-
Notifications
You must be signed in to change notification settings - Fork 63
/
train_tiktok.py
1437 lines (1328 loc) · 82.8 KB
/
train_tiktok.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
"""
Train a control net
"""
import os
import sys
import argparse
import datetime
import numpy as np
import pdb
# torch
import torch
from torch.utils.data import DataLoader
from torchvision.utils import make_grid, save_image
from torchvision import transforms as T
import torch.nn.functional as F
from torch.utils.tensorboard import SummaryWriter
#from ema_pytorch import EMA
import kornia
from einops import rearrange
import einops
# distributed
import torch.distributed as dist
from torch.distributed.optim import ZeroRedundancyOptimizer
from torch.nn.parallel import DistributedDataParallel as DDP
# data
from dataset import tiktok_video_arnold_copy
from dataset.hdfs_io import hcopy, hexists
from dataset.transforms import RemoveWhite, CenterCrop
# utils
from utils.checkpoint import load_from_pretrain, save_checkpoint_ema
from utils.utils import set_seed, count_param, print_peak_memory, anal_tensor, merge_lists_by_index
from utils.lr_scheduler import LambdaLinearScheduler
from dataset.hdfs_io import hexists, hmkdir, hopen, hcopy
from langdetect import detect
# model
import imageio
from PIL import Image
from model_lib.ControlNet.cldm.model import create_model, instantiate_from_config
import copy
TORCH_VERSION = torch.__version__.split(".")[0]
FP16_DTYPE = torch.float16 if TORCH_VERSION == "1" else torch.bfloat16
print(f"TORCH_VERSION={TORCH_VERSION} FP16_DTYPE={FP16_DTYPE}")
# torch.cuda.OutOfMemoryError: CUDA out of memory.
# Tried to allocate 120.00 MiB (GPU 0; 47.45 GiB total capacity; 43.81 GiB already allocated; 118.31 MiB free;
# 44.16 GiB reserved in total by PyTorch) If reserved memory is >> allocated memory try setting max_split_size_mb
# to avoid fragmentation. See documentation for Memory Management and PYTORCH_CUDA_ALLOC_CONF
def tensor_to_image(tensor):
# Assuming tensor shape is [1, 3, 64, 64]
# Convert the tensor to a numpy array and move the channel dimension to the last axis
image_np = tensor.squeeze(0).permute(1, 2, 0).cpu().numpy()
# Rescale the values from [-1, 1] to [0, 255]
image_np = ((image_np + 1) * 0.5 * 255).astype('uint8')
return Image.fromarray(image_np)
def build_mask_image(image,mask):
binary_mask = torch.where(mask <= 0, 0.0, 1.0)
result = binary_mask * image
return (1.0 - binary_mask)[:,0,...].unsqueeze(1), result
def build_face_mask_image(image, face_mask):
result = image * face_mask
return result, (1.0 - face_mask)
def is_english(text):
try:
lang = detect(text)
return lang == 'en'
except:
return False
def write_list_to_file(lst, filename):
with open(filename, 'w') as file:
for item in lst:
file.write(item + '\n')
def delete_zero_conv_in_controlmodel(state_dict):
keys_to_delete = [key for key in state_dict.keys() if (key.startswith('control_model.zero_convs') or key.startswith('control_model.middle_block_out')) ]
for key in keys_to_delete:
del state_dict[key]
return state_dict
# You can now use the modified state_dict without the deleted keys
def copy_diffusion_outputblocks(state_dict):
new_state_dict = copy.deepcopy(state_dict)
for key, value in state_dict.items():
if key.startswith('model.diffusion_model.output_blocks'):
new_key = key.replace('model.diffusion_model.output_blocks', 'control_model.output_blocks')
new_state_dict[new_key] = value
return new_state_dict
def load_state_dict(model, ckpt_path, reinit_hint_block=False, strict=True, map_location="cpu"):
print(f"Loading model state dict from {ckpt_path} ...")
state_dict = load_from_pretrain(ckpt_path, map_location=map_location)
state_dict = state_dict.get('state_dict', state_dict)
if reinit_hint_block:
print("Ignoring hint block parameters from checkpoint!")
for k in list(state_dict.keys()):
if k.startswith("control_model.input_hint_block"):
state_dict.pop(k)
model.load_state_dict(state_dict, strict=strict)
del state_dict
def load_state_dict_pretrained_everything_mm(model, pretrained_everything_path, mm_ckpt_path,strict=False, map_location="cpu"):
print(f"Loading pretrained_everything_path state dict from {pretrained_everything_path} ...")
print(f"Loading motion module state dict from {mm_ckpt_path} ...")
mm_state_dict = torch.load(mm_ckpt_path, map_location=map_location)
mm_state_dict = mm_state_dict.get('state_dict', mm_state_dict)
image_state_dict = load_from_pretrain(pretrained_everything_path, map_location=map_location)
image_state_dict = image_state_dict.get('state_dict', image_state_dict)
state_dict_final = merge_state_dict_mm(image_state_dict, mm_state_dict)
model.load_state_dict(state_dict_final, strict=True)
del image_state_dict, mm_state_dict, state_dict_final
def replace_keys_in_state_dict(state_dict,str1,str2):
new_state_dict = {}
for key, value in state_dict.items():
if str1 in key:
new_key = key.replace(str1, str2)
new_state_dict[new_key] = value
else:
new_state_dict[key] = value
return new_state_dict
def merge_state_dict(image_state_dict,pose_state_dict_new):
for key, value in pose_state_dict_new.items():
if "pose_control_model" in key:
image_state_dict[key] = value
else:
continue
return image_state_dict
def merge_state_dict_mm(image_state_dict,mm_state_dict):
for key, value in mm_state_dict.items():
if "motion_modules" in key:
if "down_blocks.0.motion_modules.0" in key:
image_state_dict[key.replace("down_blocks.0.motion_modules.0","model.diffusion_model.input_blocks_motion_module.0.0")] = value
elif "down_blocks.0.motion_modules.1" in key:
image_state_dict[key.replace("down_blocks.0.motion_modules.1","model.diffusion_model.input_blocks_motion_module.1.0")] = value
elif "down_blocks.1.motion_modules.0" in key:
image_state_dict[key.replace("down_blocks.1.motion_modules.0","model.diffusion_model.input_blocks_motion_module.2.0")] = value
elif "down_blocks.1.motion_modules.1" in key:
image_state_dict[key.replace("down_blocks.1.motion_modules.1","model.diffusion_model.input_blocks_motion_module.3.0")] = value
elif "down_blocks.2.motion_modules.0" in key:
image_state_dict[key.replace("down_blocks.2.motion_modules.0","model.diffusion_model.input_blocks_motion_module.4.0")] = value
