAE to run in float32 for high image resolution pipelines, such as SD-XL. VAE
|
61 |
+
can be fine-tuned / trained to a lower range without loosing too much precision in which case
|
62 |
+
`force_upcast` can be set to `False` - see: https://huggingface.co/madebyollin/sdxl-vae-fp16-fix
|
63 |
+
"""
|
64 |
+
|
65 |
+
_supports_gradient_checkpointing = True
|
66 |
+
|
67 |
+
@register_to_config
|
68 |
+
def __init__(
|
69 |
+
self,
|
70 |
+
in_channels: int = 3,
|
71 |
+
out_channels: int = 3,
|
72 |
+
down_block_types: Tuple[str] = ("DownEncoderBlock2D",),
|
73 |
+
up_block_types: Tuple[str] = ("UpDecoderBlock2D",),
|
74 |
+
block_out_channels: Tuple[int] = (64,),
|
75 |
+
layers_per_block: int = 1,
|
76 |
+
act_fn: str = "silu",
|
77 |
+
latent_channels: int = 4,
|
78 |
+
norm_num_groups: int = 32,
|
79 |
+
sample_size: int = 32,
|
80 |
+
scaling_factor: float = 0.18215,
|
81 |
+
force_upcast: float = True,
|
82 |
+
):
|
83 |
+
super().__init__()
|
84 |
+
|
85 |
+
# pass init params to Encoder
|
86 |
+
self.encoder = Encoder(
|
87 |
+
in_channels=in_channels,
|
88 |
+
out_channels=latent_channels,
|
89 |
+
down_block_types=down_block_types,
|
90 |
+
block_out_channels=block_out_channels,
|
91 |
+
layers_per_block=layers_per_block,
|
92 |
+
act_fn=act_fn,
|
93 |
+
norm_num_groups=norm_num_groups,
|
94 |
+
double_z=True,
|
95 |
+
)
|
96 |
+
|
97 |
+
# pass init params to Decoder
|
98 |
+
self.decoder = Decoder(
|
99 |
+
in_channels=latent_channels,
|
100 |
+
out_channels=out_channels,
|
101 |
+
up_block_types=up_block_types,
|
102 |
+
block_out_channels=block_out_channels,
|
103 |
+
layers_per_block=layers_per_block,
|
104 |
+
norm_num_groups=norm_num_groups,
|
105 |
+
act_fn=act_fn,
|
106 |
+
)
|
107 |
+
|
108 |
+
self.quant_conv = nn.Conv2d(2 * latent_channels, 2 * latent_channels, 1)
|
109 |
+
self.post_quant_conv = nn.Conv2d(latent_channels, latent_channels, 1)
|
110 |
+
|
111 |
+
self.use_slicing = False
|
112 |
+
self.use_tiling = False
|
113 |
+
|
114 |
+
# only relevant if vae tiling is enabled
|
115 |
+
self.tile_sample_min_size = self.config.sample_size
|
116 |
+
sample_size = (
|
117 |
+
self.config.sample_size[0]
|
118 |
+
if isinstance(self.config.sample_size, (list, tuple))
|
119 |
+
else self.config.sample_size
|
120 |
+
)
|
121 |
+
self.tile_latent_min_size = int(sample_size / (2 ** (len(self.config.block_out_channels) - 1)))
|
122 |
+
self.tile_overlap_factor = 0.25
|
123 |
+
|
124 |
+
def _set_gradient_checkpointing(self, module, value=False):
|
125 |
+
if isinstance(module, (Encoder, Decoder)):
|
126 |
+
module.gradient_checkpointing = value
|
127 |
+
|
128 |
+
def enable_tiling(self, use_tiling: bool = True):
|
129 |
+
r"""
|
130 |
+
Enable tiled VAE decoding. When this option is enabled, the VAE will split the input tensor into tiles to
|
131 |
+
compute decoding and encoding in several steps. This is useful for saving a large amount of memory and to allow
|
132 |
+
processing larger images.
