{{ message }}
Repository navigation
Expand file tree
/
Copy pathloaders.py
More file actions
1053 lines (888 loc) · 49.1 KB
/
Copy pathloaders.py
File metadata and controls
1053 lines (888 loc) · 49.1 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
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
# Copyright 2023 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 collections import defaultdict
from typing import Callable, Dict, List, Optional, Union
import torch
from .models.attention_processor import LoRAAttnProcessor
from .utils import (
DIFFUSERS_CACHE,
HF_HUB_OFFLINE,
TEXT_ENCODER_TARGET_MODULES,
_get_model_file,
deprecate,
is_safetensors_available,
is_transformers_available,
logging,
)
if is_safetensors_available():
import safetensors
if is_transformers_available():
from transformers import PreTrainedModel, PreTrainedTokenizer
logger = logging.get_logger(__name__)
LORA_WEIGHT_NAME = "pytorch_lora_weights.bin"
LORA_WEIGHT_NAME_SAFE = "pytorch_lora_weights.safetensors"
TEXT_INVERSION_NAME = "learned_embeds.bin"
TEXT_INVERSION_NAME_SAFE = "learned_embeds.safetensors"
class AttnProcsLayers(torch.nn.Module):
def __init__(self, state_dict: Dict[str, torch.Tensor]):
super().__init__()
self.layers = torch.nn.ModuleList(state_dict.values())
self.mapping = dict(enumerate(state_dict.keys()))
self.rev_mapping = {v: k for k, v in enumerate(state_dict.keys())}
# we add a hook to state_dict() and load_state_dict() so that the
# naming fits with `unet.attn_processors`
def map_to(module, state_dict, *args, **kwargs):
new_state_dict = {}
for key, value in state_dict.items():
num = int(key.split(".")[1]) # 0 is always "layers"
new_key = key.replace(f"layers.{num}", module.mapping[num])
new_state_dict[new_key] = value
return new_state_dict
def map_from(module, state_dict, *args, **kwargs):
all_keys = list(state_dict.keys())
for key in all_keys:
replace_key = key.split(".processor")[0] + ".processor"
new_key = key.replace(replace_key, f"layers.{module.rev_mapping[replace_key]}")
state_dict[new_key] = state_dict[key]
del state_dict[key]
self._register_state_dict_hook(map_to)
self._register_load_state_dict_pre_hook(map_from, with_module=True)
class UNet2DConditionLoadersMixin:
def load_attn_procs(self, pretrained_model_name_or_path_or_dict: Union[str, Dict[str, torch.Tensor]], **kwargs):
r"""
Load pretrained attention processor layers into `UNet2DConditionModel`. Attention processor layers have to be
defined in
[`cross_attention.py`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/cross_attention.py)
and be a `torch.nn.Module` class.
<Tip warning={true}>
This function is experimental and might change in the future.
</Tip>
Parameters:
pretrained_model_name_or_path_or_dict (`str` or `os.PathLike` or `dict`):
Can be either:
- A string, the *model id* of a pretrained model hosted inside a model repo on huggingface.co.
Valid model ids should have an organization name, like `google/ddpm-celebahq-256`.
- A path to a *directory* containing model weights saved using [`~ModelMixin.save_config`], e.g.,
`./my_model_directory/`.
- A [torch state
dict](https://pytorch.org/tutorials/beginner/saving_loading_models.html#what-is-a-state-dict).
cache_dir (`Union[str, os.PathLike]`, *optional*):
Path to a directory in which a downloaded pretrained model configuration should be cached if the
standard cache should not be used.
force_download (`bool`, *optional*, defaults to `False`):
Whether or not to force the (re-)download of the model weights and configuration files, overriding the
cached versions if they exist.
resume_download (`bool`, *optional*, defaults to `False`):
Whether or not to delete incompletely received files. Will attempt to resume the download if such a
file exists.
proxies (`Dict[str, str]`, *optional*):
A dictionary of proxy servers to use by protocol or endpoint, e.g., `{'http': 'foo.bar:3128',
'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request.
local_files_only(`bool`, *optional*, defaults to `False`):
Whether or not to only look at local files (i.e., do not try to download the model).
use_auth_token (`str` or *bool*, *optional*):
The token to use as HTTP bearer authorization for remote files. If `True`, will use the token generated
when running `diffusers-cli login` (stored in `~/.huggingface`).
revision (`str`, *optional*, defaults to `"main"`):
The specific model version to use. It can be a branch name, a tag name, or a commit id, since we use a
git-based system for storing models and other artifacts on huggingface.co, so `revision` can be any
identifier allowed by git.
subfolder (`str`, *optional*, defaults to `""`):
In case the relevant files are located inside a subfolder of the model repo (either remote in
huggingface.co or downloaded locally), you can specify the folder name here.
mirror (`str`, *optional*):
Mirror source to accelerate downloads in China. If you are from China and have an accessibility
problem, you can set this option to resolve it. Note that we do not guarantee the timeliness or safety.
