-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathadaptors.py
1548 lines (1250 loc) · 50.3 KB
/
adaptors.py
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
import copy
import torch
import torch.nn as nn
import torch.nn.functional as F
from peft import (
LoraModel,
LoraConfig,
get_peft_model_state_dict,
set_peft_model_state_dict,
)
from peft import OFTModel, OFTConfig
from peft import LoKrModel, LoKrConfig
from peft import LoHaModel, LoHaConfig
from diffusers.loaders import AttnProcsLayers
from diffusers.models.attention_processor import LoRAAttnProcessor
from diffusers import (
AutoencoderKL,
DDPMScheduler,
DiffusionPipeline,
StableDiffusionPipeline,
DPMSolverMultistepScheduler,
)
from svdiff.utils import (
load_unet_for_svdiff,
load_text_encoder_for_svdiff,
SCHEDULER_MAPPING,
)
from diffusers.loaders import AttnProcsLayers
from difffit.unet_2d_condition import (
UNet2DConditionModel as UNet2DConditionModelForDiffFit,
)
from difffit.attention_processor import (
DiffFitAttnProcessor,
DiffFitXFormersAttnProcessor,
)
from difffit.modeling_clip import CLIPTextModel as CLIPTextModelForDiffFit
# from Scale_Shift_Features.unet_2d_condition import UNet2DConditionModel as UNet2DConditionModelForSSF
# from Scale_Shift_Features.attention_processor import SSFAttnProcessor, SSFXFormersAttnProcessor
# from Scale_Shift_Features.modeling_clip import CLIPTextModel as CLIPTextModelForDiffFit
from typing import Dict
import inspect
import accelerate
from accelerate.utils import set_module_tensor_to_device
from transformers import CLIPTextModel, CLIPTextConfig
from diffusers import UNet2DConditionModel
from safetensors.torch import safe_open
import huggingface_hub
def enable_grad(module):
for p in module.parameters():
p.requires_grad = True
def disable_grad(module):
for p in module.parameters():
p.requires_grad = False
def disable_grad_svdiff(module):
for n, p in module.named_parameters():
if "delta" in n:
p.requires_grad = False
def disable_grad_difffit(module):
for n, p in module.named_parameters():
if "gamma_" in n:
p.requires_grad = False
def disable_grad_attention(module):
for n, p in module.named_parameters():
if "attn" in n:
p.requires_grad = False
def set_module_grad_status(module, flag=False):
if isinstance(module, list):
print("list", module)
for m in module:
set_module_grad_status(m, flag)
else:
print("not a list", module)
for p in module.parameters():
p.requires_grad = flag
def return_param_sum(model, verbose=False):
total_magnitude = 0
# print("Calculating sum of trainable parameters")
for p in model.parameters():
if p.requires_grad:
total_magnitude += p.sum().item()
return round(total_magnitude, 3)
def return_blockwise_param_sum(model, verbose=False):
sum_dict_down_blocks = {}
sum_dict_mid_block = {}
sum_dict_up_blocks = {}
# print("Calculating the sum of parameters in U-net Down Blocks")
for idx, block in enumerate(model.down_blocks):
sum_dict_down_blocks["ALL_Down_Block_" + str(idx)] = return_param_sum(
block, False
)
# print("Calculating the sum of parameters in U-net Mid Block")
sum_dict_mid_block["ALL_Mid_Block"] = return_param_sum(model.mid_block)
# print("Calculating the sum of parameters in U-net Up Blocks")
for idx, block in enumerate(model.up_blocks):
sum_dict_up_blocks["ALL_Up_Block_" + str(idx)] = return_param_sum(block, False)