|
133 |
+
"""
|
134 |
+
self.use_tiling = use_tiling
|
135 |
+
|
136 |
+
def disable_tiling(self):
|
137 |
+
r"""
|
138 |
+
Disable tiled VAE decoding. If `enable_tiling` was previously enabled, this method will go back to computing
|
139 |
+
decoding in one step.
|
140 |
+
"""
|
141 |
+
self.enable_tiling(False)
|
142 |
+
|
143 |
+
def enable_slicing(self):
|
144 |
+
r"""
|
145 |
+
Enable sliced VAE decoding. When this option is enabled, the VAE will split the input tensor in slices to
|
146 |
+
compute decoding in several steps. This is useful to save some memory and allow larger batch sizes.
|
147 |
+
"""
|
148 |
+
self.use_slicing = True
|
149 |
+
|
150 |
+
def disable_slicing(self):
|
151 |
+
r"""
|
152 |
+
Disable sliced VAE decoding. If `enable_slicing` was previously enabled, this method will go back to computing
|
153 |
+
decoding in one step.
|
154 |
+
"""
|
155 |
+
self.use_slicing = False
|
156 |
+
|
157 |
+
@property
|
158 |
+
# Copied from diffusers.models.unet_2d_condition.UNet2DConditionModel.attn_processors
|
159 |
+
def attn_processors(self) -> Dict[str, AttentionProcessor]:
|
160 |
+
r"""
|
161 |
+
Returns:
|
162 |
+
`dict` of attention processors: A dictionary containing all attention processors used in the model with
|
163 |
+
indexed by its weight name.
|
164 |
+
"""
|
165 |
+
# set recursively
|
166 |
+
processors = {}
|
167 |
+
|
168 |
+
def fn_recursive_add_processors(name: str, module: torch.nn.Module, processors: Dict[str, AttentionProcessor]):
|
169 |
+
if hasattr(module, "get_processor"):
|
170 |
+
processors[f"{name}.processor"] = module.get_processor(return_deprecated_lora=True)
|
171 |
+
|
172 |
+
for sub_name, child in module.named_children():
|
173 |
+
fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)
|
174 |
+
|
175 |
+
return processors
|
176 |
+
|
177 |
+
for name, module in self.named_children():
|
178 |
+
fn_recursive_add_processors(name, module, processors)
|
179 |
+
|
180 |
+
return processors
|
181 |
+
|
182 |
+
# Copied from diffusers.models.unet_2d_condition.UNet2DConditionModel.set_attn_processor
|
183 |
+
def set_attn_processor(
|
184 |
+
self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]], _remove_lora=False
|
185 |
+
):
|
186 |
+
r"""
|
187 |
+
Sets the attention processor to use to compute attention.
|
188 |
+
Parameters:
|
189 |
+
processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
|
190 |
+
The instantiated processor class or a dictionary of processor classes that will be set as the processor
|
191 |
+
for **all** `Attention` layers.
|
192 |
+
If `processor` is a dict, the key needs to define the path to the corresponding cross attention
|
193 |
+
processor. This is strongly recommended when setting trainable attention processors.
|
194 |
+
"""
|
195 |
+
count = len(self.attn_processors.keys())
|
196 |
+
|
197 |
+
if isinstance(processor, dict) and len(processor) != count:
|
198 |
+
raise ValueError(
|
199 |
+
f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
|
200 |
+
f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
|
201 |
+
)
|
202 |
+
|
203 |
+
def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor):
|
204 |
+
if hasattr(module, "set_processor"):
|
205 |
+
if not isinstance(processor, dict):
|
206 |
+
module.set_processor(processor, _remove_lora=_remove_lora)
|
207 |
+
else:
|
208 |
+
module.set_processor(processor.pop(f"{name}.processor"), _remove_lora=_remove_lora)
|
209 |
+
|
210 |
+
for sub_name, child in module.named_children():
|
211 |
+
fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)
|
212 |
+
|
213 |
+
for name, module in self.named_children():
|
214 |
+
fn_recursive_attn_processor(name, module, processor)
|
215 |
+
|
216 |
+
# Copied from diffusers.models.unet_2d_condition.UNet2DConditionModel.set_default_attn_processor
|
217 |
+
def set_default_attn_processor(self):
|
218 |
+
"""
|
219 |
+
Disables custom attention processors and sets the default attention implementation.