Please refer to the mirror site for more information.
<Tip>
It is required to be logged in (`huggingface-cli login`) when you want to use private or [gated
models](https://huggingface.co/docs/hub/models-gated#gated-models).
</Tip>
"""
cache_dir = kwargs.pop("cache_dir", DIFFUSERS_CACHE)
force_download = kwargs.pop("force_download", False)
resume_download = kwargs.pop("resume_download", False)
proxies = kwargs.pop("proxies", None)
local_files_only = kwargs.pop("local_files_only", HF_HUB_OFFLINE)
use_auth_token = kwargs.pop("use_auth_token", None)
revision = kwargs.pop("revision", None)
subfolder = kwargs.pop("subfolder", None)
weight_name = kwargs.pop("weight_name", None)
use_safetensors = kwargs.pop("use_safetensors", None)
if use_safetensors and not is_safetensors_available():
raise ValueError(
"`use_safetensors`=True but safetensors is not installed. Please install safetensors with `pip install safetenstors"
)
allow_pickle = False
if use_safetensors is None:
use_safetensors = is_safetensors_available()
allow_pickle = True
user_agent = {
"file_type": "attn_procs_weights",
"framework": "pytorch",
}
model_file = None
if not isinstance(pretrained_model_name_or_path_or_dict, dict):
# Let's first try to load .safetensors weights
if (use_safetensors and weight_name is None) or (
weight_name is not None and weight_name.endswith(".safetensors")
):
try:
model_file = _get_model_file(
pretrained_model_name_or_path_or_dict,
weights_name=weight_name or LORA_WEIGHT_NAME_SAFE,
cache_dir=cache_dir,
force_download=force_download,
resume_download=resume_download,
proxies=proxies,
local_files_only=local_files_only,
use_auth_token=use_auth_token,
revision=revision,
subfolder=subfolder,
user_agent=user_agent,
)
state_dict = safetensors.torch.load_file(model_file, device="cpu")
except IOError as e:
if not allow_pickle:
raise e
# try loading non-safetensors weights
pass
if model_file is None:
model_file = _get_model_file(
pretrained_model_name_or_path_or_dict,
weights_name=weight_name or LORA_WEIGHT_NAME,
cache_dir=cache_dir,
force_download=force_download,
resume_download=resume_download,
proxies=proxies,
local_files_only=local_files_only,
use_auth_token=use_auth_token,
revision=revision,
subfolder=subfolder,
user_agent=user_agent,
)
state_dict = torch.load(model_file, map_location="cpu")
else:
state_dict = pretrained_model_name_or_path_or_dict
# fill attn processors
attn_processors = {}
is_lora = all("lora" in k for k in state_dict.keys())
if is_lora:
lora_grouped_dict = defaultdict(dict)
for key, value in state_dict.items():
attn_processor_key, sub_key = ".".join(key.split(".")[:-3]), ".".join(key.split(".")[-3:])
lora_grouped_dict[attn_processor_key][sub_key] = value
for key, value_dict in lora_grouped_dict.items():
rank = value_dict["to_k_lora.down.weight"].shape[0]
cross_attention_dim = value_dict["to_k_lora.down.weight"].shape[1]
hidden_size = value_dict["to_k_lora.up.weight"].shape[0]
attn_processors[key] = LoRAAttnProcessor(
hidden_size=hidden_size, cross_attention_dim=cross_attention_dim, rank=rank
)
attn_processors[key].load_state_dict(value_dict)
else:
raise ValueError(f"{model_file} does not seem to be in the correct format expected by LoRA training.")
# set correct dtype & device
attn_processors = {k: v.to(device=self.device, dtype=self.dtype) for k, v in attn_processors.items()}
# set layers
self.set_attn_processor(attn_processors)
def save_attn_procs(
self,
save_directory: Union[str, os.PathLike],
is_main_process: bool = True,
weight_name: str = None,
save_function: Callable = None,
safe_serialization: bool = False,
**kwargs,
):
r"""
Save an attention processor to a directory, so that it can be re-loaded using the
[`~loaders.UNet2DConditionLoadersMixin.load_attn_procs`] 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`.