return sum_dict_down_blocks, sum_dict_mid_block, sum_dict_up_blocks
def return_norm_param_sum(model, verbose=False):
total_magnitude = 0
for n, p in model.named_parameters():
if p.requires_grad and "norm" in n:
total_magnitude += p.sum().item()
return round(total_magnitude, 3)
def return_blockwise_norm_param_sum(model, verbose=False):
sum_dict_down_blocks = {}
sum_dict_mid_block = {}
sum_dict_up_blocks = {}
# print("Calculating the sum of NORM parameters in U-net Down Blocks")
for idx, block in enumerate(model.down_blocks):
sum_dict_down_blocks["NORM_Down_Block_" + str(idx)] = return_norm_param_sum(
block, False
)
# print("Calculating the sum of NORM parameters in U-net Mid Block")
sum_dict_mid_block["NORM_Mid_Block"] = return_norm_param_sum(model.mid_block)
# print("Calculating the sum of NORM parameters in U-net Up Blocks")
for idx, block in enumerate(model.up_blocks):
sum_dict_up_blocks["NORM_Up_Block_" + str(idx)] = return_norm_param_sum(
block, False
)
return sum_dict_down_blocks, sum_dict_mid_block, sum_dict_up_blocks
def return_bias_param_sum(model, verbose=False):
total_magnitude = 0
for n, p in model.named_parameters():
if p.requires_grad and "bias" in n:
total_magnitude += p.sum().item()
return round(total_magnitude, 3)
def return_blockwise_bias_param_sum(model, verbose=False):
sum_dict_down_blocks = {}
sum_dict_mid_block = {}
sum_dict_up_blocks = {}
# print("Calculating the sum of BIAS parameters in U-net Down Blocks")
for idx, block in enumerate(model.down_blocks):
sum_dict_down_blocks["BIAS_Down_Block_" + str(idx)] = return_bias_param_sum(
block, False
)
# print("Calculating the sum of BIAS parameters in U-net Mid Block")
sum_dict_mid_block["BIAS_Mid_Block"] = return_bias_param_sum(model.mid_block)
# print("Calculating the sum of BIAS parameters in U-net Up Blocks")
for idx, block in enumerate(model.up_blocks):
sum_dict_up_blocks["BIAS_Up_Block_" + str(idx)] = return_bias_param_sum(
block, False
)
return sum_dict_down_blocks, sum_dict_mid_block, sum_dict_up_blocks
def return_attn_param_sum(model, verbose=False):
total_magnitude = 0
for n, p in model.named_parameters():
if p.requires_grad and "attentions" in n:
total_magnitude += p.sum().item()
return round(total_magnitude, 3)
def return_blockwise_attn_param_sum(model, verbose=False):
sum_dict_down_blocks = {}
sum_dict_mid_block = {}
sum_dict_up_blocks = {}
# print("Calculating the sum of ATTENTION parameters in U-net Down Blocks")
for idx, block in enumerate(model.down_blocks):
sum_dict_down_blocks["ATTN_Down_Block_" + str(idx)] = return_attn_param_sum(
block, False
)
# print("Calculating the sum of ATTENTION parameters in U-net Mid Block")
sum_dict_mid_block["ATTN_Mid_Block"] = return_attn_param_sum(model.mid_block)
# print("Calculating the sum of ATTENTION parameters in U-net Up Blocks")
for idx, block in enumerate(model.up_blocks):
sum_dict_up_blocks["ATTN_Up_Block_" + str(idx)] = return_attn_param_sum(
block, False
)
return sum_dict_down_blocks, sum_dict_mid_block, sum_dict_up_blocks
def check_params(model):
print("The following parameters are trainable:")
for n, p in model.named_parameters():
if p.requires_grad:
print(n)
def check_tunable_params(model, verbose=True):
"""
Prints the number of trainable parameters in the model.