elif "down_blocks.2.motion_modules.1" in key:
image_state_dict[key.replace("down_blocks.2.motion_modules.1","model.diffusion_model.input_blocks_motion_module.5.0")] = value
elif "down_blocks.3.motion_modules.0" in key:
image_state_dict[key.replace("down_blocks.3.motion_modules.0","model.diffusion_model.input_blocks_motion_module.6.0")] = value
elif "down_blocks.3.motion_modules.1" in key:
image_state_dict[key.replace("down_blocks.3.motion_modules.1","model.diffusion_model.input_blocks_motion_module.7.0")] = value
elif "up_blocks.0.motion_modules.0" in key:
image_state_dict[key.replace("up_blocks.0.motion_modules.0","model.diffusion_model.output_blocks_motion_module.0.0")] = value
elif "up_blocks.0.motion_modules.1" in key:
image_state_dict[key.replace("up_blocks.0.motion_modules.1","model.diffusion_model.output_blocks_motion_module.1.0")] = value
elif "up_blocks.0.motion_modules.2" in key:
image_state_dict[key.replace("up_blocks.0.motion_modules.2","model.diffusion_model.output_blocks_motion_module.2.0")] = value
elif "up_blocks.1.motion_modules.0" in key:
image_state_dict[key.replace("up_blocks.1.motion_modules.0","model.diffusion_model.output_blocks_motion_module.3.0")] = value
elif "up_blocks.1.motion_modules.1" in key:
image_state_dict[key.replace("up_blocks.1.motion_modules.1","model.diffusion_model.output_blocks_motion_module.4.0")] = value
elif "up_blocks.1.motion_modules.2" in key:
image_state_dict[key.replace("up_blocks.1.motion_modules.2","model.diffusion_model.output_blocks_motion_module.5.0")] = value
elif "up_blocks.2.motion_modules.0" in key:
image_state_dict[key.replace("up_blocks.2.motion_modules.0","model.diffusion_model.output_blocks_motion_module.6.0")] = value
elif "up_blocks.2.motion_modules.1" in key:
image_state_dict[key.replace("up_blocks.2.motion_modules.1","model.diffusion_model.output_blocks_motion_module.7.0")] = value
elif "up_blocks.2.motion_modules.2" in key:
image_state_dict[key.replace("up_blocks.2.motion_modules.2","model.diffusion_model.output_blocks_motion_module.8.0")] = value
elif "up_blocks.3.motion_modules.0" in key:
image_state_dict[key.replace("up_blocks.3.motion_modules.0","model.diffusion_model.output_blocks_motion_module.9.0")] = value
elif "up_blocks.3.motion_modules.1" in key:
image_state_dict[key.replace("up_blocks.3.motion_modules.1","model.diffusion_model.output_blocks_motion_module.10.0")] = value
elif "up_blocks.3.motion_modules.2" in key:
image_state_dict[key.replace("up_blocks.3.motion_modules.2","model.diffusion_model.output_blocks_motion_module.11.0")] = value
else:
continue
return image_state_dict
def load_state_dict_image_pose(model, image_ckpt_path, pose_ckpt_path, strict=False, map_location="cpu"):
print(f"Loading appearance model state dict from {image_ckpt_path} ...")
print(f"Loading pose model state dict from {pose_ckpt_path} ...")
state_dict = load_from_pretrain(image_ckpt_path, map_location=map_location)
state_dict = state_dict.get('state_dict', state_dict)
image_state_dict = replace_keys_in_state_dict(state_dict,"control_model","appearance_control_model")
# pdb.set_trace()
pose_state_dict = load_from_pretrain(pose_ckpt_path, map_location=map_location)
# pdb.set_trace()
pose_state_dict = pose_state_dict.get('state_dict', pose_state_dict)
# pdb.set_trace()
pose_state_dict_new = replace_keys_in_state_dict(pose_state_dict,"control_model","pose_control_model")
# pdb.set_trace()
state_dict_final = merge_state_dict(image_state_dict,pose_state_dict_new)
# pdb.set_trace()
model.load_state_dict(state_dict_final, strict=strict)
del state_dict, image_state_dict, pose_state_dict, pose_state_dict_new, state_dict_final
def load_state_dict_image_pose_mm(model, image_ckpt_path, pose_ckpt_path, mm_ckpt_path,strict=False, map_location="cpu"):
print(f"Loading appearance model state dict from {image_ckpt_path} ...")
print(f"Loading pose model state dict from {pose_ckpt_path} ...")
print(f"Loading pose model state dict from {mm_ckpt_path} ...")
pdb.set_trace()
mm_state_dict = torch.load(mm_ckpt_path, map_location=map_location)
# pdb.set_trace()
mm_state_dict = mm_state_dict.get('state_dict', mm_state_dict)
# pdb.set_trace()
model.load_state_dict(mm_state_dict, strict=False)
pdb.set_trace()
state_dict = load_from_pretrain(image_ckpt_path, map_location=map_location)
state_dict = state_dict.get('state_dict', state_dict)
image_state_dict = replace_keys_in_state_dict(state_dict,"control_model","appearance_control_model")
pose_state_dict = load_from_pretrain(pose_ckpt_path, map_location=map_location)
pose_state_dict = pose_state_dict.get('state_dict', pose_state_dict)
pose_state_dict_new = replace_keys_in_state_dict(pose_state_dict,"control_model","pose_control_model")
state_dict_final = merge_state_dict(image_state_dict,pose_state_dict_new)
model.load_state_dict(state_dict_final, strict=strict)
del state_dict, image_state_dict, pose_state_dict, pose_state_dict_new, state_dict_final
def load_state_dict_reference_only(model, ckpt_path, reinit_hint_block=False, strict=True, map_location="cpu"):
print(f"Loading model state dict from {ckpt_path} ...")
state_dict = load_from_pretrain(ckpt_path, map_location=map_location)
state_dict = state_dict.get('state_dict', state_dict)
state_dict = delete_zero_conv_in_controlmodel(state_dict)
new_state_dict = copy_diffusion_outputblocks(state_dict)
if reinit_hint_block:
print("Ignoring hint block parameters from checkpoint!")
for k in list(state_dict.keys()):
if k.startswith("control_model.input_hint_block"):
state_dict.pop(k)
model.load_state_dict(new_state_dict, strict=strict)
del state_dict
def load_state_dict_reference_only_mask(model, ckpt_path, reinit_hint_block=False, strict=True, map_location="cpu"):
print(f"Loading model state dict from {ckpt_path} ...")
state_dict = load_from_pretrain(ckpt_path, map_location=map_location)
state_dict = state_dict.get('state_dict', state_dict)
state_dict = delete_zero_conv_in_controlmodel(state_dict)
new_state_dict = copy_diffusion_outputblocks(state_dict)
conv_input_weight = new_state_dict.pop('control_model.input_blocks.0.0.weight')
if reinit_hint_block:
print("Ignoring hint block parameters from checkpoint!")