|
220 |
+
"""
|
221 |
+
if all(proc.__class__ in ADDED_KV_ATTENTION_PROCESSORS for proc in self.attn_processors.values()):
|
222 |
+
processor = AttnAddedKVProcessor()
|
223 |
+
elif all(proc.__class__ in CROSS_ATTENTION_PROCESSORS for proc in self.attn_processors.values()):
|
224 |
+
processor = AttnProcessor()
|
225 |
+
else:
|
226 |
+
raise ValueError(
|
227 |
+
f"Cannot call `set_default_attn_processor` when attention processors are of type {next(iter(self.attn_processors.values()))}"
|
228 |
+
)
|
229 |
+
|
230 |
+
self.set_attn_processor(processor, _remove_lora=True)
|
231 |
+
|
232 |
+
@apply_forward_hook
|
233 |
+
def encode(
|
234 |
+
self, x: torch.FloatTensor, return_dict: bool = True
|
235 |
+
) -> Union[AutoencoderKLOutput, Tuple[DiagonalGaussianDistribution]]:
|
236 |
+
"""
|
237 |
+
Encode a batch of images into latents.
|
238 |
+
Args:
|
239 |
+
x (`torch.FloatTensor`): Input batch of images.
|
240 |
+
return_dict (`bool`, *optional*, defaults to `True`):
|
241 |
+
Whether to return a [`~models.autoencoder_kl.AutoencoderKLOutput`] instead of a plain tuple.
|
242 |
+
Returns:
|
243 |
+
The latent representations of the encoded images. If `return_dict` is True, a
|
244 |
+
[`~models.autoencoder_kl.AutoencoderKLOutput`] is returned, otherwise a plain `tuple` is returned.
|
245 |
+
"""
|
246 |
+
if self.use_tiling and (x.shape[-1] > self.tile_sample_min_size or x.shape[-2] > self.tile_sample_min_size):
|
247 |
+
return self.tiled_encode(x, return_dict=return_dict)
|
248 |
+
|
249 |
+
if self.use_slicing and x.shape[0] > 1:
|
250 |
+
encoded_slices = [self.encoder(x_slice) for x_slice in x.split(1)]
|
251 |
+
h = torch.cat(encoded_slices)
|
252 |
+
else:
|
253 |
+
h = self.encoder(x)
|
254 |
+
|
255 |
+
moments = self.quant_conv(h)
|
256 |
+
posterior = DiagonalGaussianDistribution(moments)
|
257 |
+
|
258 |
+
if not return_dict:
|
259 |
+
return (posterior,)
|
260 |
+
|
261 |
+
return AutoencoderKLOutput(latent_dist=posterior)
|
262 |
+
|
263 |
+
def _decode(self, z: torch.FloatTensor, return_dict: bool = True) -> Union[DecoderOutput, torch.FloatTensor]:
|
264 |
+
if self.use_tiling and (z.shape[-1] > self.tile_latent_min_size or z.shape[-2] > self.tile_latent_min_size):
|
265 |
+
return self.tiled_decode(z, return_dict=return_dict)
|
266 |
+
|
267 |
+
z = self.post_quant_conv(z)
|
268 |
+
dec = self.decoder(z)
|
269 |
+
|
270 |
+
if not return_dict:
|
271 |
+
return (dec,)
|
272 |
+
|
273 |
+
return DecoderOutput(sample=dec)
|
274 |
+
|
275 |
+
@apply_forward_hook
|
276 |
+
def decode(
|
277 |
+
self, z: torch.FloatTensor, return_dict: bool = True, generator=None
|
278 |
+
) -> Union[DecoderOutput, torch.FloatTensor]:
|
279 |
+
"""
|
280 |
+
Decode a batch of images.