"""
trainable_params = 0
all_param = 0
for name, param in model.named_parameters():
all_param += param.numel()
if param.requires_grad:
if verbose:
print(name)
trainable_params += param.numel()
# return f"trainable params: {trainable_params} || all params: {all_param} || trainable%: {100 * trainable_params / all_param:.5f} + \n"
print("\n")
print(
f"trainable params: {trainable_params} || all params: {all_param} || trainable%: {100 * trainable_params / all_param:.5f}"
)
return round(100 * trainable_params / all_param, 5)
# class conv_tsa(nn.Module):
# def __init__(self, orig_conv, ad_type):
# super().__init__()
# # the original conv layer
# self.conv = copy.deepcopy(orig_conv)
# self.conv.weight.requires_grad = False
# planes, in_planes, _, _ = self.conv.weight.size()
# stride, _ = self.conv.stride
# self.alpha = nn.Parameter(torch.ones(planes, in_planes, 1, 1))
# self.alpha.requires_grad = True
# self.ad_type = ad_type
# def forward(self, x):
# y = self.conv(x)
# if self.ad_type == "residual":
# if self.alpha.size(0) > 1:
# # residual adaptation in matrix form
# y = y + F.conv2d(x, self.alpha, stride=self.conv.stride)
# else:
# # residual adaptation in channel-wise (vector)
# y = y + x * self.alpha
# elif self.ad_type == "serial":
# if self.alpha.size(0) > 1:
# # serial adaptation in matrix form
# y = F.conv2d(y, self.alpha) + self.alpha_bias
# else:
# # serial adaptation in channel-wise (vector)
# y = y * self.alpha + self.alpha_bias
# return y
# class pa(nn.Module):
# """
# pre-classifier alignment (PA) mapping from 'Universal Representation Learning from Multiple Domains for Few-shot Classification'
# (https://arxiv.org/pdf/2103.13841.pdf)
# """
# def __init__(self, feat_dim):
# super(pa, self).__init__()
# # define pre-classifier alignment mapping
# self.weight = nn.Parameter(torch.ones(feat_dim, feat_dim, 1, 1))
# self.weight.requires_grad = True
# def forward(self, x):
# if len(list(x.size())) == 2:
# x = x.unsqueeze(-1).unsqueeze(-1)
# x = F.conv2d(x, self.weight.to(x.device)).flatten(1)
# return x
# def apply_tsa_to_unet(unet, apply_tsa_downsampler=False, apply_tsa_upsampler=False):
# print("Applying TSA to UNet")
# # Attaching TSA in Down blocks
# print("Iterating in Down Blocks")
# i = 0
# for block in unet.down_blocks:
# i += 1
# print("Block ", str(i))
# resnets = block.resnets
# if apply_tsa_downsampler:
# # Down present only in Blocks 1,2,3
# if i < 4:
# downsampler = block.downsamplers[0]
# setattr(downsampler, "conv", conv_tsa(downsampler.conv, "residual"))
# for resnet in resnets:
# setattr(resnet, "conv1", conv_tsa(resnet.conv1, "residual"))
# setattr(resnet, "conv2", conv_tsa(resnet.conv2, "residual"))
# # Attaching TSA in Mid blocks
# for resnet in unet.mid_block.resnets:
# setattr(resnet, "conv1", conv_tsa(resnet.conv1, "residual"))
# setattr(resnet, "conv2", conv_tsa(resnet.conv2, "residual"))
# print("\n")
# print("Iterating in Up Blocks")
# i = 0
# for block in unet.up_blocks:
# i += 1
# print("Block ", str(i))
# resnets = block.resnets
# if apply_tsa_upsampler:
# # Upsampler present only in Blocks 1,2,3
# if i < 4:
# upsampler = block.upsamplers[0]
# setattr(upsampler, "conv", conv_tsa(upsampler.conv, "residual"))
# for resnet in resnets:
# setattr(resnet, "conv1", conv_tsa(resnet.conv1, "residual"))
# setattr(resnet, "conv2", conv_tsa(resnet.conv2, "residual"))
# return unet
# def apply_tsa_to_vae(vae, apply_tsa_downsampler=False, apply_tsa_upsampler=False):
# print("Applying TSA to VAE")
# # Attaching TSA in Encoder
# i = 0
# for block in