-
Notifications
You must be signed in to change notification settings - Fork 32
/
Copy pathtrain.py
1085 lines (923 loc) · 53.3 KB
/
train.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 pi-GAN. Supports distributed training."""
import copy
import json
import os
import random
import time
from datetime import datetime
import joblib
import numpy as np
import torch
import torch.distributed as dist
from torch.nn.parallel import DistributedDataParallel as DDP
from torch_ema import ExponentialMovingAverage
from torch_fidelity import calculate_metrics
from torchvision.utils import save_image
from tqdm import tqdm
import gmpi.curriculums as curriculums
import gmpi.datasets as datasets
import gmpi.fid_evaluation as fid_evaluation
from gmpi.core.light_renderer import LightRenderer
from gmpi.core.mpi_renderer import MPIRenderer
from gmpi.train_helpers import STYLEGAN2_CFG_SPECS, cleanup, find_worst_view_per_z, modify_curriculums, setup, z_sampler
from gmpi.utils import TensorboardWriter, convert_cfg_to_dict, logger
def train(rank, world_size, config, master_port, run_dataset):
xyz_coords_only_z = False
# https://discuss.pytorch.org/t/extra-10gb-memory-on-gpu-0-in-ddp-tutorial/118113/2
torch.cuda.set_device(rank)
torch.cuda.empty_cache()
# reproducibility set up
random.seed(config.SEED)
np.random.seed(config.SEED)
torch.manual_seed(config.SEED)
torch.cuda.manual_seed(config.SEED)
torch.cuda.manual_seed_all(config.SEED)
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
torch_g = torch.Generator()
torch_g.manual_seed(config.SEED)
modify_curriculums(config, run_dataset=run_dataset)
# save config for reproducibility
torch.save(
{"config": convert_cfg_to_dict(config)},
os.path.join(config.CHECKPOINT_FOLDER, "config.pth"),
)
if rank == 0:
logger.add_filehandler(config.LOG_FILE)
opt = config.GMPI.TRAIN
opt.defrost()
opt.port = master_port
opt.freeze()
setup(rank, world_size, opt.port)
device = torch.device(rank)
curriculum = getattr(curriculums, opt.curriculum)
metadata = curriculums.extract_metadata(curriculum, 0)
if rank == 0:
logger.info(f"config: {config}\n\n")
logger.info(f"curriculum: {curriculum}\n")
writer = TensorboardWriter(config.TENSORBOARD_DIR, flush_secs=10)
fixed_z = z_sampler((25, metadata["latent_dim"]), device="cpu", dist=metadata["z_dist"])
if opt.load_dir != "":
raise NotImplementedError
else:
n_g_out_channels = 4
n_g_out_planes = config.GMPI.MPI.n_gen_planes
# fmt: off
if config.GMPI.MODEL.STYLEGAN2.torgba_cond_on_pos_enc != "none":
if config.GMPI.MODEL.STYLEGAN2.torgba_cond_on_pos_enc == "depth2alpha":
print("\nGenerator comes from depth2alpha\n")
from gmpi.models.networks.networks_vanilla_depth2alpha import Generator as StyleGAN2Generator
elif config.GMPI.MODEL.STYLEGAN2.torgba_cond_on_pos_enc == "normalize_add_z":
if config.GMPI.MODEL.STYLEGAN2.torgba_cond_on_pos_enc_embed_func in ["learnable_param"]:
print("\nGenerator comes from learnable_param\n")
from gmpi.models.networks.networks_pos_enc_learnable_param import Generator as StyleGAN2Generator
elif config.GMPI.MODEL.STYLEGAN2.torgba_cond_on_pos_enc_embed_func in ["modulated_lrelu"]:
print("\nGenerator comes from cond_on_depth\n")