for k in list(state_dict.keys()):
if k.startswith("control_model.input_hint_block"):
state_dict.pop(k)
model.load_state_dict(new_state_dict, strict=strict)
# pdb.set_trace()
model.control_model.input_blocks[0][0].weight.data[:,0,...] = torch.zeros(320,3,3)
model.control_model.input_blocks[0][0].weight.data[:,1:,...] = conv_input_weight
# pdb.set_trace()
del state_dict
def get_cond_inpaint(model, batch_data, device, batch_size=None, blank_mask=False):
randommask = batch_data["randommask"][:batch_size]
if blank_mask:
randommask = torch.ones_like(randommask)
masked_image = batch_data["image"][:batch_size] * (1-randommask)
image = model.get_first_stage_encoding(model.encode_first_stage(masked_image.to(device)))
mask = torch.nn.functional.interpolate(randommask.to(device), size=image.shape[-2:])
inpaint = torch.cat((mask, image), 1)
return inpaint, masked_image
def get_cond_control(args, batch_data, control_type, device, model=None, batch_size=None, train=True):
control_type = copy.deepcopy(control_type)[0]
if control_type == "body+hand+face" :
if train:
if args.train_dataset != "tiktok_video_mm":
assert "pose_map" in batch_data
pose_map = batch_data["pose_map"].cuda()
c_cat = pose_map
cond_image = batch_data["condition_image"].cuda()
if args.mask_bg:
mask_bg, cond_image = build_mask_image(cond_image, batch_data['condition_mask'].cuda())
if args.mask_face:
cond_image, mask_face = build_face_mask_image(cond_image, batch_data['face_mask'].cuda())
if args.random_mask:
randommask = batch_data["randommask"].cuda()
cond_image = cond_image * (1-randommask)
cond_image = model.get_first_stage_encoding(model.encode_first_stage(cond_image))
if args.mask_bg:
mask = torch.nn.functional.interpolate(mask_bg.to(device), size=cond_image.shape[-2:])
cond_image = torch.cat((mask, cond_image), 1)
if args.random_mask:
mask = torch.nn.functional.interpolate(randommask.to(device), size=cond_image.shape[-2:])
cond_image = torch.cat((mask, cond_image), 1)
if args.mask_face:
mask = torch.nn.functional.interpolate(mask_face.to(device), size=cond_image.shape[-2:])
cond_image = torch.cat((mask, cond_image), 1)
cond_img_cat = cond_image
elif args.train_dataset == "tiktok_video_mm":
assert "pose_map_list" in batch_data
pose_map_list = batch_data["pose_map_list"].cuda()
# pdb.set_trace()
c_cat_list = [pose_map_list[:,i,:,:,:].cuda() for i in range(pose_map_list.shape[1])]
# pdb.set_trace()
cond_image = batch_data["condition_image"].cuda()
cond_image = model.get_first_stage_encoding(model.encode_first_stage(cond_image))
cond_img_cat = cond_image
else:
raise NotImplementedError
else:
if args.train_dataset != "tiktok_video_mm":
assert "pose_map_list" in batch_data
pose_map_list = batch_data["pose_map_list"]
print("Get Inference Control Type")
c_cat_list = [pose_map.cuda() for pose_map in pose_map_list]
cond_image = batch_data["condition_image"].cuda()
if args.mask_bg:
mask_bg, cond_image = build_mask_image(cond_image, batch_data['condition_mask'].cuda())
if args.mask_face:
cond_image, mask_face = build_face_mask_image(cond_image, batch_data['face_mask'].cuda())
if args.random_mask:
randommask = batch_data["randommask"].cuda()
cond_image = cond_image * (1-randommask)
cond_image = model.get_first_stage_encoding(model.encode_first_stage(cond_image))
if args.mask_bg:
mask = torch.nn.functional.interpolate(mask_bg.to(device), size=cond_image.shape[-2:])
cond_image = torch.cat((mask, cond_image), 1)
if args.random_mask:
mask = torch.nn.functional.interpolate(randommask.to(device), size=cond_image.shape[-2:])
cond_image = torch.cat((mask, cond_image), 1)
if args.mask_face:
mask = torch.nn.functional.interpolate(mask_face.to(device), size=cond_image.shape[-2:])
cond_image = torch.cat((mask, cond_image), 1)
cond_img_cat = cond_image
elif args.train_dataset == "tiktok_video_mm":
assert "pose_map_list" in batch_data
pose_map_list = batch_data["pose_map_list"].cuda()
c_cat_list = [pose_map_list[:,i,:,:,:].cuda() for i in range(pose_map_list.shape[1])]
cond_image = batch_data["condition_image"].cuda()
cond_image = model.get_first_stage_encoding(model.encode_first_stage(cond_image))
cond_img_cat = cond_image
else:
raise NotImplementedError
else:
raise NotImplementedError(f"cond_type={control_type} not supported!")
