{{ message }}
Repository navigation
Expand file tree
/
Copy pathadapter.py
More file actions
513 lines (414 loc) · 21.1 KB
/
Copy pathadapter.py
File metadata and controls
513 lines (414 loc) · 21.1 KB
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
# Copyright 2022 The HuggingFace Team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import os
from typing import Callable, List, Optional, Union
import torch
import torch.nn as nn
from ..configuration_utils import ConfigMixin, register_to_config
from ..utils import logging
from .modeling_utils import ModelMixin
from .resnet import Downsample2D
logger = logging.get_logger(__name__)
class MultiAdapter(ModelMixin):
r"""
MultiAdapter is a wrapper model that contains multiple adapter models and merges their outputs according to
user-assigned weighting.
This model inherits from [`ModelMixin`]. Check the superclass documentation for the generic methods the library
implements for all the model (such as downloading or saving, etc.)
Parameters:
adapters (`List[T2IAdapter]`, *optional*, defaults to None):
A list of `T2IAdapter` model instances.
"""
def __init__(self, adapters: List["T2IAdapter"]):
super(MultiAdapter, self).__init__()
self.num_adapter = len(adapters)
self.adapters = nn.ModuleList(adapters)
if len(adapters) == 0:
raise ValueError("Expecting at least one adapter")
if len(adapters) == 1:
raise ValueError("For a single adapter, please use the `T2IAdapter` class instead of `MultiAdapter`")
# The outputs from each adapter are added together with a weight
# This means that the change in dimenstions from downsampling must
# be the same for all adapters. Inductively, it also means the total
# downscale factor must also be the same for all adapters.
first_adapter_total_downscale_factor = adapters[0].total_downscale_factor
for idx in range(1, len(adapters)):
adapter_idx_total_downscale_factor = adapters[idx].total_downscale_factor
if adapter_idx_total_downscale_factor != first_adapter_total_downscale_factor:
raise ValueError(
f"Expecting all adapters to have the same total_downscale_factor, "
f"but got adapters[0].total_downscale_factor={first_adapter_total_downscale_factor} and "
f"adapter[`{idx}`]={adapter_idx_total_downscale_factor}"
)
self.total_downscale_factor = adapters[0].total_downscale_factor
def forward(self, xs: torch.Tensor, adapter_weights: Optional[List[float]] = None) -> List[torch.Tensor]:
r"""
Args:
xs (`torch.Tensor`):
(batch, channel, height, width) input images for multiple adapter models concated along dimension 1,
`channel` should equal to `num_adapter` * "number of channel of image".
adapter_weights (`List[float]`, *optional*, defaults to None):
List of floats representing the weight which will be multiply to each adapter's output before adding
them together.
"""
if adapter_weights is None:
adapter_weights = torch.tensor([1 / self.num_adapter] * self.num_adapter)
else:
adapter_weights = torch.tensor(adapter_weights)
accume_state = None
for x, w, adapter in zip(xs, adapter_weights, self.adapters):
features = adapter(x)
if accume_state is None:
accume_state = features
for i in range(len(accume_state)):
accume_state[i] = w * accume_state[i]
else:
for i in range(len(features)):
accume_state[i] += w * features[i]
return accume_state
def save_pretrained(
self,
save_directory: Union[str, os.PathLike],
is_main_process: bool = True,
save_function: Callable = None,
safe_serialization: bool = True,
variant: Optional[str] = None,
):
"""
Save a model and its configuration file to a directory, so that it can be re-loaded using the
`[`~models.adapter.MultiAdapter.from_pretrained`]` class method.
Arguments:
save_directory (`str` or `os.PathLike`):
Directory to which to save. Will be created if it doesn't exist.
is_main_process (`bool`, *optional*, defaults to `True`):
Whether the process calling this is the main process or not. Useful when in distributed training like
TPUs and need to call this function on all processes. In this case, set `is_main_process=True` only on
the main process to avoid race conditions.
save_function (`Callable`):
The function to use to save the state dictionary. Useful on distributed training like TPUs when one
need to replace `torch.save` by another method. Can be configured with the environment variable
`DIFFUSERS_SAVE_MODE`.
safe_serialization (`bool`, *optional*, defaults to `True`):
Whether to save the model using `safetensors` or the traditional PyTorch way (that uses `pickle`).
variant (`str`, *optional*):
If specified, weights are saved in the format pytorch_model.<variant>.bin.