vae.encoder.down_blocks:
# i += 1
# print("Block ", str(i))
# resnets = block.resnets
# if apply_tsa_downsampler:
# # Downsampler present only in Block 1,2,3
# if i < 4:
# downsampler = block.downsamplers[0]
# setattr(downsampler, "conv", conv_tsa(downsampler.conv, "residual"))
# # Iterating in the resnet blocks of the encoder
# for resnet in resnets:
# setattr(resnet, "conv1", conv_tsa(resnet.conv1, "residual"))
# setattr(resnet, "conv2", conv_tsa(resnet.conv2, "residual"))
# for resnet in vae.encoder.mid_block.resnets:
# setattr(resnet, "conv1", conv_tsa(resnet.conv1, "residual"))
# setattr(resnet, "conv2", conv_tsa(resnet.conv2, "residual"))
# # Attaching TSA in Decoder
# i = 0
# for block in vae.decoder.up_blocks:
# i += 1
# print("Block ", str(i))
# resnets = block.resnets
# if apply_tsa_upsampler:
# # Upsampler present only in Block 1,2,3
# if i < 4:
# upsampler = block.upsamplers[0]
# setattr(upsampler, "conv", conv_tsa(upsampler.conv, "residual"))
# # Iterating in the resnet blocks of the decoder
# for resnet in resnets:
# setattr(resnet, "conv1", conv_tsa(resnet.conv1, "residual"))
# setattr(resnet, "conv2", conv_tsa(resnet.conv2, "residual"))
# for resnet in vae.decoder.mid_block.resnets:
# setattr(resnet, "conv1", conv_tsa(resnet.conv1, "residual"))
# setattr(resnet, "conv2", conv_tsa(resnet.conv2, "residual"))
# return vae
def get_lora_config(text_lora_rank=16):
config = {
"peft_config": {
"peft_type": "LORA",
"task_type": None,
"inference_mode": True,
"r": 16,
"target_modules": ["to_q", "to_v", "query", "value"],
"lora_alpha": 27,
"lora_dropout": 0.0,
"merge_weights": False,
"fan_in_fan_out": False,
"enable_lora": None,
"bias": "none",
},
"text_encoder_peft_config": {
"peft_type": "LORA",
"task_type": None,
"inference_mode": True,
"r": text_lora_rank,
"target_modules": ["q_proj", "v_proj"],
"lora_alpha": 17,
"lora_dropout": 0.0,
"merge_weights": False,
"fan_in_fan_out": False,
"enable_lora": None,
"bias": "none",
},
"vae_peft_config": {
"peft_type": "LORA",
"task_type": None,
"inference_mode": True,
"r": 16,
"target_modules": ["query", "value"],
"lora_alpha": 17,
"lora_dropout": 0.0,
"merge_weights": False,
"fan_in_fan_out": False,
"enable_lora": None,
"bias": "none",
},
}
return config
# def apply_lora_to_unet(unet, lora_config=None):
# """
# Usage:
# pipe.unet = apply_lora_to_unet(pipe.unet)
# Number of trainable parameters should be around 0.18%
# """
# print("Applying LORA to UNet")
# if lora_config is None:
# lora_config = get_lora_config()
# unet_config = LoraConfig(**lora_config["peft_config"])
# print("Tunable parameters before applying LORA: ")
# check_tunable_params(unet, verbose=False)
# unet = LoraModel(unet_config, unet)
# print("Tunable parameters after applying LORA: ")
# check_tunable_params(unet, verbose=False)
# return unet
def apply_lora_to_unetv2(unet, image_lora_rank, dtype):
# Set correct lora layers
# Snippet obtained from https://github.com/huggingface/diffusers/blob/7e6886f5e93ca9bb1e6d4beece46fe1e43b819c2/examples/text_to_image/train_text_to_image_lora.py#L446
unet.requires_grad_(False)
lora_attn_procs = {}
for name in unet.attn_processors.keys():
cross_attention_dim = (
None
if name.endswith("attn1.processor")
else unet.config.cross_attention_dim
)
if name.startswith("mid_block"):
hidden_size = unet.config.block_out_channels[-1]
elif name.startswith("up_blocks"):
block_id = int(name[len("up_blocks.")])
hidden_size = list(reversed(unet.config.block_out_channels))[block_id]
elif name.startswith("down_blocks"):
block_id = int(name[len("down_blocks.")])