from gmpi.models.networks.networks_cond_on_pos_enc import Generator as StyleGAN2Generator
else:
raise NotImplementedError
else:
raise NotImplementedError
else:
print("\nGenerator comes from vanilla\n")
from gmpi.models.networks.networks_vanilla import Generator as StyleGAN2Generator
# fmt: on
synthesis_kwargs = convert_cfg_to_dict(config.GMPI.MODEL.STYLEGAN2.synthesis_kwargs)
synthesis_kwargs_D = convert_cfg_to_dict(config.GMPI.MODEL.STYLEGAN2.synthesis_kwargs)
# NOTE: ref: https://github.com/NVlabs/stylegan2-ada-pytorch/blob/6f160b3d22b8b178ebe533a50d4d5e63aedba21d/train.py#L178
synthesis_kwargs["channel_base"] = int(
STYLEGAN2_CFG_SPECS[str(config.GMPI.MODEL.STYLEGAN2.max_out_dim)]["fmaps"]
* synthesis_kwargs["channel_base"]
)
synthesis_kwargs_D["channel_base"] = int(
STYLEGAN2_CFG_SPECS[str(config.GMPI.MODEL.STYLEGAN2.max_out_dim_D)]["fmaps"]
* synthesis_kwargs_D["channel_base"]
)
# For clamping, ref:
# - Sec. D.1 Mixed-precision training of https://arxiv.org/abs/2006.06676
# - https://github.com/NVlabs/stylegan2-ada-pytorch/blob/6f160b3d22b8b178ebe533a50d4d5e63aedba21d/train.py#L333
if config.GMPI.MODEL.STYLEGAN2.max_out_dim <= 128:
synthesis_kwargs["num_fp16_res"] = 0
synthesis_kwargs["conv_clamp"] = None
synthesis_kwargs_D["num_fp16_res"] = 0
synthesis_kwargs_D["conv_clamp"] = None
else:
synthesis_kwargs["conv_clamp"] = 256
# Ref: https://github.com/NVlabs/stylegan2-ada-pytorch/blob/d4b2afe9c27e3c305b721bc886d2cb5229458eba/train.py#L181
synthesis_kwargs["num_fp16_res"] = 4
synthesis_kwargs_D["conv_clamp"] = 256
synthesis_kwargs_D["num_fp16_res"] = 4
if config.GMPI.TRAIN.normalized_xyz_range == "01":
depth2alpha_z_range = 1.0
elif config.GMPI.TRAIN.normalized_xyz_range == "-11":
depth2alpha_z_range = 2.0
else:
raise ValueError
generator = StyleGAN2Generator(
z_dim=metadata["latent_dim"], # config.GMPI.MODEL.STYLEGAN2.z_dim,
c_dim=metadata["generator_label_dim"], # config.GMPI.MODEL.STYLEGAN2.label_dim,
w_dim=metadata["stylegan2_w_dim"], # config.GMPI.MODEL.STYLEGAN2.w_dim,
img_resolution=config.GMPI.MODEL.STYLEGAN2.max_out_dim,
n_planes=n_g_out_planes,
plane_channels=n_g_out_channels,
pos_enc_multires=config.GMPI.MODEL.STYLEGAN2.pos_enc_multires,
torgba_cond_on_pos_enc=config.GMPI.MODEL.STYLEGAN2.torgba_cond_on_pos_enc,
torgba_cond_on_pos_enc_embed_func=config.GMPI.MODEL.STYLEGAN2.torgba_cond_on_pos_enc_embed_func,
torgba_sep_background=config.GMPI.MODEL.STYLEGAN2.torgba_sep_background,
build_background_from_rgb=config.GMPI.MODEL.STYLEGAN2.build_background_from_rgb,
build_background_from_rgb_ratio=config.GMPI.MODEL.STYLEGAN2.build_background_from_rgb_ratio,
cond_on_pos_enc_only_alpha=config.GMPI.MODEL.STYLEGAN2.cond_on_pos_enc_only_alpha,
gen_alpha_largest_res=config.GMPI.MODEL.STYLEGAN2.gen_alpha_largest_res,
background_alpha_full=config.GMPI.MODEL.STYLEGAN2.background_alpha_full,
G_final_img_act=config.GMPI.MODEL.STYLEGAN2.G_final_img_act,
depth2alpha_z_range=depth2alpha_z_range,
depth2alpha_n_z_bins=config.GMPI.MPI.depth2alpha_n_z_bins,
mapping_kwargs=convert_cfg_to_dict(config.GMPI.MODEL.STYLEGAN2.mapping_kwargs),
synthesis_kwargs=synthesis_kwargs,
).to(device)
if config.GMPI.TRAIN.D_from_stylegan2:
assert not config.GMPI.TRAIN.D_pred_pos