if train:
if args.train_dataset != "tiktok_video_mm":
if args.control_dropout > 0:
mask = torch.rand((c_cat.shape[0],1,1,1)) > args.control_dropout
c_cat = c_cat * mask.type(torch.float32).to(device)
return [ c_cat[:batch_size].to(device) ], [cond_img_cat[:batch_size].to(device) ]
else:
return_list = []
for c_cat in c_cat_list:
if args.control_dropout > 0:
mask = torch.rand((c_cat.shape[0],1,1,1)) > args.control_dropout
c_cat = c_cat * mask.type(torch.float32).to(device)
# pdb.set_trace()
return_list.append([c_cat[:batch_size].to(device)])
# pdb.set_trace()
return return_list, [cond_img_cat[:batch_size].to(device) ]
else:
return_list = []
for c_cat in c_cat_list:
if args.control_dropout > 0:
mask = torch.rand((c_cat.shape[0],1,1,1)) > args.control_dropout
c_cat = c_cat * mask.type(torch.float32).to(device)
# pdb.set_trace()
return_list.append([c_cat[:batch_size].to(device)])
# pdb.set_trace()
return return_list, [cond_img_cat[:batch_size].to(device) ]
def visualize(args, name, batch_data, tb_writer, infer_model, global_step, nSample, nTest=1):
infer_model.eval()
if args.train_dataset != "tiktok_video_mm":
real_image_list = [real_image.cuda() for real_image in batch_data["image_list"]]
else:
image_list = batch_data["image_list"].to(args.device) # B,F,C,H,W
rec_image_list = []
if not args.v4:
text = [" "] * nSample
else:
text = [" "] * nSample
if not args.with_text:
for i in range(len(text)):
text[i] = ""
if args.text_prompt is not None:
for i in range(len(text)):
text[i] = args.text_prompt
if args.train_dataset == "tiktok_video_mm":
c_cat_list, cond_img_cat = get_cond_control(args, batch_data, args.control_type, args.device, model=infer_model, batch_size=nSample,train=False)
pose_map_list = [pose_map[0].cpu() for pose_map in c_cat_list]
pose_tensor_list = torch.cat([c_cat[0] for c_cat in c_cat_list],0)
c_cat_list = [pose_tensor_list]
c_cross = infer_model.get_learned_conditioning(text)[:nSample]
uc_cross = infer_model.get_unconditional_conditioning(nSample)
c_cross = einops.repeat(c_cross, 'b h w -> (repeat b) h w', repeat=args.frame_num)
uc_cross = einops.repeat(uc_cross, 'b h w -> (repeat b) h w', repeat=args.frame_num)
cond_img_cat = [einops.repeat(cond_img_cat[0], 'b c h w -> (repeat b) c h w', repeat=args.frame_num)]
c = {"c_concat": c_cat_list, "c_crossattn": [c_cross], "image_control": cond_img_cat}
if args.control_mode == "controlnet_important":
uc = {"c_concat": c_cat_list, "c_crossattn": [uc_cross]}
else:
uc = {"c_concat": c_cat_list, "c_crossattn": [uc_cross], "image_control":cond_img_cat}
if args.wonoise:
c['wonoise'] = True
uc['wonoise'] = True
else:
c['wonoise'] = False
uc['wonoise'] = False
noise_shape = (c_cross.shape[0], infer_model.channels, infer_model.image_size, infer_model.image_size)
noise = torch.randn(noise_shape, device=c_cross.device)
with torch.cuda.amp.autocast(enabled=args.use_fp16, dtype=FP16_DTYPE):
# for _ in range(nTest):
infer_model.to(args.device)
infer_model.eval()
gene_img, _ = infer_model.sample_log(cond=c,
batch_size=c_cross.shape[0], ddim=True,
ddim_steps=50, eta=args.eta,
unconditional_guidance_scale=7,
unconditional_conditioning=uc,
inpaint=None,
x_T=noise
)
gene_img = infer_model.decode_first_stage( gene_img )
gene_img = gene_img.clamp(-1, 1).cpu()
image_list = rearrange(image_list, "b f c h w -> (b f) c h w")
latent = infer_model.get_first_stage_encoding(infer_model.encode_first_stage(image_list))
rec_image = infer_model.decode_first_stage(latent)
gene_img_list = [gene_img[i,:,:,:].unsqueeze(0).cpu() for i in range(gene_img.shape[0])]
rec_image_list = [rec_image[i,:,:,:].unsqueeze(0).cpu() for i in range(rec_image.shape[0])]
grid = make_grid( torch.cat([rec_img.cpu() for rec_img in rec_image_list]+ pose_map_list + gene_img_list + [batch_data["condition_image"].cpu()]).clamp(-1,1).add(1).mul(0.5), nrow=len(rec_image_list))
if args.with_text:
write_list_to_file(text, f"{args.local_image_dir}/{name}_{global_step}.txt")
save_image(grid, f"{args.local_image_dir}/{name}_{global_step}.jpg")
imageio.mimsave(f"{args.local_image_dir}/{name}_{global_step}_gen.gif", [tensor_to_image(tensor.cpu()) for tensor in gene_img_list], duration=args.gif_time) # Adjust the duration between frames as needed
imageio.mimsave(f"{args.local_image_dir}/{name}_{global_step}_pose.gif", [tensor_to_image(tensor.clamp(-1, 1).cpu()) for tensor in pose_map_list], duration=args.gif_time)
imageio.mimsave(f"{args.local_image_dir}/{name}_{global_step}_gt.gif", [tensor_to_image(tensor.clamp(-1, 1).cpu()) for tensor in rec_image_list], duration=args.gif_time) # Adjust the duration between frames as needed
else:
c_cat_list, cond_img_cat = get_cond_control(args, batch_data, args.control_type, args.device, model=infer_model, batch_size=nSample,train=False)
c_cross = infer_model.get_learned_conditioning(text)[:nSample]
uc_cross = infer_model.get_unconditional_conditioning(nSample)
gene_img_list = []
pose_map_list = []
for img_num in range(len(real_image_list)):
print("Generate Image {} in {} images".format(img_num,len(real_image_list)))
# pdb.set_trace()
c = {"c_concat": c_cat_list[img_num], "c_crossattn": [c_cross], "image_control":cond_img_cat}
if args.control_mode == "controlnet_important":
uc = {"c_concat": c_cat_list[img_num], "c_crossattn": [uc_cross]}
else:
uc = {"c_concat": c_cat_list[img_num], "c_crossattn": [uc_cross], "image_control":cond_img_cat}
if args.wonoise:
c['wonoise'] = True
uc['wonoise'] = True
else:
c['wonoise'] = False
uc['wonoise'] = False
pose_map_list.append(c_cat_list[img_num][0][:,:3].cpu())
if args.inpaint_unet:
inpaint, masked_image = get_cond_inpaint(infer_model, batch_data, args.device, batch_size=nSample, blank_mask=False)
inpaint_list = [masked_image.cpu()]
else:
inpaint = None
inpaint_list = []
noise_shape = (c_cross.shape[0], infer_model.channels, infer_model.image_size, infer_model.image_size)
noise = torch.randn(noise_shape, device=c_cross.device)
# generate images
with torch.cuda.amp.autocast(enabled=args.use_fp16, dtype=FP16_DTYPE):
# for _ in range(nTest):
infer_model.to(args.device)
infer_model.eval()
gene_img, _ = infer_model.sample_log(cond=c,
batch_size=nSample, ddim=True,
ddim_steps=50, eta=args.eta,
unconditional_guidance_scale=7,
unconditional_conditioning=uc,
inpaint=inpaint,
x_T=noise,
)
gene_img = infer_model.decode_first_stage( gene_img )
gene_img_list.append(gene_img.clamp(-1, 1).cpu())
image = real_image_list[img_num][:nSample]
# mask = real_mask_list[img_num][:nSample]
latent = infer_model.get_first_stage_encoding(infer_model.encode_first_stage(image))
rec_image = infer_model.decode_first_stage(latent)
rec_image_list.append(rec_image)
if args.random_mask:
cond_image = batch_data["condition_image"].cpu() * (1-batch_data["randommask"].cpu())
elif args.mask_bg:
_, cond_image = build_mask_image(batch_data["condition_image"].cuda(), batch_data['condition_mask'].cuda())
elif args.mask_face:
cond_image, _ = build_face_mask_image(batch_data["condition_image"].cuda(), batch_data['face_mask'].cuda())
else:
cond_image = batch_data["condition_image"].cpu()
grid = make_grid( torch.cat([rec_img.cpu() for rec_img in rec_image_list]+ pose_map_list + gene_img_list + [cond_image.cpu()]).clamp(-1,1).add(1).mul(0.5), nrow=len(rec_image_list))
if args.with_text:
write_list_to_file(text, f"{args.local_image_dir}/{name}_{args.control_type}_{global_step}.txt")
save_image(grid, f"{args.local_image_dir}/{name}_{args.control_type}_{global_step}.jpg")
tb_writer.add_image(f'{name}_{args.control_type}', grid, global_step)
tb_writer.add_text(f'{name}_{args.control_type}_caption',f"{str(text)}", global_step )
def estimate_deviation(args, infer_model, tb_writer, global_step):
def _calc_dist(model1, model2, keyword, replace=None):
with torch.no_grad():
model1 = model1.state_dict()
model2 = model2.state_dict()
if replace is None: replace = lambda x: x
distance = 0
keys = [k for k in model1.keys() if keyword in k and replace(k) in model2.keys()]
for k in keys:
distance += (model1[k].float() - model2[replace(k)].float()).pow(2).sum()
return distance / len(keys)
if args.local_rank == 0:
log_output = "Deviation: "
def main(args):
# ******************************
# initialize training
# ******************************
# assign rank
args.world_size = int(os.environ['WORLD_SIZE'])
args.local_rank = int(os.environ['LOCAL_RANK'])
args.rank = int(os.environ['RANK'])
args.device = torch.device("cuda", args.local_rank)
os.makedirs(args.local_image_dir,exist_ok=True)
os.makedirs(args.local_log_dir,exist_ok=True)
if args.rank == 0:
print(args)
# initial distribution comminucation
dist.init_process_group("nccl", rank=args.rank, world_size=args.world_size)
torch.backends.cuda.matmul.allow_tf32 = True
torch.backends.cudnn.benchmark = True
# set seed for reproducibility
set_seed(args.seed)
# visdom / tensorboard
if args.rank == 0:
tb_writer = SummaryWriter(log_dir=args.local_log_dir)
else:
tb_writer = None
# ******************************
# create model
# ******************************
model = create_model(args.model_config).cpu()
model.sd_locked = args.sd_locked
model.only_mid_control = args.only_mid_control
model.to(args.local_rank)
if args.local_rank == 0:
print('Total base parameters {:.02f}M'.format(count_param([model])))
if args.ema_rate is not None and args.ema_rate > 0 and args.rank == 0:
print(f"Creating EMA model at ema_rate={args.ema_rate}")
model_ema = EMA(model, beta=args.ema_rate, update_after_step=0, update_every=1)
else:
model_ema = None
# ******************************
# load pre-trained models
# ******************************
optimizer_state_dict = None
global_step = args.global_step
args.resume_dir = args.resume_dir or args.output_dir
if args.resume_dir is not None and hexists(os.path.join(args.resume_dir, "optimizer_state_latest.th")):
print('find optimizer state dict from {} ...'.format(os.path.join(args.resume_dir, "optimizer_state_latest.th")))
optimizer_state_dict = load_from_pretrain(os.path.join(args.resume_dir, "optimizer_state_latest.th"), map_location="cpu")
if global_step == 0:
global_step = optimizer_state_dict["step"]
if args.local_rank == 0:
assert hexists(os.path.join(args.resume_dir, f"model_state-{global_step}.th"))
load_state_dict(model, os.path.join(args.resume_dir, f"model_state-{global_step}.th"))
if args.rank == 0 and model_ema is not None:
if hexists(os.path.join(args.resume_dir, f"model_state-{global_step}_ema_{args.ema_rate}.th")):
load_state_dict(model_ema.ema_model, os.path.join(args.resume_dir, f"model_state-{global_step}_ema_{args.ema_rate}.th"))
else:
model_ema.copy_params_from_model_to_ema()
else:
# Initializing the model with SD weights
if args.image_pretrain_dir is None:
assert args.init_path is not None
if args.pose_pretrain_dir is not None:
print("Load appearance coontrol model weights from init path.")
print('find pose_pretrain model from {} ...'.format(os.path.join(args.pose_pretrain_dir, "optimizer_state_latest.th")))
if not args.pose_pretrain_dir.endswith(".pth"):
optimizer_state_dict_pose = load_from_pretrain(os.path.join(args.pose_pretrain_dir, "optimizer_state_latest.th"), map_location="cpu")
else:
optimizer_state_dict_pose = None
if global_step == 0:
if not args.pose_pretrain_dir.endswith(".pth"):
global_step_pose = optimizer_state_dict_pose["step"]
else:
global_step_pose = None
optimizer_state_dict_pose = None
if args.local_rank == 0:
if not args.pose_pretrain_dir.endswith(".pth"):
assert hexists(os.path.join(args.pose_pretrain_dir, f"model_state-{global_step_pose}.th"))
if "model_lib/ControlNet/models/cldm_v15_reference_only_temporal_pose.yaml" in args.model_config:
print("Load pretrained reference only controlnet and pose controlnet for motion module+pose controlnet training")
if args.pose_pretrain_dir.endswith(".pth"):
load_state_dict_image_pose(model, args.init_path, args.pose_pretrain_dir, strict=False)
else:
load_state_dict_image_pose(model, args.init_path, os.path.join(args.pose_pretrain_dir, f"model_state-{global_step_pose}.th"), strict=False)
else:
if args.pose_pretrain_dir.endswith(".pth"):
load_state_dict_image_pose(model, args.init_path, args.pose_pretrain_dir, strict=False)
else:
load_state_dict_image_pose(model, args.init_path, os.path.join(args.pose_pretrain_dir, f"model_state-{global_step_pose}.th"), strict=False)
if args.rank == 0 and model_ema is not None:
model_ema.copy_params_from_model_to_ema()
else:
if args.pretrained_everything is not None:
print('find model state dict from {} ...'.format(args.pretrained_everything))
if args.local_rank == 0:
assert hexists(args.pretrained_everything)
if args.mm_pretrain_dir is not None:
print('find pertrained checkpoint from {} ...'.format(args.pretrained_everything))
print('find motion module from {} ...'.format(args.mm_pretrain_dir))
load_state_dict_pretrained_everything_mm(model, args.pretrained_everything, args.mm_pretrain_dir, strict=False)
else:
load_state_dict(model, args.pretrained_everything ,strict=False)
if model_ema is not None:
model_ema.copy_params_from_model_to_ema()
else:
print("Load model weights from init path.")
if args.model_config == "model_lib/ControlNet/models/cldm_v15_reference_only.yaml" or args.model_config == "./model_lib/ControlNet/models/cldm_v15_reference_only.yaml":
print("Reference Only Control.")
load_state_dict_reference_only(model, args.init_path, reinit_hint_block=args.reinit_hint_block, strict=True)
elif args.model_config == "model_lib/ControlNet/models/cldm_v15_reference_only_mask.yaml" or args.model_config == "./model_lib/ControlNet/models/cldm_v15_reference_only_mask.yaml":
print("Reference Only Control with Mask.")