"""
weight_name = weight_name or deprecate(
"weights_name",
"0.18.0",
"`weights_name` is deprecated, please use `weight_name` instead.",
take_from=kwargs,
)
if os.path.isfile(save_directory):
logger.error(f"Provided path ({save_directory}) should be a directory, not a file")
return
if save_function is None:
if safe_serialization:
def save_function(weights, filename):
return safetensors.torch.save_file(weights, filename, metadata={"format": "pt"})
else:
save_function = torch.save
os.makedirs(save_directory, exist_ok=True)
model_to_save = AttnProcsLayers(self.attn_processors)
# Save the model
state_dict = model_to_save.state_dict()
if weight_name is None:
if safe_serialization:
weight_name = LORA_WEIGHT_NAME_SAFE
else:
weight_name = LORA_WEIGHT_NAME
# Save the model
save_function(state_dict, os.path.join(save_directory, weight_name))
logger.info(f"Model weights saved in {os.path.join(save_directory, weight_name)}")
class TextualInversionLoaderMixin:
r"""
Mixin class for loading textual inversion tokens and embeddings to the tokenizer and text encoder.
"""
def maybe_convert_prompt(self, prompt: Union[str, List[str]], tokenizer: "PreTrainedTokenizer"):
r"""
Maybe convert a prompt into a "multi vector"-compatible prompt. If the prompt includes a token that corresponds
to a multi-vector textual inversion embedding, this function will process the prompt so that the special token
is replaced with multiple special tokens each corresponding to one of the vectors. If the prompt has no textual
inversion token or a textual inversion token that is a single vector, the input prompt is simply returned.
Parameters:
prompt (`str` or list of `str`):
The prompt or prompts to guide the image generation.
tokenizer (`PreTrainedTokenizer`):
The tokenizer responsible for encoding the prompt into input tokens.
Returns:
`str` or list of `str`: The converted prompt
"""
if not isinstance(prompt, List):
prompts = [prompt]
else:
prompts = prompt
prompts = [self._maybe_convert_prompt(p, tokenizer) for p in prompts]
if not isinstance(prompt, List):
return prompts[0]
return prompts
def _maybe_convert_prompt(self, prompt: str, tokenizer: "PreTrainedTokenizer"):
r"""
Maybe convert a prompt into a "multi vector"-compatible prompt. If the prompt includes a token that corresponds
to a multi-vector textual inversion embedding, this function will process the prompt so that the special token
is replaced with multiple special tokens each corresponding to one of the vectors. If the prompt has no textual
inversion token or a textual inversion token that is a single vector, the input prompt is simply returned.
Parameters:
prompt (`str`):
The prompt to guide the image generation.
tokenizer (`PreTrainedTokenizer`):
The tokenizer responsible for encoding the prompt into input tokens.
Returns:
`str`: The converted prompt
"""
tokens = tokenizer.tokenize(prompt)
for token in tokens:
if token in tokenizer.added_tokens_encoder:
replacement = token
i = 1
while f"{token}_{i}" in tokenizer.added_tokens_encoder:
replacement += f"{token}_{i}"
i += 1
prompt = prompt.replace(token, replacement)
return prompt
def load_textual_inversion(
self, pretrained_model_name_or_path: Union[str, Dict[str, torch.Tensor]], token: Optional[str] = None, **kwargs
):
r"""
Load textual inversion embeddings into the text encoder of stable diffusion pipelines. Both `diffusers` and
`Automatic1111` formats are supported (see example below).
<Tip warning={true}>
This function is experimental and might change in the future.
</Tip>
Parameters:
pretrained_model_name_or_path (`str` or `os.PathLike`):
Can be either:
- A string, the *model id* of a pretrained model hosted inside a model repo on huggingface.co.
Valid model ids should have an organization name, like
`"sd-concepts-library/low-poly-hd-logos-icons"`.
- A path to a *directory* containing textual inversion weights, e.g.
`./my_text_inversion_directory/`.
weight_name (`str`, *optional*):
Name of a custom weight file. This should be used in two cases:
- The saved textual inversion file is in `diffusers` format, but was saved under a specific weight
name, such as `text_inv.bin`.