from gmpi.models.networks.networks_cond_on_pos_enc import Discriminator as DiscriminatorStyleGAN2
if config.GMPI.TRAIN.D_cond_on_pose:
stylegan2_cdim = config.GMPI.TRAIN.D_cond_pose_dim
else:
stylegan2_cdim = 0
img_channels = 3
discriminator = DiscriminatorStyleGAN2(
stylegan2_cdim,
config.GMPI.MODEL.STYLEGAN2.max_out_dim_D,
img_channels,
channel_base=synthesis_kwargs_D["channel_base"],
channel_max=config.GMPI.MODEL.STYLEGAN2.synthesis_kwargs.channel_max,
num_fp16_res=synthesis_kwargs_D["num_fp16_res"],
conv_clamp=synthesis_kwargs_D["conv_clamp"],
cmap_dim=config.GMPI.MODEL.STYLEGAN2.discriminator.cmap_dim,
D_stylegan2_ori_mapping=config.GMPI.MODEL.STYLEGAN2.discriminator.use_ori_mapping,
use_mbstd_in_D=config.GMPI.MODEL.STYLEGAN2.discriminator.use_mbstd_in_D,
epilogue_kwargs={
"mbstd_group_size": config.GMPI.MODEL.STYLEGAN2.discriminator.mbstd_group_size,
# "mbstd_num_channels": config.GMPI.MODEL.STYLEGAN2.discriminator.mbstd_num_channels,
},
).to(device)
else:
raise NotImplementedError
ema = ExponentialMovingAverage(generator.parameters(), decay=0.999)
ema2 = ExponentialMovingAverage(generator.parameters(), decay=0.9999)
if config.GMPI.MODEL.use_pretrained_ckpt:
import gmpi.models.legacy as stylegan2_legacy
import gmpi.models.torch_utils.misc as stylegan2_misc
from gmpi.models.dnnlib.util import open_url
print(f'[rank {rank}] Resuming from "{config.GMPI.MODEL.pretrained}"')
# NOTE: it requires torch_utils folder in PYTHONPATH
with open_url(config.GMPI.MODEL.pretrained) as f:
resume_data = stylegan2_legacy.load_network_pkl(f)
print(f'[rank {rank}] Resuming D from "{config.GMPI.MODEL.pretrained_D}"')
with open_url(config.GMPI.MODEL.pretrained_D) as f:
resume_data_D = stylegan2_legacy.load_network_pkl(f)
generator = generator.train().requires_grad_(False)
discriminator = discriminator.train().requires_grad_(False)
ema_placeholder = copy.deepcopy(generator)
# NOTE: we resume G from G_ema
# for name, module in [('G_ema', generator), ('G_ema', ema_placeholder)]:
for name, module in [("G_ema", generator), ("G_ema", ema_placeholder)]:
logger.info(f"\n\n[rank {rank}] Resume {name}\n")
stylegan2_misc.copy_params_and_buffers(resume_data[name], module, require_all=False)
if config.GMPI.TRAIN.D_from_stylegan2:
logger.info(f"\n\n[rank {rank}] Resume D\n")
stylegan2_misc.copy_params_and_buffers(resume_data_D["D"], discriminator, require_all=False)
generator = generator.train().requires_grad_(True)
discriminator = discriminator.train().requires_grad_(True)
ema = ExponentialMovingAverage(ema_placeholder.parameters(), decay=0.999)
ema2 = ExponentialMovingAverage(ema_placeholder.parameters(), decay=0.9999)
generator_ddp = DDP(
generator,
device_ids=[rank],
find_unused_parameters=True,
broadcast_buffers=False,
)
discriminator_ddp = DDP(
discriminator,
device_ids=[rank],
find_unused_parameters=True,
broadcast_buffers=False,
)
generator = generator_ddp.module
discriminator = discriminator_ddp.module
# NOTE: setup tuning strategy here
if config.GMPI.TRAIN.only_tune:
raise NotImplementedError
else:
assert config.GMPI.TRAIN.D_train
enable_mapping_grad = True
enable_syn_feat_net_grad = True
assert not metadata.get("unique_lr", False)
optimizer_G = torch.optim.Adam(
generator_ddp.parameters(),