load_state_dict_reference_only_mask(model, args.init_path, reinit_hint_block=args.reinit_hint_block, strict=False)
else:
load_state_dict(model, args.init_path, reinit_hint_block=args.reinit_hint_block, strict=False)
if model_ema is not None:
model_ema.copy_params_from_model_to_ema()
# Load weights from Image Model in Step 1
else:
if args.pose_pretrain_dir is None:
print('find image_pretrain model from {} ...'.format(os.path.join(args.image_pretrain_dir, "optimizer_state_latest.th")))
optimizer_state_dict = load_from_pretrain(os.path.join(args.image_pretrain_dir, "optimizer_state_latest.th"), map_location="cpu")
if global_step == 0:
global_step = optimizer_state_dict["step"]
optimizer_state_dict = None
if args.local_rank == 0:
assert hexists(os.path.join(args.image_pretrain_dir, f"model_state-{global_step}.th"))
if "model_lib/ControlNet/models/cldm_v15_reference_only_temporal.yaml" in args.model_config:
print("Load pretrained reference only controlnet for motion module training")
load_state_dict(model, os.path.join(args.image_pretrain_dir, f"model_state-{global_step}.th"),strict=False)
else:
load_state_dict(model, os.path.join(args.image_pretrain_dir, f"model_state-{global_step}.th"),strict=False)
if args.rank == 0 and model_ema is not None:
if hexists(os.path.join(args.image_pretrain_dir, f"model_state-{global_step}_ema_{args.ema_rate}.th")):
load_state_dict(model_ema.ema_model, os.path.join(args.image_pretrain_dir, f"model_state-{global_step}_ema_{args.ema_rate}.th"),strict=False)
else:
model_ema.copy_params_from_model_to_ema()
else:
if args.mm_pretrain_dir is not None:
print('find image_pretrain model from {} ...'.format(os.path.join(args.image_pretrain_dir, "optimizer_state_latest.th")))
print('find pose_pretrain model from {} ...'.format(os.path.join(args.pose_pretrain_dir, "optimizer_state_latest.th")))
print('find motion module from {} ...'.format(args.mm_pretrain_dir))
optimizer_state_dict_image = load_from_pretrain(os.path.join(args.image_pretrain_dir, "optimizer_state_latest.th"), map_location="cpu")
if not args.pose_pretrain_dir.endswith(".pth"):
optimizer_state_dict_pose = load_from_pretrain(os.path.join(args.pose_pretrain_dir, "optimizer_state_latest.th"), map_location="cpu")
else:
optimizer_state_dict_pose = None
if global_step == 0:
global_step_image = optimizer_state_dict_image["step"]
if not args.pose_pretrain_dir.endswith(".pth"):
global_step_pose = optimizer_state_dict_pose["step"]
else:
global_step_pose = None
optimizer_state_dict_image, optimizer_state_dict_pose = None, None
if args.local_rank == 0:
assert hexists(os.path.join(args.image_pretrain_dir, f"model_state-{global_step_image}.th"))
if not args.pose_pretrain_dir.endswith(".pth"):
assert hexists(os.path.join(args.pose_pretrain_dir, f"model_state-{global_step_pose}.th"))
if "model_lib/ControlNet/models/cldm_v15_reference_only_temporal_pose.yaml" in args.model_config:
print("Load pretrained reference only controlnet, pose controlnet and motion module for motion module+pose controlnet training")
if args.pose_pretrain_dir.endswith(".pth"):
load_state_dict_image_pose_mm(model, os.path.join(args.image_pretrain_dir, f"model_state-{global_step_image}.th"), args.pose_pretrain_dir, args.mm_pretrain_dir, strict=False)
else:
load_state_dict_image_pose_mm(model, os.path.join(args.image_pretrain_dir, f"model_state-{global_step_image}.th"), os.path.join(args.pose_pretrain_dir, f"model_state-{global_step_pose}.th"), args.mm_pretrain_dir, strict=False)
else:
if args.pose_pretrain_dir.endswith(".pth"):
load_state_dict_image_pose_mm(model, os.path.join(args.image_pretrain_dir, f"model_state-{global_step_image}.th"), args.pose_pretrain_dir, args.mm_pretrain_dir, strict=False)
else:
load_state_dict_image_pose_mm(model, os.path.join(args.image_pretrain_dir, f"model_state-{global_step_image}.th"), os.path.join(args.pose_pretrain_dir, f"model_state-{global_step_pose}.th"), args.mm_pretrain_dir, strict=False)
if args.rank == 0 and model_ema is not None:
model_ema.copy_params_from_model_to_ema()
else:
print('find image_pretrain model from {} ...'.format(os.path.join(args.image_pretrain_dir, "optimizer_state_latest.th")))
print('find pose_pretrain model from {} ...'.format(os.path.join(args.pose_pretrain_dir, "optimizer_state_latest.th")))
optimizer_state_dict_image = load_from_pretrain(os.path.join(args.image_pretrain_dir, "optimizer_state_latest.th"), map_location="cpu")
if not args.pose_pretrain_dir.endswith(".pth"):
optimizer_state_dict_pose = load_from_pretrain(os.path.join(args.pose_pretrain_dir, "optimizer_state_latest.th"), map_location="cpu")
else:
optimizer_state_dict_pose = None
if global_step == 0:
global_step_image = optimizer_state_dict_image["step"]
if not args.pose_pretrain_dir.endswith(".pth"):
global_step_pose = optimizer_state_dict_pose["step"]
else:
global_step_pose = None
optimizer_state_dict_image, optimizer_state_dict_pose = None, None
if args.local_rank == 0:
assert hexists(os.path.join(args.image_pretrain_dir, f"model_state-{global_step_image}.th"))
if not args.pose_pretrain_dir.endswith(".pth"):
assert hexists(os.path.join(args.pose_pretrain_dir, f"model_state-{global_step_pose}.th"))
if "model_lib/ControlNet/models/cldm_v15_reference_only_temporal_pose.yaml" in args.model_config:
print("Load pretrained reference only controlnet and pose controlnet for motion module+pose controlnet training")
if args.pose_pretrain_dir.endswith(".pth"):
load_state_dict_image_pose(model, os.path.join(args.image_pretrain_dir, f"model_state-{global_step_image}.th"), args.pose_pretrain_dir, strict=False)
else:
load_state_dict_image_pose(model, os.path.join(args.image_pretrain_dir, f"model_state-{global_step_image}.th"), os.path.join(args.pose_pretrain_dir, f"model_state-{global_step_pose}.th"), strict=False)
else:
if args.pose_pretrain_dir.endswith(".pth"):
load_state_dict_image_pose(model, os.path.join(args.image_pretrain_dir, f"model_state-{global_step_image}.th"), args.pose_pretrain_dir, strict=False)
else:
load_state_dict_image_pose(model, os.path.join(args.image_pretrain_dir, f"model_state-{global_step_image}.th"), os.path.join(args.pose_pretrain_dir, f"model_state-{global_step_pose}.th"), strict=False)
if args.rank == 0 and model_ema is not None:
model_ema.copy_params_from_model_to_ema()
global_step = 0
torch.cuda.empty_cache()
# ******************************
# create optimizer
# ******************************
if args.finetune_all:
params = list(model.control_model.parameters())
params += list(model.model.diffusion_model.parameters())
try:
params += list(model.image_control_model.parameters())
except:
pass
print("Finetune All.")