hidden_size = unet.config.block_out_channels[block_id]
lora_attn_procs[name] = LoRAAttnProcessor(
hidden_size=hidden_size,
cross_attention_dim=cross_attention_dim,
rank=image_lora_rank,
dtype=dtype,
)
unet.set_attn_processor(lora_attn_procs)
lora_layers = AttnProcsLayers(unet.attn_processors)
return unet, lora_layers
def apply_svdiff_to_unet(args, cache_dir):
# Adapted from https://github.com/mkshing/svdiff-pytorch/blob/a78f69e14410c1963318806050a566d262eca9f8/train_svdiff.py#L717
if args is not None:
pretrained_model_name_or_path = args.pretrained_model_name_or_path
else:
pretrained_model_name_or_path = "runwayml/stable-diffusion-v1-5"
unet = load_unet_for_svdiff(
pretrained_model_name_or_path, subfolder="unet", cache_dir=cache_dir
)
unet.requires_grad_(False)
optim_params = []
optim_params_1d = []
for n, p in unet.named_parameters():
if "delta" in n:
p.requires_grad = True
if "norm" in n:
optim_params_1d.append(p)
else:
optim_params.append(p)
return unet, optim_params, optim_params_1d
def apply_svdiff_to_text_encoder(args):
pretrained_model_name_or_path = args.pretrained_model_name_or_path
text_encoder = load_text_encoder_for_svdiff(
pretrained_model_name_or_path, subfolder="text_encoder"
)
text_encoder.requires_grad_(False)
optim_params = []
optim_params_1d = []
for n, p in text_encoder.named_parameters():
if "delta" in n:
p.requires_grad = True
if "norm" in n:
optim_params_1d.append(p)
else:
optim_params.append(p)
return text_encoder, optim_params, optim_params_1d
def apply_lora_to_text_encoder(text_encoder, text_lora_rank=16, lora_config=None):
"""
Usage:
pipe.text_encoder = apply_lora_to_text_encoder(pipe.text_encoder)
Number of trainable parameters should be around 0.47%
"""
print("Applying LORA to Text Encoder")
if lora_config is None:
lora_config = get_lora_config(text_lora_rank=text_lora_rank)
text_encoder_config = LoraConfig(**lora_config["text_encoder_peft_config"])
print("Tunable parameters before applying LORA: ")
check_tunable_params(text_encoder, verbose=False)
text_encoder = LoraModel(text_encoder_config, text_encoder)
print("Tunable parameters after applying LORA to text encoder: ")
check_tunable_params(text_encoder, False)
return text_encoder
# def apply_lora_to_vae(vae, lora_config=None):
# """
# Usage:
# pipe.vae = apply_lora_to_vae(pipe.vae)
# Number of trainable parameters should be around
# """
# print("Applying LORA to VAE")
# if lora_config is None:
# lora_config = get_lora_config()
# vae_config = LoraConfig(**lora_config["vae_peft_config"])
# print("Tunable parameters before applying LORA: ")
# check_tunable_params(vae, verbose=False)
# vae = LoraModel(vae_config, vae)
# print("Tunable parameters after applying LORA")
# check_tunable_params(vae, False)
# return vae
def enable_bias_update(model):
print("Enabling Bias layers")
for m in model.modules():
for name, param in m.named_parameters():
if name == "bias":
param.requires_grad = True
def mark_only_biases_as_trainable(model: nn.Module, is_bitfit=False):
trainable_names = ["bias", "norm", "gamma", "y_embed"]
for par_name, par_tensor in model.named_parameters():
par_tensor.requires_grad = any([kw in par_name for kw in trainable_names])
return model
def enable_norm_update(model):
print("Enabling Normalization layers")
for m in model.modules():
for name, param in m.named_parameters():
if "norm" in name:
param.requires_grad = True
def enable_attention_update(model):
print("Enabling Attention layers")
for m in model.modules():
for name, param in m.named_parameters():
if "attentions" in name:
param.requires_grad = True
def disable_attention_update(model):
print("Disabling Attention layers")
for m in model.modules():
for name, param in m.named_parameters():
if "attentions" in name:
param.requires_grad = False
def enable_blockwise_attention_update(unet, block_idx=None):