lr=metadata["gen_lr"],
betas=metadata["betas"],
weight_decay=metadata["weight_decay"],
)
optimizer_D = torch.optim.Adam(
discriminator_ddp.parameters(),
lr=metadata["disc_lr"],
betas=metadata["betas"],
weight_decay=metadata["weight_decay"],
)
generator_losses = []
discriminator_losses = []
generator.set_device(device)
# set up MPI renderer
mpi_renderer = MPIRenderer(
n_mpi_planes=config.GMPI.MPI.n_gen_planes,
plane_min_d=metadata["ray_start"],
plane_max_d=metadata["ray_end"],
plan_spatial_enlarge_factor=config.GMPI.MPI.CAM_SETUP.spatial_enlarge_factor,
plane_distances_sample_method=config.GMPI.MPI.distance_sample_method,
cam_fov=metadata["fov"],
sphere_center_z=config.GMPI.MPI.CAM_SETUP.cam_sphere_center_z,
sphere_r=config.GMPI.MPI.CAM_SETUP.cam_sphere_r,
horizontal_mean=metadata["h_mean"],
horizontal_std=metadata["h_stddev"],
vertical_mean=metadata["v_mean"],
vertical_std=metadata["v_stddev"],
cam_pose_n_truncated_stds=config.GMPI.MPI.CAM_SETUP.cam_pose_n_truncated_stds,
cam_sample_method=config.GMPI.MPI.CAM_SETUP.cam_pose_sample_method,
mpi_align_corners=config.GMPI.MPI.align_corners,
use_xyz_ztype=config.GMPI.TRAIN.use_xyz_ztype,
use_normalized_xyz=config.GMPI.TRAIN.use_normalized_xyz,
normalized_xyz_range=config.GMPI.TRAIN.normalized_xyz_range,
use_confined_volume=config.GMPI.MPI.use_confined_volume,
device=device,
)
light_renderer = LightRenderer(
sphere_center_z=config.GMPI.MPI.CAM_SETUP.cam_sphere_center_z,
sphere_r=config.GMPI.MPI.CAM_SETUP.cam_sphere_r,
ka_max=config.GMPI.TRAIN.aug_with_lighting_max_ka,
kd_max=config.GMPI.TRAIN.aug_with_lighting_max_kd,
n_grow_iters=config.GMPI.TRAIN.aug_with_lighting_grow_n_iters,
)
mpi_return_single_res_xyz = False
# ----------
# Training
# ----------
if rank == 0:
with open(os.path.join(opt.output_dir, "options.txt"), "w") as f:
f.write(str(opt))
f.write("\n\n")
f.write(str(generator))
f.write("\n\n")
f.write(str(discriminator))
f.write("\n\n")
f.write(str(curriculum))
# log information about the model
import gmpi.models.torch_utils.misc as stylegan2_misc
z = torch.empty([1, generator.z_dim], device=device)
c = torch.empty([1, generator.c_dim], device=device)
mpi_tex_pix_xyz, mpi_tex_pix_normalized_xyz = mpi_renderer.get_xyz(
metadata["tex_size"], metadata["tex_size"], ret_single_res=mpi_return_single_res_xyz
)
if config.GMPI.TRAIN.use_normalized_xyz:
mpi_xyz_input = mpi_tex_pix_normalized_xyz
else:
mpi_xyz_input = mpi_tex_pix_xyz
with torch.no_grad():
_ = stylegan2_misc.print_module_summary(
generator, [z, c, mpi_xyz_input, xyz_coords_only_z, config.GMPI.MPI.n_gen_planes]
)
torch.cuda.empty_cache()
n_g_params = sum(p.numel() for p in generator.parameters())
logger.info(f"#params of G: {n_g_params}\n")
if config.GMPI.TRAIN.D_from_stylegan2:
c = torch.empty([1, discriminator.c_dim], device=device)
img_placeholder = torch.empty([1, 3, metadata["img_size"], metadata["img_size"]], device=device)
alpha_placeholder = 1.0
stylegan2_misc.print_module_summary(discriminator, [img_placeholder, alpha_placeholder, c])
else:
logger.info(f"discriminator: {discriminator}")
n_d_params = sum(p.numel() for p in discriminator.parameters())
logger.info(f"#params of D: {n_d_params}\n")
torch.manual_seed(rank)
dataloader = None