- The saved textual inversion file is in the "Automatic1111" form.
cache_dir (`Union[str, os.PathLike]`, *optional*):
Path to a directory in which a downloaded pretrained model configuration should be cached if the
standard cache should not be used.
force_download (`bool`, *optional*, defaults to `False`):
Whether or not to force the (re-)download of the model weights and configuration files, overriding the
cached versions if they exist.
resume_download (`bool`, *optional*, defaults to `False`):
Whether or not to delete incompletely received files. Will attempt to resume the download if such a
file exists.
proxies (`Dict[str, str]`, *optional*):
A dictionary of proxy servers to use by protocol or endpoint, e.g., `{'http': 'foo.bar:3128',
'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request.
local_files_only(`bool`, *optional*, defaults to `False`):
Whether or not to only look at local files (i.e., do not try to download the model).
use_auth_token (`str` or *bool*, *optional*):
The token to use as HTTP bearer authorization for remote files. If `True`, will use the token generated
when running `diffusers-cli login` (stored in `~/.huggingface`).
revision (`str`, *optional*, defaults to `"main"`):
The specific model version to use. It can be a branch name, a tag name, or a commit id, since we use a
git-based system for storing models and other artifacts on huggingface.co, so `revision` can be any
identifier allowed by git.
subfolder (`str`, *optional*, defaults to `""`):
In case the relevant files are located inside a subfolder of the model repo (either remote in
huggingface.co or downloaded locally), you can specify the folder name here.
mirror (`str`, *optional*):
Mirror source to accelerate downloads in China. If you are from China and have an accessibility
problem, you can set this option to resolve it. Note that we do not guarantee the timeliness or safety.
Please refer to the mirror site for more information.
<Tip>
It is required to be logged in (`huggingface-cli login`) when you want to use private or [gated
models](https://huggingface.co/docs/hub/models-gated#gated-models).
</Tip>
Example:
To load a textual inversion embedding vector in `diffusers` format:
```py
from diffusers import StableDiffusionPipeline
import torch
model_id = "runwayml/stable-diffusion-v1-5"
pipe = StableDiffusionPipeline.from_pretrained(model_id, torch_dtype=torch.float16).to("cuda")
pipe.load_textual_inversion("sd-concepts-library/cat-toy")
prompt = "A <cat-toy> backpack"
image = pipe(prompt, num_inference_steps=50).images[0]
image.save("cat-backpack.png")
```
To load a textual inversion embedding vector in Automatic1111 format, make sure to first download the vector,
e.g. from [civitAI](https://civitai.com/models/3036?modelVersionId=9857) and then load the vector locally:
```py
from diffusers import StableDiffusionPipeline
import torch
model_id = "runwayml/stable-diffusion-v1-5"
pipe = StableDiffusionPipeline.from_pretrained(model_id, torch_dtype=torch.float16).to("cuda")
pipe.load_textual_inversion("./charturnerv2.pt")
prompt = "charturnerv2, multiple views of the same character in the same outfit, a character turnaround of a woman wearing a black jacket and red shirt, best quality, intricate details."
image = pipe(prompt, num_inference_steps=50).images[0]
image.save("character.png")
```
"""
if not hasattr(self, "tokenizer") or not isinstance(self.tokenizer, PreTrainedTokenizer):
raise ValueError(
f"{self.__class__.__name__} requires `self.tokenizer` of type `PreTrainedTokenizer` for calling"
f" `{self.load_textual_inversion.__name__}`"
)
if not hasattr(self, "text_encoder") or not isinstance(self.text_encoder, PreTrainedModel):
raise ValueError(
f"{self.__class__.__name__} requires `self.text_encoder` of type `PreTrainedModel` for calling"
f" `{self.load_textual_inversion.__name__}`"
)
cache_dir = kwargs.pop("cache_dir", DIFFUSERS_CACHE)
force_download = kwargs.pop("force_download", False)
resume_download = kwargs.pop("resume_download", False)
proxies = kwargs.pop("proxies", None)
local_files_only = kwargs.pop("local_files_only", HF_HUB_OFFLINE)
use_auth_token = kwargs.pop("use_auth_token", None)
revision = kwargs.pop("revision", None)
subfolder = kwargs.pop("subfolder", None)
weight_name = kwargs.pop("weight_name", None)
use_safetensors = kwargs.pop("use_safetensors", None)
if use_safetensors and not is_safetensors_available():
raise ValueError(
"`use_safetensors`=True but safetensors is not installed. Please install safetensors with `pip install safetenstors"
)
allow_pickle = False
if use_safetensors is None:
use_safetensors = is_safetensors_available()
allow_pickle = True
user_agent = {
"file_type": "text_inversion",
"framework": "pytorch",
}
# 1. Load textual inversion file
model_file = None
# Let's first try to load .safetensors weights
if (use_safetensors and weight_name is None) or (
weight_name is not None and weight_name.endswith(".safetensors")
):
try:
model_file = _get_model_file(
pretrained_model_name_or_path,
weights_name=weight_name or TEXT_INVERSION_NAME_SAFE,
cache_dir=cache_dir,
force_download=force_download,
resume_download=resume_download,
proxies=proxies,
local_files_only=local_files_only,
use_auth_token=use_auth_token,
revision=revision,
subfolder=subfolder,
user_agent=user_agent,
)
state_dict = safetensors.torch.load_file(model_file, device="cpu")
except Exception as e:
if not allow_pickle:
raise e
model_file = None
if model_file is None:
model_file = _get_model_file(
pretrained_model_name_or_path,
weights_name=weight_name or TEXT_INVERSION_NAME,
cache_dir=cache_dir,
force_download=force_download,
resume_download=resume_download,
proxies=proxies,
local_files_only=local_files_only,
use_auth_token=use_auth_token,
revision=revision,
subfolder=subfolder,
user_agent=user_agent,
)
state_dict = torch.load(model_file, map_location="cpu")
# 2. Load token and embedding correcly from file
if isinstance(state_dict, torch.Tensor):
if token is None:
raise ValueError(
"You are trying to load a textual inversion embedding that has been saved as a PyTorch tensor. Make sure to pass the name of the corresponding token in this case: `token=...`."