if block_idx is not None:
assert block_idx <= 3, "Block index should be between 0 and 3"
if block_idx is None:
# Update all blocks
enable_attention_update(unet)
else:
# Update only the specified block
enable_attention_update(unet[block_idx])
def enable_attention_update_text_encoder(text_encoder):
print("Enabling Attention layers")
for n, p in text_encoder.text_model.named_parameters():
if "self_attn" in n:
p.requires_grad = True
def get_adapted_unet(unet, method, args, **kwargs):
if method == "attention":
disable_grad(unet)
enable_attention_update(unet)
verbose = True
elif method == "attention_down_blocks":
disable_grad(unet)
enable_blockwise_attention_update(
unet.down_blocks, block_idx=kwargs["unet_block_idx"]
)
verbose = True
elif method == "attention_up_blocks":
disable_grad(unet)
enable_blockwise_attention_update(
unet.up_blocks, block_idx=kwargs["unet_block_idx"]
)
verbose = True
elif method == "norm":
disable_grad(unet)
enable_norm_update(unet)
verbose = True
elif method == "bias":
disable_grad(unet)
enable_bias_update(unet)
verbose = True
elif method == "norm_bias":
disable_grad(unet)
enable_norm_update(unet)
enable_bias_update(unet)
verbose = True
elif method == "norm_bias_attention":
disable_grad(unet)
enable_norm_update(unet)
enable_bias_update(unet)
enable_attention_update(unet)
verbose = True
elif method == "lora":
unet = apply_lora_to_unet(unet)
verbose = True
elif method == "lorav2":
unet, lora_layers = apply_lora_to_unetv2(
unet, image_lora_rank=kwargs["image_lora_rank"], dtype=kwargs["dtype"]
)
return unet, lora_layers
elif method == "svdiff":
unet, optim_params, optim_params_1d = apply_svdiff_to_unet(
args=args, cache_dir=kwargs["cache_dir"]
)
return unet, optim_params, optim_params_1d
elif method == "difffit":
print("Loading model for DIFFFIT")
unet = load_unet_for_difffit(
pretrained_model_name_or_path=kwargs["pretrained_model_name_or_path"],
efficient_weights_ckpt=None,
hf_hub_kwargs=None,
is_bitfit=False,
cache_dir=kwargs["cache_dir"],
)
unet = mark_only_biases_as_trainable(unet, is_bitfit=False)
elif method == "ssf":
unet = load_unet_for_ssf(
pretrained_model_name_or_path=kwargs["pretrained_model_name_or_path"],
efficient_weights_ckpt=None,
hf_hub_kwargs=None,
is_bitfit=False,
)
unet = mark_only_ssf_as_trainable(unet)
elif method == "oft":
config_unet = OFTConfig(
r=args.image_oft_rank,
target_modules=[
"proj_in",
"proj_out",
"to_k",
"to_q",
"to_v",
"to_out.0",
"ff.net.0.proj",
"ff.net.2",
],
module_dropout=0.0,
init_weights=True,
)
unet = OFTModel(unet, config_unet, "default")
elif method == "loha":
# https://arxiv.org/abs/2108.06098
config_unet = LoHaConfig(
r=args.image_loha_rank,
alpha=32,
target_modules=[
"proj_in",
"proj_out",
"to_k",
"to_q",
"to_v",
"to_out.0",
"ff.net.0.proj",
"ff.net.2",
],
rank_dropout=0.0,
module_dropout=0.0,
init_weights=True,
use_effective_conv2d=True,
)
unet = LoHaModel(unet, config_unet, "default")
elif method == "lokr":
# https://arxiv.org/abs/2108.06098
# https://arxiv.org/abs/2309.14859
config_unet = LoKrConfig(
r=args.image_lokr_rank,
alpha=32,
target_modules=[
"proj_in",
"proj_out",
"to_k",
"to_q",
"to_v",
"to_out.0",
"ff.net.0.proj",
"ff.net.2",
],
rank_dropout=0.0,
module_dropout=0.0,
init_weights=True,
use_effective_conv2d=True,
)
unet = LoKrModel(unet, config_unet, "default")
elif method == "full":
verbose = False
pass
elif method == "full_without_attention":
verbose = False
disable_attention_update(unet)
elif method == "freeze":
verbose = False
disable_grad(unet)
else:
raise ValueError("Method not recognized")
# check_tunable_params(unet, verbose)
return unet
def get_adapted_text_encoder(text_encoder, method, args, **kwargs):
# Training only attention was introduced in https://arxiv.org/abs/2203.09795
if method == "attention":
disable_grad(text_encoder)
enable_attention_update_text_encoder(