elif args.finetune_imagecond_unet:
params = list(model.model.diffusion_model.parameters())
try:
params += list(model.image_control_model.parameters())
except:
pass
for p in model.control_model.parameters():
p.requires_grad_(False)
print("Finetune Unet and image controlnet, freeze pose controlnet")
elif args.finetune_attn:
for b in list(model.model.diffusion_model.input_blocks) + [model.model.diffusion_model.middle_block] + list(model.model.diffusion_model.output_blocks) + list(model.model.diffusion_model.out):
for p in b.parameters():
p.requires_grad_(False)
self_attn_parameters = []
for name, param in model.model.diffusion_model.named_parameters():
if "attn1" in name:
param.requires_grad_(True)
self_attn_parameters.append(param)
if "cldm_v15_reference_only_pose.yaml" in args.model_config:
params = list(model.appearance_control_model.parameters())
params += list(model.pose_control_model.parameters())
print("Train controlnet and attention layers in UNet")
else:
params = list(model.control_model.parameters())
params+= self_attn_parameters
if not args.sd_locked:
params += list(model.model.diffusion_model.output_blocks.parameters())
params += list(model.model.diffusion_model.out.parameters())
elif args.finetune_control:
for b in list(model.model.diffusion_model.input_blocks) + [model.model.diffusion_model.middle_block] + list(model.model.diffusion_model.output_blocks) + list(model.model.diffusion_model.out):
for p in b.parameters():
p.requires_grad_(False)
if ("cldm_v15_reference_only_pose.yaml" in args.model_config) or ("cldm_v15_reference_only_pose_mask.yaml") in args.model_config:
params = list(model.appearance_control_model.parameters())
if not args.fix_hint:
params += list(model.pose_control_model.parameters())
print("Train controlnet layers in UNet")
else:
# TODO
for b in list(model.pose_control_model.input_hint_block):
for p in b.parameters():
p.requires_grad_(False)
# params += list(model.pose_control_model.parameters())
params += list(model.pose_control_model.input_blocks.parameters())
params += list(model.pose_control_model.middle_block.parameters())
params += list(model.pose_control_model.middle_block_out.parameters())
params += list(model.pose_control_model.zero_convs.parameters())
print("Train controlnet layers in UNet, fix hint block in pose controlnet.")
else:
params = list(model.control_model.parameters())
if not args.sd_locked:
params += list(model.model.diffusion_model.output_blocks.parameters())
params += list(model.model.diffusion_model.out.parameters())
elif args.finetune_pose_only:
for b in list(model.model.diffusion_model.input_blocks) + [model.model.diffusion_model.middle_block] + list(model.model.diffusion_model.output_blocks) + list(model.model.diffusion_model.out) + list(model.appearance_control_model.input_blocks) + [model.appearance_control_model.middle_block] + list(model.appearance_control_model.output_blocks) + list(model.appearance_control_model.input_hint_block):
for p in b.parameters():
p.requires_grad_(False)
if ("cldm_v15_reference_only_pose.yaml" in args.model_config) or ("cldm_v15_reference_only_pose_mask.yaml") in args.model_config:
params = list(model.pose_control_model.parameters())
print("Train pose controlnet layers in UNet, freeze appearance")
else:
params = list(model.control_model.parameters())
if not args.sd_locked:
params += list(model.model.diffusion_model.output_blocks.parameters())
params += list(model.model.diffusion_model.out.parameters())
elif args.finetune_reference_only:
for b in list(model.model.diffusion_model.input_blocks) + [model.model.diffusion_model.middle_block] + list(model.model.diffusion_model.output_blocks) + list(model.model.diffusion_model.out) + list(model.pose_control_model.input_blocks) + [model.pose_control_model.middle_block] + list(model.pose_control_model.input_hint_block) + list(model.pose_control_model.middle_block_out) + list(model.pose_control_model.zero_convs):
for p in b.parameters():
p.requires_grad_(False)
if ("cldm_v15_reference_only_pose.yaml" in args.model_config) or ("cldm_v15_reference_only_pose_mask.yaml") in args.model_config:
params = list(model.appearance_control_model.parameters())
print("Train appearance controlnet layers in UNet, freeze pose controlnet")
else:
params = list(model.control_model.parameters())
if not args.sd_locked:
params += list(model.model.diffusion_model.output_blocks.parameters())
params += list(model.model.diffusion_model.out.parameters())
elif args.finetune_mm:
if args.image_finetune: # As described in Animatediff, train motion module then finetune image Unet layers.