)
embedding = state_dict
elif len(state_dict) == 1:
# diffusers
loaded_token, embedding = next(iter(state_dict.items()))
elif "string_to_param" in state_dict:
# A1111
loaded_token = state_dict["name"]
embedding = state_dict["string_to_param"]["*"]
if token is not None and loaded_token != token:
logger.warn(f"The loaded token: {loaded_token} is overwritten by the passed token {token}.")
else:
token = loaded_token
embedding = embedding.to(dtype=self.text_encoder.dtype, device=self.text_encoder.device)
# 3. Make sure we don't mess up the tokenizer or text encoder
vocab = self.tokenizer.get_vocab()
if token in vocab:
raise ValueError(
f"Token {token} already in tokenizer vocabulary. Please choose a different token name or remove {token} and embedding from the tokenizer and text encoder."
)
elif f"{token}_1" in vocab:
multi_vector_tokens = [token]
i = 1
while f"{token}_{i}" in self.tokenizer.added_tokens_encoder:
multi_vector_tokens.append(f"{token}_{i}")
i += 1
raise ValueError(
f"Multi-vector Token {multi_vector_tokens} already in tokenizer vocabulary. Please choose a different token name or remove the {multi_vector_tokens} and embedding from the tokenizer and text encoder."
)
is_multi_vector = len(embedding.shape) > 1 and embedding.shape[0] > 1
if is_multi_vector:
tokens = [token] + [f"{token}_{i}" for i in range(1, embedding.shape[0])]
embeddings = [e for e in embedding] # noqa: C416
else:
tokens = [token]
embeddings = [embedding[0]] if len(embedding.shape) > 1 else [embedding]
# add tokens and get ids
self.tokenizer.add_tokens(tokens)
token_ids = self.tokenizer.convert_tokens_to_ids(tokens)
# resize token embeddings and set new embeddings
self.text_encoder.resize_token_embeddings(len(self.tokenizer))
for token_id, embedding in zip(token_ids, embeddings):
self.text_encoder.get_input_embeddings().weight.data[token_id] = embedding
logger.info(f"Loaded textual inversion embedding for {token}.")
class LoraLoaderMixin:
r"""
Utility class for handling the loading LoRA layers into UNet (of class [`UNet2DConditionModel`]) and Text Encoder
(of class [`CLIPTextModel`](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel)).
<Tip warning={true}>
This function is experimental and might change in the future.
</Tip>
"""
text_encoder_name = "text_encoder"
unet_name = "unet"
def load_lora_weights(self, pretrained_model_name_or_path_or_dict: Union[str, Dict[str, torch.Tensor]], **kwargs):
r"""
Load pretrained attention processor layers (such as LoRA) into [`UNet2DConditionModel`] and
[`CLIPTextModel`](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel)).
<Tip warning={true}>
This function is experimental and might change in the future.
</Tip>
Parameters:
pretrained_model_name_or_path_or_dict (`str` or `os.PathLike` or `dict`):
Can be either:
- A string, the *model id* of a pretrained model hosted inside a model repo on huggingface.co.
Valid model ids should have an organization name, like `google/ddpm-celebahq-256`.
- A path to a *directory* containing model weights saved using [`~ModelMixin.save_config`], e.g.,
`./my_model_directory/`.