total_progress_bar = tqdm(total=opt.total_iters, desc="Total progress", dynamic_ncols=True)
total_progress_bar.update(discriminator.step)
for epoch_i in range(opt.n_epochs):
if discriminator.step > opt.total_iters:
break
metadata = curriculums.extract_metadata(curriculum, discriminator.step)
# NOTE: for condition on pose
metadata["sphere_center"] = config.GMPI.MPI.CAM_SETUP.cam_sphere_center_z
metadata["sphere_r"] = config.GMPI.MPI.CAM_SETUP.cam_sphere_r
metadata["flat_pose_dim"] = config.GMPI.TRAIN.D_cond_pose_dim
# we reset camera if needed
if metadata["img_size"] != mpi_renderer.render_h or metadata["img_size"] != mpi_renderer.render_w:
mpi_renderer.set_cam(metadata["fov"], metadata["img_size"], metadata["img_size"])
# Set learning rates
for param_group in optimizer_G.param_groups:
if param_group.get("name", None) == "mapping_network":
param_group["lr"] = metadata["gen_lr"] * 5e-2
else:
param_group["lr"] = metadata["gen_lr"]
param_group["betas"] = metadata["betas"]
param_group["weight_decay"] = metadata["weight_decay"]
for param_group in optimizer_D.param_groups:
param_group["lr"] = metadata["disc_lr"]
param_group["betas"] = metadata["betas"]
param_group["weight_decay"] = metadata["weight_decay"]
if not dataloader or dataloader.dataset.img_size != metadata["img_size"]:
del dataloader
time.sleep(10)
dataloader, data_sampler, CHANNELS = datasets.get_dataset_distributed(
metadata["dataset"],
world_size,
rank,
num_workres=config.GMPI.TRAIN.n_dataloader_workers,
torch_g=torch_g,
**metadata,
)
step_next_upsample = curriculums.next_upsample_step(curriculum, discriminator.step)
step_last_upsample = curriculums.last_upsample_step(curriculum, discriminator.step)
logger.info(f"New epoch {epoch_i}.\n\n")
# NOTE: this is requred to make distributed sampler shuffle dataset.
data_sampler.set_epoch(epoch_i)
for batch_i, batch_data in enumerate(dataloader):
if discriminator.step > opt.total_iters:
break
if rank == 0:
total_progress_bar.update(1)
# pred_yaws_real/pred_pitches_real: [B, 1]
imgs, flat_w2c_mats_real, _, pred_yaws_real, pred_pitches_real = batch_data
# NOTE: we only condition on rotation
flat_w2c_mats_real = flat_w2c_mats_real.to(device)
pred_pose_angles_real = torch.cat((pred_pitches_real, pred_yaws_real), dim=1).to(device)
cur_h_stddev = metadata["h_stddev"]
cur_v_stddev = metadata["v_stddev"]
if discriminator.step % opt.model_save_interval == 0 and rank == 0:
# fmt: off
now = datetime.now()
now = now.strftime("%d--%H:%M--")
torch.save(ema.state_dict(), os.path.join(config.CHECKPOINT_FOLDER, f'ema_{discriminator.step}.pth'))
torch.save(ema2.state_dict(), os.path.join(config.CHECKPOINT_FOLDER, f'ema2_{discriminator.step}.pth'))
torch.save(generator_ddp.module.state_dict(), os.path.join(config.CHECKPOINT_FOLDER, f'generator_{discriminator.step}.pth'))
torch.save(discriminator_ddp.module.state_dict(), os.path.join(config.CHECKPOINT_FOLDER, f'discriminator_{discriminator.step}.pth'))
torch.save(optimizer_G.state_dict(), os.path.join(config.CHECKPOINT_FOLDER, f'optimizer_G_{discriminator.step}.pth'))
torch.save(optimizer_D.state_dict(), os.path.join(config.CHECKPOINT_FOLDER, f'optimizer_D_{discriminator.step}.pth'))