|
281 |
+
Args:
|
282 |
+
z (`torch.FloatTensor`): Input batch of latent vectors.
|
283 |
+
return_dict (`bool`, *optional*, defaults to `True`):
|
284 |
+
Whether to return a [`~models.vae.DecoderOutput`] instead of a plain tuple.
|
285 |
+
Returns:
|
286 |
+
[`~models.vae.DecoderOutput`] or `tuple`:
|
287 |
+
If return_dict is True, a [`~models.vae.DecoderOutput`] is returned, otherwise a plain `tuple` is
|
288 |
+
returned.
|
289 |
+
"""
|
290 |
+
if self.use_slicing and z.shape[0] > 1:
|
291 |
+
decoded_slices = [self._decode(z_slice).sample for z_slice in z.split(1)]
|
292 |
+
decoded = torch.cat(decoded_slices)
|
293 |
+
else:
|
294 |
+
decoded = self._decode(z).sample
|
295 |
+
|
296 |
+
if not return_dict:
|
297 |
+
return (decoded,)
|
298 |
+
|
299 |
+
return DecoderOutput(sample=decoded)
|
300 |
+
|
301 |
+
def blend_v(self, a: torch.Tensor, b: torch.Tensor, blend_extent: int) -> torch.Tensor:
|
302 |
+
blend_extent = min(a.shape[2], b.shape[2], blend_extent)
|
303 |
+
for y in range(blend_extent):
|
304 |
+
b[:, :, y, :] = a[:, :, -blend_extent + y, :] * (1 - y / blend_extent) + b[:, :, y, :] * (y / blend_extent)
|
305 |
+
return b
|
306 |
+
|
307 |
+
def blend_h(self, a: torch.Tensor, b: torch.Tensor, blend_extent: int) -> torch.Tensor:
|
308 |
+
blend_extent = min(a.shape[3], b.shape[3], blend_extent)
|
309 |
+
for x in range(blend_extent):
|
310 |
+
b[:, :, :, x] = a[:, :, :, -blend_extent + x] * (1 - x / blend_extent) + b[:, :, :, x] * (x / blend_extent)
|
311 |
+
return b
|
312 |
+
|
313 |
+
def tiled_encode(self, x: torch.FloatTensor, return_dict: bool = True) -> AutoencoderKLOutput:
|
314 |
+
r"""Encode a batch of images using a tiled encoder.
|
315 |
+
When this option is enabled, the VAE will split the input tensor into tiles to compute encoding in several
|
316 |
+
steps. This is useful to keep memory use constant regardless of image size. The end result of tiled encoding is
|
317 |
+
different from non-tiled encoding because each tile uses a different encoder. To avoid tiling artifacts, the
|
318 |
+
tiles overlap and are blended together to form a smooth output. You may still see tile-sized changes in the
|
319 |
+
output, but they should be much less noticeable.
|
320 |
+
Args:
|
321 |
+
x (`torch.FloatTensor`): Input batch of images.
|
322 |
+
return_dict (`bool`, *optional*, defaults to `True`):
|
323 |
+
Whether or not to return a [`~models.autoencoder_kl.AutoencoderKLOutput`] instead of a plain tuple.
|
324 |
+
Returns:
|
325 |
+
[`~models.autoencoder_kl.AutoencoderKLOutput`] or `tuple`:
|
326 |
+
If return_dict is True, a [`~models.autoencoder_kl.AutoencoderKLOutput`] is returned, otherwise a plain
|
327 |
+
`tuple` is returned.