- A [torch state
dict](https://pytorch.org/tutorials/beginner/saving_loading_models.html#what-is-a-state-dict).
cache_dir (`Union[str, os.PathLike]`, *optional*):
Path to a directory in which a downloaded pretrained model configuration should be cached if the
standard cache should not be used.
force_download (`bool`, *optional*, defaults to `False`):
Whether or not to force the (re-)download of the model weights and configuration files, overriding the
cached versions if they exist.
resume_download (`bool`, *optional*, defaults to `False`):
Whether or not to delete incompletely received files. Will attempt to resume the download if such a
file exists.
proxies (`Dict[str, str]`, *optional*):
A dictionary of proxy servers to use by protocol or endpoint, e.g., `{'http': 'foo.bar:3128',
'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request.
local_files_only(`bool`, *optional*, defaults to `False`):
Whether or not to only look at local files (i.e., do not try to download the model).
use_auth_token (`str` or *bool*, *optional*):
The token to use as HTTP bearer authorization for remote files. If `True`, will use the token generated
when running `diffusers-cli login` (stored in `~/.huggingface`).
revision (`str`, *optional*, defaults to `"main"`):
The specific model version to use. It can be a branch name, a tag name, or a commit id, since we use a
git-based system for storing models and other artifacts on huggingface.co, so `revision` can be any
identifier allowed by git.
subfolder (`str`, *optional*, defaults to `""`):
In case the relevant files are located inside a subfolder of the model repo (either remote in
huggingface.co or downloaded locally), you can specify the folder name here.
mirror (`str`, *optional*):
Mirror source to accelerate downloads in China. If you are from China and have an accessibility
problem, you can set this option to resolve it. Note that we do not guarantee the timeliness or safety.
Please refer to the mirror site for more information.
<Tip>
It is required to be logged in (`huggingface-cli login`) when you want to use private or [gated
models](https://huggingface.co/docs/hub/models-gated#gated-models).
</Tip>
"""
# Load the main state dict first which has the LoRA layers for either of
# UNet and text encoder or both.
cache_dir = kwargs.pop("cache_dir", DIFFUSERS_CACHE)
force_download = kwargs.pop("force_download", False)
resume_download = kwargs.pop("resume_download", False)
proxies = kwargs.pop("proxies", None)
local_files_only = kwargs.pop("local_files_only", HF_HUB_OFFLINE)
use_auth_token = kwargs.pop("use_auth_token", None)
revision = kwargs.pop("revision", None)
subfolder = kwargs.pop("subfolder", None)
weight_name = kwargs.pop("weight_name", None)
use_safetensors = kwargs.pop("use_safetensors", None)
if use_safetensors and not is_safetensors_available():
raise ValueError(
"`use_safetensors`=True but safetensors is not installed. Please install safetensors with `pip install safetenstors"
)
allow_pickle = False
if use_safetensors is None:
use_safetensors = is_safetensors_available()
allow_pickle = True
user_agent = {
"file_type": "attn_procs_weights",
"framework": "pytorch",
}
model_file = None
if not isinstance(pretrained_model_name_or_path_or_dict, dict):
# Let's first try to load .safetensors weights
if (use_safetensors and weight_name is None) or (
weight_name is not None and weight_name.endswith(".safetensors")
):
try:
model_file = _get_model_file(
pretrained_model_name_or_path_or_dict,
weights_name=weight_name or LORA_WEIGHT_NAME_SAFE,
cache_dir=cache_dir,
force_download=force_download,
resume_download=resume_download,
proxies=proxies,
local_files_only=local_files_only,
use_auth_token=use_auth_token,
revision=revision,
subfolder=subfolder,
user_agent=user_agent,
)
state_dict = safetensors.torch.load_file(model_file, device="cpu")
except IOError as e:
if not allow_pickle:
raise e
# try loading non-safetensors weights
pass
if model_file is None:
model_file = _get_model_file(
pretrained_model_name_or_path_or_dict,
weights_name=weight_name or LORA_WEIGHT_NAME,
cache_dir=cache_dir,
force_download=force_download,
resume_download=resume_download,
proxies=proxies,
local_files_only=local_files_only,
use_auth_token=use_auth_token,
revision=revision,
subfolder=subfolder,
user_agent=user_agent,
)
state_dict = torch.load(model_file, map_location="cpu")
else:
state_dict = pretrained_model_name_or_path_or_dict
# If the serialization format is new (introduced in https://github.com/huggingface/diffusers/pull/2918),
# then the `state_dict` keys should have `self.unet_name` and/or `self.text_encoder_name` as
# their prefixes.
keys = list(state_dict.keys())
# Load the layers corresponding to UNet.
if all(key.startswith(self.unet_name) for key in keys):
logger.info(f"Loading {self.unet_name}.")
unet_lora_state_dict = {k: v for k, v in state_dict.items() if k.startswith(self.unet_name)}
self.unet.load_attn_procs(unet_lora_state_dict)
# Load the layers corresponding to text encoder and make necessary adjustments.
elif all(key.startswith(self.text_encoder_name) for key in keys):
logger.info(f"Loading {self.text_encoder_name}.")