"""
idx = 0
model_path_to_save = save_directory
for adapter in self.adapters:
adapter.save_pretrained(
model_path_to_save,
is_main_process=is_main_process,
save_function=save_function,
safe_serialization=safe_serialization,
variant=variant,
)
idx += 1
model_path_to_save = model_path_to_save + f"_{idx}"
@classmethod
def from_pretrained(cls, pretrained_model_path: Optional[Union[str, os.PathLike]], **kwargs):
r"""
Instantiate a pretrained MultiAdapter model from multiple pre-trained adapter models.
The model is set in evaluation mode by default using `model.eval()` (Dropout modules are deactivated). To train
the model, you should first set it back in training mode with `model.train()`.
The warning *Weights from XXX not initialized from pretrained model* means that the weights of XXX do not come
pretrained with the rest of the model. It is up to you to train those weights with a downstream fine-tuning
task.
The warning *Weights from XXX not used in YYY* means that the layer XXX is not used by YYY, therefore those
weights are discarded.
Parameters:
pretrained_model_path (`os.PathLike`):
A path to a *directory* containing model weights saved using
[`~diffusers.models.adapter.MultiAdapter.save_pretrained`], e.g., `./my_model_directory/adapter`.
torch_dtype (`str` or `torch.dtype`, *optional*):
Override the default `torch.dtype` and load the model under this dtype. If `"auto"` is passed the dtype
will be automatically derived from the model's weights.
output_loading_info(`bool`, *optional*, defaults to `False`):
Whether or not to also return a dictionary containing missing keys, unexpected keys and error messages.
device_map (`str` or `Dict[str, Union[int, str, torch.device]]`, *optional*):
A map that specifies where each submodule should go. It doesn't need to be refined to each
parameter/buffer name, once a given module name is inside, every submodule of it will be sent to the
same device.
To have Accelerate compute the most optimized `device_map` automatically, set `device_map="auto"`. For
more information about each option see [designing a device
map](https://hf.co/docs/accelerate/main/en/usage_guides/big_modeling#designing-a-device-map).
max_memory (`Dict`, *optional*):
A dictionary device identifier to maximum memory. Will default to the maximum memory available for each
GPU and the available CPU RAM if unset.
low_cpu_mem_usage (`bool`, *optional*, defaults to `True` if torch version >= 1.9.0 else `False`):
Speed up model loading by not initializing the weights and only loading the pre-trained weights. This
also tries to not use more than 1x model size in CPU memory (including peak memory) while loading the
model. This is only supported when torch version >= 1.9.0. If you are using an older version of torch,
setting this argument to `True` will raise an error.
variant (`str`, *optional*):
If specified load weights from `variant` filename, *e.g.* pytorch_model.<variant>.bin. `variant` is
ignored when using `from_flax`.
use_safetensors (`bool`, *optional*, defaults to `None`):
If set to `None`, the `safetensors` weights will be downloaded if they're available **and** if the
`safetensors` library is installed. If set to `True`, the model will be forcibly loaded from
`safetensors` weights. If set to `False`, loading will *not* use `safetensors`.
"""
idx = 0
adapters = []
# load adapter and append to list until no adapter directory exists anymore
# first adapter has to be saved under `./mydirectory/adapter` to be compliant with `DiffusionPipeline.from_pretrained`
# second, third, ... adapters have to be saved under `./mydirectory/adapter_1`, `./mydirectory/adapter_2`, ...
model_path_to_load = pretrained_model_path
while os.path.isdir(model_path_to_load):
adapter = T2IAdapter.from_pretrained(model_path_to_load, **kwargs)
adapters.append(adapter)
idx += 1
model_path_to_load = pretrained_model_path + f"_{idx}"
logger.info(f"{len(adapters)} adapters loaded from {pretrained_model_path}.")
if len(adapters) == 0:
raise ValueError(
f"No T2IAdapters found under {os.path.dirname(pretrained_model_path)}. Expected at least {pretrained_model_path + '_0'}."
)
return cls(adapters)
class T2IAdapter(ModelMixin, ConfigMixin):
r"""
A simple ResNet-like model that accepts images containing control signals such as keyposes and depth. The model
generates multiple feature maps that are used as additional conditioning in [`UNet2DConditionModel`]. The model's
architecture follows the original implementation of
[Adapter](https://github.com/TencentARC/T2I-Adapter/blob/686de4681515662c0ac2ffa07bf5dda83af1038a/ldm/modules/encoders/adapter.py#L97)
and
[AdapterLight](https://github.com/TencentARC/T2I-Adapter/blob/686de4681515662c0ac2ffa07bf5dda83af1038a/ldm/modules/encoders/adapter.py#L235).