text_encoder
) # Around 28348416 params (23.036%) trainable
verbose = True
elif method == "norm":
disable_grad(text_encoder)
enable_norm_update(text_encoder)
verbose = True
# Training only bias terms in transformers was introduced in BitFit: https://arxiv.org/abs/2106.10199
elif method == "bias":
disable_grad(text_encoder)
enable_bias_update(text_encoder)
verbose = True
elif method == "norm_bias":
disable_grad(text_encoder)
enable_norm_update(text_encoder)
enable_bias_update(text_encoder)
verbose = True
elif method == "norm_bias_attention":
disable_grad(text_encoder)
enable_norm_update(text_encoder)
enable_bias_update(text_encoder)
enable_attention_update(text_encoder)
verbose = True
# LORA: https://arxiv.org/abs/2106.09685
elif method == "lora":
text_encoder = apply_lora_to_text_encoder(
text_encoder, text_lora_rank=kwargs["text_lora_rank"]
)
verbose = True
elif method == "svdiff":
text_encoder, optim_params, optim_params_1d = apply_svdiff_to_text_encoder(args)
return text_encoder, optim_params, optim_params_1d
elif method == "difffit":
text_encoder = load_text_encoder_for_difffit(
pretrained_model_name_or_path=kwargs["pretrained_model_name_or_path"],
efficient_weights_ckpt=None,
hf_hub_kwargs=None,
is_bitfit=False,
)
text_encoder = mark_only_biases_as_trainable(text_encoder, is_bitfit=False)
elif method == "full":
verbose = False
pass
elif method == "freeze":
verbose = False
disable_grad(text_encoder)
else:
raise ValueError("Method not recognized")
return text_encoder
def get_adapted_vae(vae, method):
if method == "tsa":
# disable_grad(vae)
# vae = apply_tsa_to_vae(
# vae, apply_tsa_downsampler=False, apply_tsa_upsampler=False
# ) # 8.80% trainable
# # vae = apply_tsa_to_vae(vae, apply_tsa_downsampler=True, apply_tsa_upsampler=True) # 9.72% trainable
# verbose = True
raise NotImplementedError("TSA not implemented")
elif method == "tsa2":
# disable_grad(vae)
# vae = apply_tsa_to_vae(
# vae, apply_tsa_downsampler=True, apply_tsa_upsampler=True
# )
# verbose = True
raise NotImplementedError("TSA not implemented")
elif method == "attention":
# disable_grad(vae)
# enable_attention_update(vae) # Around 2103296 params (2.51%) trainable
# verbose = True
raise NotImplementedError("Attention Tuning not implemented for VAE")
elif method == "lora":
# vae = apply_lora_to_vae(
# vae, lora_config=None
# ) # Around 65536 params (0.07%) trainable
# verbose = True
raise NotImplementedError("LoRA not implemented for VAE")
elif method == "full":
verbose = False
pass
elif method == "freeze":
verbose = False
disable_grad(vae)
# check_tunable_params(vae, verbose)
return vae
############## DiffFit ################
def mark_only_biases_as_trainable(model: nn.Module, is_bitfit=False):
trainable_names = ["bias", "norm", "gamma", "y_embed"]
for par_name, par_tensor in model.named_parameters():
par_tensor.requires_grad = any([kw in par_name for kw in trainable_names])
return model
############## SSF ################
def mark_only_ssf_as_trainable(model):
print("Enabling shift and scale parameters")
for m in model.modules():
for name, param in m.named_parameters():
# print(name)
# if "shift" in name or "scale" in name:
# param.requires_grad = True
if "ssf_" in name:
param.requires_grad = True
return model
def get_state_dict(model: nn.Module) -> Dict[str, torch.Tensor]:
return {
par_name: par_tensor
for par_name, par_tensor in model.named_parameters()
if par_tensor.requires_grad
}
def load_config_for_difffit(model_path, **kwargs):
if os.path.exists(model_path):
if "config.json" not in model_path:
model_path = os.path.join(model_path, "config.json")
else:
model_path = huggingface_hub.hf_hub_download(
model_path, filename="config.json", **kwargs
)
with open(model_path, "r", encoding="utf-8") as f:
config = json.load(f)
return config
def load_unet_for_ssf(