# fmt: on
metadata = curriculums.extract_metadata(curriculum, discriminator.step)
# NOTE: for condition on pose
metadata["sphere_center"] = config.GMPI.MPI.CAM_SETUP.cam_sphere_center_z
metadata["sphere_r"] = config.GMPI.MPI.CAM_SETUP.cam_sphere_r
metadata["flat_pose_dim"] = config.GMPI.TRAIN.D_cond_pose_dim
# if dataloader.batch_size != metadata["batch_size"]:
if dataloader.dataset.img_size != metadata["img_size"]:
time.sleep(10)
break
generator_ddp.train()
discriminator_ddp.train()
if config.GMPI.TRAIN.D_from_stylegan2:
alpha = 1.0
else:
alpha = min(1, (discriminator.step - step_last_upsample) / (metadata["fade_steps"]))
# alpha = 1.0
real_imgs = imgs.to(device, non_blocking=True)
# TRAIN DISCRIMINATOR
optimizer_D.zero_grad()
with torch.set_grad_enabled(config.GMPI.TRAIN.D_train):
mpi_tex_pix_xyz, mpi_tex_pix_normalized_xyz = mpi_renderer.get_xyz(
metadata["tex_size"], metadata["tex_size"], ret_single_res=mpi_return_single_res_xyz
)
if config.GMPI.TRAIN.use_normalized_xyz:
stylegan2_mpi_xyz_input = mpi_tex_pix_normalized_xyz
else:
stylegan2_mpi_xyz_input = mpi_tex_pix_xyz
if batch_i == 0:
logger.info(f"\nmpi_tex_pix_xyz: {stylegan2_mpi_xyz_input[4][:, 0, 0, 2]} \n")
# Generate images for discriminator training
with torch.no_grad():
real_bs = real_imgs.shape[0]
if (config.GMPI.TRAIN.n_view_per_z_in_train > 1) and (
config.GMPI.TRAIN.G_select_worse_view == "none"
):
# NOTE: if we do not select worst view for G and we use multi-view setup,
# to make sure real and fake have same number of views,
# we need to reduce the number of generated MPIs.
fake_bs = real_bs // config.GMPI.TRAIN.n_view_per_z_in_train
assert (
real_bs % config.GMPI.TRAIN.n_view_per_z_in_train == 0
), f"{real_bs}, {config.GMPI.TRAIN.n_view_per_z_in_train}"
else:
fake_bs = real_bs
z = z_sampler(
(fake_bs, metadata["latent_dim"]),
device=device,
dist=metadata["z_dist"],
)
split_batch_size_list = [
z.shape[0] // metadata["batch_split"] for _ in range(metadata["batch_split"])
]
n_diff = z.shape[0] - np.sum(split_batch_size_list)
for i in range(n_diff):
split_batch_size_list[i] += 1
gen_imgs = []
gen_positions = []
flat_w2c_mats_gen = []
start_batch_idx = 0
for split in range(metadata["batch_split"]):
split_batch_size = split_batch_size_list[split]
subset_z = z[start_batch_idx : start_batch_idx + split_batch_size]
start_batch_idx = start_batch_idx + split_batch_size
batch_mpi_rgbas = generator_ddp(
z=subset_z,
c=None,
mpi_xyz_coords=stylegan2_mpi_xyz_input,
xyz_coords_only_z=xyz_coords_only_z,
n_planes=config.GMPI.MPI.n_gen_planes,
enable_mapping_grad=enable_mapping_grad,
enable_syn_feat_net_grad=enable_syn_feat_net_grad,
truncation_psi=config.GMPI.MODEL.STYLEGAN2.truncation_psi,
)
if batch_i == 0:
logger.info(f"\n[D] batch_mpi_rgbas: {batch_mpi_rgbas.shape}\n")
if (
config.GMPI.TRAIN.aug_with_lighting
and discriminator.step > config.GMPI.TRAIN.aug_with_lighting_start_iter
):
tmp_mpi_tex_pix_xyz = mpi_tex_pix_xyz[metadata["tex_size"]]
batch_mpi_rgbas = light_renderer.render(
batch_mpi_rgbas, mpi_renderer.static_mpi_plane_dhws, tmp_mpi_tex_pix_xyz
)
# [B, #planes, 4, H, W]
# print("\nbatch_mpi_rgbas: ", batch_mpi_rgbas.shape, "\n")
tmp_bs, tmp_n_planes, _, tmp_h, tmp_w = batch_mpi_rgbas.shape