This model inherits from [`ModelMixin`]. Check the superclass documentation for the generic methods the library
implements for all the model (such as downloading or saving, etc.)
Parameters:
in_channels (`int`, *optional*, defaults to 3):
Number of channels of Aapter's input(*control image*). Set this parameter to 1 if you're using gray scale
image as *control image*.
channels (`List[int]`, *optional*, defaults to `(320, 640, 1280, 1280)`):
The number of channel of each downsample block's output hidden state. The `len(block_out_channels)` will
also determine the number of downsample blocks in the Adapter.
num_res_blocks (`int`, *optional*, defaults to 2):
Number of ResNet blocks in each downsample block
"""
@register_to_config
def __init__(
self,
in_channels: int = 3,
channels: List[int] = [320, 640, 1280, 1280],
num_res_blocks: int = 2,
downscale_factor: int = 8,
adapter_type: str = "full_adapter",
):
super().__init__()
if adapter_type == "full_adapter":
self.adapter = FullAdapter(in_channels, channels, num_res_blocks, downscale_factor)
elif adapter_type == "full_adapter_xl":
self.adapter = FullAdapterXL(in_channels, channels, num_res_blocks, downscale_factor)
elif adapter_type == "light_adapter":
self.adapter = LightAdapter(in_channels, channels, num_res_blocks, downscale_factor)
else:
raise ValueError(
f"Unsupported adapter_type: '{adapter_type}'. Choose either 'full_adapter' or "
"'full_adapter_xl' or 'light_adapter'."
)
def forward(self, x: torch.Tensor) -> List[torch.Tensor]:
r"""
This function processes the input tensor `x` through the adapter model and returns a list of feature tensors,
each representing information extracted at a different scale from the input. The length of the list is
determined by the number of downsample blocks in the Adapter, as specified by the `channels` and
`num_res_blocks` parameters during initialization.
"""
return self.adapter(x)
@property
def total_downscale_factor(self):
return self.adapter.total_downscale_factor
# full adapter
class FullAdapter(nn.Module):
def __init__(
self,
in_channels: int = 3,
channels: List[int] = [320, 640, 1280, 1280],
num_res_blocks: int = 2,
downscale_factor: int = 8,
):
super().__init__()
in_channels = in_channels * downscale_factor**2
self.unshuffle = nn.PixelUnshuffle(downscale_factor)
self.conv_in = nn.Conv2d(in_channels, channels[0], kernel_size=3, padding=1)
self.body = nn.ModuleList(
[
AdapterBlock(channels[0], channels[0], num_res_blocks),
*[
AdapterBlock(channels[i - 1], channels[i], num_res_blocks, down=True)
for i in range(1, len(channels))
],
]
)
self.total_downscale_factor = downscale_factor * 2 ** (len(channels) - 1)
def forward(self, x: torch.Tensor) -> List[torch.Tensor]:
r"""
This method processes the input tensor `x` through the FullAdapter model and performs operations including
pixel unshuffling, convolution, and a stack of AdapterBlocks. It returns a list of feature tensors, each
capturing information at a different stage of processing within the FullAdapter model. The number of feature
tensors in the list is determined by the number of downsample blocks specified during initialization.
"""
x = self.unshuffle(x)
x = self.conv_in(x)
features = []
for block in self.body:
x = block(x)
features.append(x)
return features
class FullAdapterXL(nn.Module):
def __init__(
self,
in_channels: int = 3,
channels: List[int] = [320, 640, 1280, 1280],
num_res_blocks: int = 2,
downscale_factor: int = 16,
):
super().__init__()
in_channels = in_channels * downscale_factor**2
self.unshuffle = nn.PixelUnshuffle(downscale_factor)
self.conv_in = nn.Conv2d(in_channels, channels[0], kernel_size=3, padding=1)
self.body = []
# blocks to extract XL features with dimensions of [320, 64, 64], [640, 64, 64], [1280, 32, 32], [1280, 32, 32]
for i in range(len(channels)):
if i == 1:
self.body.append(AdapterBlock(channels[i - 1], channels[i], num_res_blocks))
elif i == 2:
self.body.append(AdapterBlock(channels[i - 1], channels[i], num_res_blocks, down=True))
else:
self.body.append(AdapterBlock(channels[i], channels[i], num_res_blocks))
self.body = nn.ModuleList(self.body)
# XL has only one downsampling AdapterBlock.
self.total_downscale_factor = downscale_factor * 2
def forward(self, x: torch.Tensor) -> List[torch.Tensor]:
r"""
This method takes the tensor x as input and processes it through FullAdapterXL model. It consists of operations
including unshuffling pixels, applying convolution layer and appending each block into list of feature tensors.