|
328 |
+
"""
|
329 |
+
overlap_size = int(self.tile_sample_min_size * (1 - self.tile_overlap_factor))
|
330 |
+
blend_extent = int(self.tile_latent_min_size * self.tile_overlap_factor)
|
331 |
+
row_limit = self.tile_latent_min_size - blend_extent
|
332 |
+
|
333 |
+
# Split the image into 512x512 tiles and encode them separately.
|
334 |
+
rows = []
|
335 |
+
for i in range(0, x.shape[2], overlap_size):
|
336 |
+
row = []
|
337 |
+
for j in range(0, x.shape[3], overlap_size):
|
338 |
+
tile = x[:, :, i : i + self.tile_sample_min_size, j : j + self.tile_sample_min_size]
|
339 |
+
tile = self.encoder(tile)
|
340 |
+
tile = self.quant_conv(tile)
|
341 |
+
row.append(tile)
|
342 |
+
rows.append(row)
|
343 |
+
result_rows = []
|
344 |
+
for i, row in enumerate(rows):
|
345 |
+
result_row = []
|
346 |
+
for j, tile in enumerate(row):
|
347 |
+
# blend the above tile and the left tile
|
348 |
+
# to the current tile and add the current tile to the result row
|
349 |
+
if i > 0:
|
350 |
+
tile = self.blend_v(rows[i - 1][j], tile, blend_extent)
|
351 |
+
if j > 0:
|
352 |
+
tile = self.blend_h(row[j - 1], tile, blend_extent)
|
353 |
+
result_row.append(tile[:, :, :row_limit, :row_limit])
|
354 |
+
result_rows.append(torch.cat(result_row, dim=3))
|
355 |
+
|
356 |
+
moments = torch.cat(result_rows, dim=2)
|
357 |
+
posterior = DiagonalGaussianDistribution(moments)
|
358 |
+
|
359 |
+
if not return_dict:
|
360 |
+
return (posterior,)
|
361 |
+
|
362 |
+
return AutoencoderKLOutput(latent_dist=posterior)
|
363 |
+
|
364 |
+
def tiled_decode(self, z: torch.FloatTensor, return_dict: bool = True) -> Union[DecoderOutput, torch.FloatTensor]:
|
365 |
+
r"""
|
366 |
+
Decode a batch of images using a tiled decoder.
|
367 |
+
Args:
|
368 |
+
z (`torch.FloatTensor`): Input batch of latent vectors.
|
369 |
+
return_dict (`bool`, *optional*, defaults to `True`):
|
370 |
+
Whether or not to return a [`~models.vae.DecoderOutput`] instead of a plain tuple.
|
371 |
+
Returns:
|
372 |
+
[`~models.vae.DecoderOutput`] or `tuple`:
|
373 |
+
If return_dict is True, a [`~models.vae.DecoderOutput`] is returned, otherwise a plain `tuple` is
|
374 |
+
returned.
|
375 |
+
"""
|
376 |
+
overlap_size = int(self.tile_latent_min_size * (1 - self.tile_overlap_factor))
|
377 |
+
blend_extent = int(self.tile_sample_min_size * self.tile_overlap_factor)