text_encoder_lora_state_dict = {
k: v for k, v in state_dict.items() if k.startswith(self.text_encoder_name)
}
attn_procs_text_encoder = self.load_attn_procs(text_encoder_lora_state_dict)
self._modify_text_encoder(attn_procs_text_encoder)
# Otherwise, we're dealing with the old format. This means the `state_dict` should only
# contain the module names of the `unet` as its keys WITHOUT any prefix.
elif not all(
key.startswith(self.unet_name) or key.startswith(self.text_encoder_name) for key in state_dict.keys()
):
self.unet.load_attn_procs(state_dict)
deprecation_message = "You have saved the LoRA weights using the old format. This will be"
" deprecated soon. To convert the old LoRA weights to the new format, you can first load them"
" in a dictionary and then create a new dictionary like the following:"
" `new_state_dict = {f'unet'.{module_name}: params for module_name, params in old_state_dict.items()}`."
deprecate("legacy LoRA weights", "1.0.0", deprecation_message, standard_warn=False)
def _modify_text_encoder(self, attn_processors: Dict[str, LoRAAttnProcessor]):
r"""
Monkey-patches the forward passes of attention modules of the text encoder.
Parameters:
attn_processors: Dict[str, `LoRAAttnProcessor`]:
A dictionary mapping the module names and their corresponding [`~LoRAAttnProcessor`].
"""
# Loop over the original attention modules.
for name, _ in self.text_encoder.named_modules():
if any([x in name for x in TEXT_ENCODER_TARGET_MODULES]):
# Retrieve the module and its corresponding LoRA processor.
module = self.text_encoder.get_submodule(name)
# Construct a new function that performs the LoRA merging. We will monkey patch
# this forward pass.
lora_layer = getattr(attn_processors[name], self._get_lora_layer_attribute(name))
old_forward = module.forward
def new_forward(x):
return old_forward(x) + lora_layer(x)
# Monkey-patch.
module.forward = new_forward
def _get_lora_layer_attribute(self, name: str) -> str:
if "q_proj" in name:
return "to_q_lora"
elif "v_proj" in name:
return "to_v_lora"
elif "k_proj" in name:
return "to_k_lora"
else:
return "to_out_lora"
def load_attn_procs(self, pretrained_model_name_or_path_or_dict: Union[str, Dict[str, torch.Tensor]], **kwargs):
r"""
Load pretrained attention processor layers for
[`CLIPTextModel`](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel).
<Tip warning={true}>
This function is experimental and might change in the future.
</Tip>
Parameters:
pretrained_model_name_or_path_or_dict (`str` or `os.PathLike` or `dict`):
Can be either:
- A string, the *model id* of a pretrained model hosted inside a model repo on huggingface.co.
Valid model ids should have an organization name, like `google/ddpm-celebahq-256`.
- A path to a *directory* containing model weights saved using [`~ModelMixin.save_config`], e.g.,
`./my_model_directory/`.
- A [torch state
dict](https://pytorch.org/tutorials/beginner/saving_loading_models.html#what-is-a-state-dict).
cache_dir (`Union[str, os.PathLike]`, *optional*):
Path to a directory in which a downloaded pretrained model configuration should be cached if the
standard cache should not be used.
force_download (`bool`, *optional*, defaults to `False`):
Whether or not to force the (re-)download of the model weights and configuration files, overriding the
cached versions if they exist.
resume_download (`bool`, *optional*, defaults to `False`):
Whether or not to delete incompletely received files. Will attempt to resume the download if such a
file exists.
proxies (`Dict[str, str]`, *optional*):
A dictionary of proxy servers to use by protocol or endpoint, e.g., `{'http': 'foo.bar:3128',
'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request.
local_files_only(`bool`, *optional*, defaults to `False`):
Whether or not to only look at local files (i.e., do not try to download the model).
use_auth_token (`str` or *bool*, *optional*):
The token to use as HTTP bearer authorization for remote files. If `True`, will use the token generated
when running `diffusers-cli login` (stored in `~/.huggingface`).
revision (`str`, *optional*, defaults to `"main"`):
The specific model version to use. It can be a branch name, a tag name, or a commit id, since we use a
git-based system for storing models and other artifacts on huggingface.co, so `revision` can be any
identifier allowed by git.
subfolder (`str`, *optional*, defaults to `""`):
In case the relevant files are located inside a subfolder of the model repo (either remote in
huggingface.co or downloaded locally), you can specify the folder name here.
mirror (`str`, *optional*):
Mirror source to accelerate downloads in China. If you are from China and have an accessibility
problem, you can set this option to resolve it. Note that we do not guarantee the timeliness or safety.