"""
x = self.unshuffle(x)
x = self.conv_in(x)
features = []
for block in self.body:
x = block(x)
features.append(x)
return features
class AdapterBlock(nn.Module):
def __init__(self, in_channels, out_channels, num_res_blocks, down=False):
super().__init__()
self.downsample = None
if down:
self.downsample = Downsample2D(in_channels)
self.in_conv = None
if in_channels != out_channels:
self.in_conv = nn.Conv2d(in_channels, out_channels, kernel_size=1)
self.resnets = nn.Sequential(
*[AdapterResnetBlock(out_channels) for _ in range(num_res_blocks)],
)
def forward(self, x):
r"""
This method takes tensor x as input and performs operations downsampling and convolutional layers if the
self.downsample and self.in_conv properties of AdapterBlock model are specified. Then it applies a series of
residual blocks to the input tensor.
"""
if self.downsample is not None:
x = self.downsample(x)
if self.in_conv is not None:
x = self.in_conv(x)
x = self.resnets(x)
return x
class AdapterResnetBlock(nn.Module):
def __init__(self, channels):
super().__init__()
self.block1 = nn.Conv2d(channels, channels, kernel_size=3, padding=1)
self.act = nn.ReLU()
self.block2 = nn.Conv2d(channels, channels, kernel_size=1)
def forward(self, x):
r"""
This method takes input tensor x and applies a convolutional layer, ReLU activation, and another convolutional
layer on the input tensor. It returns addition with the input tensor.
"""
h = x
h = self.block1(h)
h = self.act(h)
h = self.block2(h)
return h + x
# light adapter
class LightAdapter(nn.Module):
def __init__(
self,
in_channels: int = 3,
channels: List[int] = [320, 640, 1280],
num_res_blocks: int = 4,
downscale_factor: int = 8,
):
super().__init__()
in_channels = in_channels * downscale_factor**2
self.unshuffle = nn.PixelUnshuffle(downscale_factor)
self.body = nn.ModuleList(
[
LightAdapterBlock(in_channels, channels[0], num_res_blocks),
*[
LightAdapterBlock(channels[i], channels[i + 1], num_res_blocks, down=True)
for i in range(len(channels) - 1)
],
LightAdapterBlock(channels[-1], channels[-1], num_res_blocks, down=True),
]
)
self.total_downscale_factor = downscale_factor * (2 ** len(channels))
def forward(self, x):
r"""
This method takes the input tensor x and performs downscaling and appends it in list of feature tensors. Each
feature tensor corresponds to a different level of processing within the LightAdapter.
"""
x = self.unshuffle(x)
features = []
for block in self.body:
x = block(x)
features.append(x)
return features
class LightAdapterBlock(nn.Module):
def __init__(self, in_channels, out_channels, num_res_blocks, down=False):
super().__init__()
mid_channels = out_channels // 4
self.downsample = None
if down:
self.downsample = Downsample2D(in_channels)
self.in_conv = nn.Conv2d(in_channels, mid_channels, kernel_size=1)
self.resnets = nn.Sequential(*[LightAdapterResnetBlock(mid_channels) for _ in range(num_res_blocks)])
self.out_conv = nn.Conv2d(mid_channels, out_channels, kernel_size=1)
def forward(self, x):
r"""
This method takes tensor x as input and performs downsampling if required. Then it applies in convolution
layer, a sequence of residual blocks, and out convolutional layer.
"""
if self.downsample is not None:
x = self.downsample(x)
x = self.in_conv(x)
x = self.resnets(x)
x = self.out_conv(x)
return x
class LightAdapterResnetBlock(nn.Module):
def __init__(self, channels):
super().__init__()
self.block1 = nn.Conv2d(channels, channels, kernel_size=3, padding=1)
self.act = nn.ReLU()
self.block2 = nn.Conv2d(channels, channels, kernel_size=3, padding=1)
def forward(self, x):
r"""
This function takes input tensor x and processes it through one convolutional layer, ReLU activation, and
another convolutional layer and adds it to input tensor.
"""
h = x
h = self.block1(h)
h = self.act(h)
h = self.block2(h)
return h + x
You can’t perform that action at this time.