|
378 |
+
row_limit = self.tile_sample_min_size - blend_extent
|
379 |
+
|
380 |
+
# Split z into overlapping 64x64 tiles and decode them separately.
|
381 |
+
# The tiles have an overlap to avoid seams between tiles.
|
382 |
+
rows = []
|
383 |
+
for i in range(0, z.shape[2], overlap_size):
|
384 |
+
row = []
|
385 |
+
for j in range(0, z.shape[3], overlap_size):
|
386 |
+
tile = z[:, :, i : i + self.tile_latent_min_size, j : j + self.tile_latent_min_size]
|
387 |
+
tile = self.post_quant_conv(tile)
|
388 |
+
decoded = self.decoder(tile)
|
389 |
+
row.append(decoded)
|
390 |
+
rows.append(row)
|
391 |
+
result_rows = []
|
392 |
+
for i, row in enumerate(rows):
|
393 |
+
result_row = []
|
394 |
+
for j, tile in enumerate(row):
|
395 |
+
# blend the above tile and the left tile
|
396 |
+
# to the current tile and add the current tile to the result row
|
397 |
+
if i > 0:
|
398 |
+
tile = self.blend_v(rows[i - 1][j], tile, blend_extent)
|
399 |
+
if j > 0:
|
400 |
+
tile = self.blend_h(row[j - 1], tile, blend_extent)
|
401 |
+
result_row.append(tile[:, :, :row_limit, :row_limit])
|
402 |
+
result_rows.append(torch.cat(result_row, dim=3))
|
403 |
+
|
404 |
+
dec = torch.cat(result_rows, dim=2)
|
405 |
+
if not return_dict:
|
406 |
+
return (dec,)
|
407 |
+
|
408 |
+
return DecoderOutput(sample=dec)
|
409 |
+
|
410 |
+
def forward(
|
411 |
+
self,
|
412 |
+
sample: torch.FloatTensor,
|
413 |
+
sample_posterior: bool = False,
|
414 |
+
return_dict: bool = True,
|
415 |
+
generator: Optional[torch.Generator] = None,
|
416 |
+
) -> Union[DecoderOutput, torch.FloatTensor]:
|
417 |
+
r"""
|
418 |
+
Args:
|
419 |
+
sample (`torch.FloatTensor`): Input sample.
|
420 |
+
sample_posterior (`bool`, *optional*, defaults to `False`):
|
421 |
+
Whether to sample from the posterior.
|
422 |
+
return_dict (`bool`, *optional*, defaults to `True`):
|
423 |
+
Whether or not to return a [`DecoderOutput`] instead of a plain tuple.
|
424 |
+
"""
|
425 |
+
x = sample
|
426 |
+
posterior = self.encode(x).latent_dist
|
427 |
+
if sample_posterior:
|
428 |
+
z = posterior.sample(generator=generator)
|
429 |
+
else:
|
430 |
+
z = posterior.mode()
|
431 |
+
dec = self.decode(z).sample
|
432 |
+
|
433 |
+
if not return_dict:
|
434 |
+
return (dec,)
|
435 |
+
|
436 |
+
return DecoderOutput(sample=dec)
|
437 |
+
|
438 |
+
# Copied from diffusers.models.unet_2d_condition.UNet2DConditionModel.fuse_qkv_projections
|
439 |
+
def fuse_qkv_projections(self):
|
440 |
+
"""
|
441 |
+
Enables fused QKV projections. For self-attention modules, all projection matrices (i.e., query,
|
442 |
+
key, value) are fused. For cross-attention modules, key and value projection matrices are fused.
|
443 |
+
<Tip warning={true}>
|
444 |
+
This API is 🧪 experimental.
|
445 |
+
</Tip>
|
446 |
+
"""
|
447 |
+
self.original_attn_processors = None
|
448 |
+
|
449 |
+
for _, attn_processor in self.attn_processors.items():
|
450 |
+
if "Added" in str(attn_processor.__class__.__name__):
|
451 |
+
raise ValueError("`fuse_qkv_projections()` is not supported for models having added KV projections.")
|
452 |
+
|
453 |
+
self.original_attn_processors = self.attn_processors
|
454 |
+
|
455 |
+
for module in self.modules():
|
456 |
+
if isinstance(module, Attention):
|
457 |
+
module.fuse_projections(fuse=True)
|
458 |
+
|
459 |
+
# Copied from diffusers.models.unet_2d_condition.UNet2DConditionModel.unfuse_qkv_projections
|
460 |
+
def unfuse_qkv_projections(self):
|
461 |
+
"""Disables the fused QKV projection if enabled.
|
462 |
+
<Tip warning={true}>
|
463 |
+
This API is 🧪 experimental.
|
464 |
+
</Tip>
|
465 |
+
"""
|
466 |
+
if self.original_attn_processors is not None:
|
467 |
+
self.set_attn_processor(self.original_attn_processors)
|