Please refer to the mirror site for more information.
Returns:
`Dict[name, LoRAAttnProcessor]`: Mapping between the module names and their corresponding
[`LoRAAttnProcessor`].
<Tip>
It is required to be logged in (`huggingface-cli login`) when you want to use private or [gated
models](https://huggingface.co/docs/hub/models-gated#gated-models).
</Tip>
"""
cache_dir = kwargs.pop("cache_dir", DIFFUSERS_CACHE)
force_download = kwargs.pop("force_download", False)
resume_download = kwargs.pop("resume_download", False)
proxies = kwargs.pop("proxies", None)
local_files_only = kwargs.pop("local_files_only", HF_HUB_OFFLINE)
use_auth_token = kwargs.pop("use_auth_token", None)
revision = kwargs.pop("revision", None)
subfolder = kwargs.pop("subfolder", None)
weight_name = kwargs.pop("weight_name", None)
use_safetensors = kwargs.pop("use_safetensors", None)
if use_safetensors and not is_safetensors_available():
raise ValueError(
"`use_safetensors`=True but safetensors is not installed. Please install safetensors with `pip install safetenstors"
)
allow_pickle = False
if use_safetensors is None:
use_safetensors = is_safetensors_available()
allow_pickle = True
user_agent = {
"file_type": "attn_procs_weights",
"framework": "pytorch",
}
model_file = None
if not isinstance(pretrained_model_name_or_path_or_dict, dict):
# Let's first try to load .safetensors weights
if (use_safetensors and weight_name is None) or (
weight_name is not None and weight_name.endswith(".safetensors")
):
try:
model_file = _get_model_file(
pretrained_model_name_or_path_or_dict,
weights_name=weight_name or LORA_WEIGHT_NAME_SAFE,
cache_dir=cache_dir,
force_download=force_download,
resume_download=resume_download,
proxies=proxies,
local_files_only=local_files_only,
use_auth_token=use_auth_token,
revision=revision,
subfolder=subfolder,
user_agent=user_agent,
)
state_dict = safetensors.torch.load_file(model_file, device="cpu")
except IOError as e:
if not allow_pickle:
raise e
# try loading non-safetensors weights
pass
if model_file is None:
model_file = _get_model_file(
pretrained_model_name_or_path_or_dict,
weights_name=weight_name or LORA_WEIGHT_NAME,
cache_dir=cache_dir,
force_download=force_download,
resume_download=resume_download,
proxies=proxies,
local_files_only=local_files_only,
use_auth_token=use_auth_token,
revision=revision,
subfolder=subfolder,
user_agent=user_agent,
)
state_dict = torch.load(model_file, map_location="cpu")
else:
state_dict = pretrained_model_name_or_path_or_dict
# fill attn processors
attn_processors = {}
is_lora = all("lora" in k for k in state_dict.keys())
if is_lora:
lora_grouped_dict = defaultdict(dict)
for key, value in state_dict.items():
attn_processor_key, sub_key = ".".join(key.split(".")[:-3]), ".".join(key.split(".")[-3:])
lora_grouped_dict[attn_processor_key][sub_key] = value
for key, value_dict in lora_grouped_dict.items():
rank = value_dict["to_k_lora.down.weight"].shape[0]
cross_attention_dim = value_dict["to_k_lora.down.weight"].shape[1]
hidden_size = value_dict["to_k_lora.up.weight"].shape[0]
attn_processors[key] = LoRAAttnProcessor(
hidden_size=hidden_size, cross_attention_dim=cross_attention_dim, rank=rank
)
attn_processors[key].load_state_dict(value_dict)
else:
raise ValueError(f"{model_file} does not seem to be in the correct format expected by LoRA training.")
# set correct dtype & device
attn_processors = {
k: v.to(device=self.device, dtype=self.text_encoder.dtype) for k, v in attn_processors.items()
}
return attn_processors
@classmethod
def save_lora_weights(
self,
save_directory: Union[str, os.PathLike],
unet_lora_layers: Dict[str, torch.nn.Module] = None,
text_encoder_lora_layers: Dict[str, torch.nn.Module] = None,
is_main_process: bool = True,
weight_name: str = None,
save_function: Callable = None,
safe_serialization: bool = False,
):
r"""
Save the LoRA parameters corresponding to the UNet and the text encoder.
Arguments:
save_directory (`str` or `os.PathLike`):
Directory to which to save. Will be created if it doesn't exist.
unet_lora_layers (`Dict[str, torch.nn.Module`]):
State dict of the LoRA layers corresponding to the UNet. Specifying this helps to make the
You can’t perform that action at this time.