for b in list(model.model.diffusion_model.input_blocks) + [model.model.diffusion_model.middle_block] + list(model.model.diffusion_model.output_blocks) + list(model.model.diffusion_model.out):
for p in b.parameters():
p.requires_grad_(False)
print("Freeze SD UNet")
if ("cldm_v15_reference_only_temporal_pose.yaml") in args.model_config:
for b in list(model.appearance_control_model.input_blocks) + [model.appearance_control_model.middle_block] + list(model.appearance_control_model.output_blocks):
for p in b.parameters():
p.requires_grad_(True)
print("Train appearance controlnet")
for b in list(model.pose_control_model.input_blocks) + [model.pose_control_model.middle_block] + list(model.pose_control_model.input_hint_block) + list(model.pose_control_model.middle_block_out) + list(model.pose_control_model.zero_convs):
for p in b.parameters():
p.requires_grad_(True)
print("Train pose controlnet")
for b in list(model.model.diffusion_model.input_blocks_motion_module) + list(model.model.diffusion_model.output_blocks_motion_module):
for p in b.parameters():
p.requires_grad_(False)
print("Freeze Motion Module")
params = list(model.pose_control_model.parameters())
params += list(model.appearance_control_model.input_blocks.parameters())
params += list(model.appearance_control_model.middle_block.parameters())
params += list(model.appearance_control_model.output_blocks.parameters())
elif args.image_finetune_unet:
for b in list(model.model.diffusion_model.input_blocks) + [model.model.diffusion_model.middle_block] + list(model.model.diffusion_model.output_blocks) + list(model.model.diffusion_model.out):
for p in b.parameters():
p.requires_grad_(True)
print("Train SD UNet")
if ("cldm_v15_reference_only_temporal_pose.yaml") in args.model_config:
for b in list(model.appearance_control_model.input_blocks) + [model.appearance_control_model.middle_block] + list(model.appearance_control_model.output_blocks):
for p in b.parameters():
p.requires_grad_(False)
print("Freeze appearance controlnet")
for b in list(model.pose_control_model.input_blocks) + [model.pose_control_model.middle_block] + list(model.pose_control_model.input_hint_block) + list(model.pose_control_model.middle_block_out) + list(model.pose_control_model.zero_convs):
for p in b.parameters():
p.requires_grad_(False)
print("Freeze pose controlnet")
for b in list(model.model.diffusion_model.input_blocks_motion_module) + list(model.model.diffusion_model.output_blocks_motion_module):
for p in b.parameters():
p.requires_grad_(False)
print("Freeze Motion Module")
params = list(model.model.diffusion_model.parameters())
elif args.unet_pose_app:
for b in list(model.model.diffusion_model.input_blocks) + [model.model.diffusion_model.middle_block] + list(model.model.diffusion_model.output_blocks) + list(model.model.diffusion_model.out):
for p in b.parameters():
p.requires_grad_(True)
print("Train SD UNet")
if ("cldm_v15_reference_only_temporal_pose.yaml") in args.model_config:
for b in list(model.appearance_control_model.input_blocks) + [model.appearance_control_model.middle_block] + list(model.appearance_control_model.output_blocks):
for p in b.parameters():
p.requires_grad_(True)
print("Train appearance controlnet")
for b in list(model.pose_control_model.input_blocks) + [model.pose_control_model.middle_block] + list(model.pose_control_model.input_hint_block) + list(model.pose_control_model.middle_block_out) + list(model.pose_control_model.zero_convs):
for p in b.parameters():
p.requires_grad_(True)
print("Train pose controlnet")
for b in list(model.model.diffusion_model.input_blocks_motion_module) + list(model.model.diffusion_model.output_blocks_motion_module):
for p in b.parameters():
p.requires_grad_(False)
print("Freeze Motion Module")
params = list(model.model.diffusion_model.parameters())
params += list(model.pose_control_model.parameters())
params += list(model.appearance_control_model.input_blocks.parameters())
params += list(model.appearance_control_model.middle_block.parameters())
params += list(model.appearance_control_model.output_blocks.parameters())
else:
for b in list(model.model.diffusion_model.input_blocks) + [model.model.diffusion_model.middle_block] + list(model.model.diffusion_model.output_blocks) + list(model.model.diffusion_model.out):
for p in b.parameters():
p.requires_grad_(False)
print("Freeze SD UNet")
if ("cldm_v15_reference_only_temporal.yaml" in args.model_config):
for b in list(model.control_model.input_blocks) + [model.control_model.middle_block] + list(model.control_model.output_blocks) + list(model.control_model.input_hint_block):
for p in b.parameters():
p.requires_grad_(False)
print("Freeze appearance controlnet")
params = list(model.model.diffusion_model.input_blocks_motion_module.parameters()) + list(model.model.diffusion_model.output_blocks_motion_module.parameters())
print("Train motion module in UNet")
if ("cldm_v15_reference_only_temporal_pose.yaml") in args.model_config:
if not args.finetune_reference:
for b in list(model.appearance_control_model.input_blocks) + [model.appearance_control_model.middle_block] + list(model.appearance_control_model.output_blocks) + list(model.appearance_control_model.input_hint_block):
for p in b.parameters():
p.requires_grad_(False)
print("Freeze appearance controlnet")
else:
for b in list(model.appearance_control_model.input_blocks) + [model.appearance_control_model.middle_block] + list(model.appearance_control_model.output_blocks):
for p in b.parameters():
p.requires_grad_(True)
print("Train appearance controlnet")
if args.finetune_mm_only:
for b in list(model.pose_control_model.input_blocks) + [model.pose_control_model.middle_block] + list(model.pose_control_model.input_hint_block) + list(model.pose_control_model.middle_block_out) + list(model.pose_control_model.zero_convs):
for p in b.parameters():
p.requires_grad_(False)
print("Freeze pose controlnet")
params = list(model.model.diffusion_model.input_blocks_motion_module.parameters()) + list(model.model.diffusion_model.output_blocks_motion_module.parameters())
print("Train Motion Module")
if not args.finetune_mm_only:
params += list(model.pose_control_model.parameters())
print("Train Pose controlnet")
if args.finetune_reference:
params += list(model.appearance_control_model.input_blocks.parameters())
params += list(model.appearance_control_model.middle_block.parameters())
params += list(model.appearance_control_model.output_blocks.parameters())
print("Train reference controlnet in UNet")
if not args.sd_locked:
params += list(model.model.diffusion_model.output_blocks.parameters())
params += list(model.model.diffusion_model.out.parameters())
else:
for b in list(model.model.diffusion_model.input_blocks) + [model.model.diffusion_model.middle_block] + list(model.model.diffusion_model.output_blocks) + list(model.model.diffusion_model.out):
for p in b.parameters():
p.requires_grad_(False)
params = list(model.control_model.parameters())
try:
params += list(model.image_control_model.parameters())
except:
pass
if not args.sd_locked:
params += list(model.model.diffusion_model.output_blocks.parameters())
params += list(model.model.diffusion_model.out.parameters())
print("Train controlnet")
optimizer = ZeroRedundancyOptimizer(
params = params,
lr=args.lr,
optimizer_class=torch.optim.AdamW,
weight_decay=args.weight_decay,
)
scaler = torch.cuda.amp.GradScaler(enabled=(args.use_fp16 and FP16_DTYPE==torch.float16))
# load optimizer state dict
if optimizer_state_dict is not None and args.load_optimizer_state:
try:
optimizer.load_state_dict(optimizer_state_dict["state_dict"])
if 'scaler_state_dict' in optimizer_state_dict:
scaler.load_state_dict(optimizer_state_dict["scaler_state_dict"])
print("load optimizer done")
except:
print("WARNING: load optimizer failed")
del optimizer_state_dict
# create lr scheduler
lr_scheduler = LambdaLinearScheduler(warm_up_steps=[ args.lr_anneal_steps ], cycle_lengths=[ 10000000000000 ], # incredibly large number to prevent corner cases
f_start=[ 1.e-6 ], f_max=[ 1. ], f_min=[ 1. ])
lr_scheduler = torch.optim.lr_scheduler.LambdaLR(optimizer, lr_lambda=lr_scheduler.schedule)
lr_scheduler.step(global_step)
# ******************************
# create DDP model
# ******************************
if args.compile and TORCH_VERSION == "2":
model = torch.compile(model)