tmp_n_view_per_z_in_train = config.GMPI.TRAIN.n_view_per_z_in_train
if (config.GMPI.TRAIN.n_view_per_z_in_train > 1) and (
config.GMPI.TRAIN.G_select_worse_view != "none"
):
# NOTE: we need to make sure D is exposed to same number of real/fake images
tmp_n_view_per_z_in_train = 1
batch_mpi_rgbas = batch_mpi_rgbas.unsqueeze(0).expand(
tmp_n_view_per_z_in_train, -1, -1, -1, -1, -1
)
batch_mpi_rgbas = batch_mpi_rgbas.reshape(
(tmp_n_view_per_z_in_train * tmp_bs, tmp_n_planes, 4, tmp_h, tmp_w)
)
g_imgs, _, g_c2w_mats, g_pos = mpi_renderer.render(
batch_mpi_rgbas,
metadata["img_size"],
metadata["img_size"],
horizontal_std=cur_h_stddev,
vertical_std=cur_v_stddev,
)
if config.GMPI.TRAIN.D_cond_pose_dim == 9:
g_w2c_mats = torch.inverse(g_c2w_mats[:, :3, :3])
elif config.GMPI.TRAIN.D_cond_pose_dim == 16:
g_w2c_mats = torch.inverse(g_c2w_mats)
else:
raise ValueError
flat_g_w2c_mats = g_w2c_mats.reshape((split_batch_size * tmp_n_view_per_z_in_train, -1))
flat_w2c_mats_gen.append(flat_g_w2c_mats)
gen_imgs.append(g_imgs)
gen_positions.append(g_pos)
gen_imgs = torch.cat(gen_imgs, axis=0)
gen_positions = torch.cat(gen_positions, axis=0)
flat_w2c_mats_gen = torch.cat(flat_w2c_mats_gen, dim=0)
real_imgs.requires_grad = True
if config.GMPI.TRAIN.D_cond_on_pose:
r_preds, z_preds, pose_preds = discriminator_ddp(real_imgs, alpha, flat_w2c_mats_real, **metadata)
else:
r_preds, z_preds, pose_preds = discriminator_ddp(real_imgs, alpha, None, **metadata)
if config.GMPI.TRAIN.D_train and metadata["r1_lambda"] > 0:
# Gradient penalty
grad_real = torch.autograd.grad(
outputs=r_preds.sum(),
inputs=real_imgs,
create_graph=True,
)
grad_real = [p for p in grad_real][0]
if config.GMPI.TRAIN.D_train and metadata["r1_lambda"] > 0:
grad_penalty = (grad_real.view(grad_real.size(0), -1).norm(2, dim=1) ** 2).mean()
grad_penalty = 0.5 * metadata["r1_lambda"] * grad_penalty
else:
grad_penalty = torch.zeros(1, device=device)
if batch_i == 0:
logger.info(f"[D] grad_penalty: {grad_penalty}\n")
if config.GMPI.TRAIN.D_cond_on_pose:
g_preds, g_pred_latent, g_pred_position = discriminator_ddp(
gen_imgs, alpha, flat_w2c_mats_gen, **metadata
)
else:
g_preds, g_pred_latent, g_pred_position = discriminator_ddp(gen_imgs, alpha, None, **metadata)
d_gan_loss_real = torch.nn.functional.softplus(-r_preds).mean()
d_gan_loss_fake = torch.nn.functional.softplus(g_preds).mean()
d_gan_loss = d_gan_loss_real + d_gan_loss_fake
d_loss = d_gan_loss + grad_penalty
discriminator_losses.append(d_loss.item())
optimizer_D.zero_grad()
if config.GMPI.TRAIN.D_train:
d_loss.backward()
torch.nn.utils.clip_grad_norm_(discriminator_ddp.parameters(), metadata["grad_clip"])
optimizer_D.step()
# TRAIN GENERATOR
optimizer_G.zero_grad()
split_batch_size_list = [fake_bs // metadata["batch_split"] for _ in range(metadata["batch_split"])]
n_diff = fake_bs - np.sum(split_batch_size_list)
for i in range(n_diff):
split_batch_size_list[i] += 1
assert fake_bs % metadata["batch_split"] == 0, f"{fake_bs}, {metadata['batch_split']}"
for tmp_g_iter in range(config.GMPI.TRAIN.G_iters):
z = z_sampler(
(fake_bs, metadata["latent_dim"]),
device=device,
dist=metadata["z_dist"],
)
if (config.GMPI.TRAIN.n_view_per_z_in_train > 1) and (config.GMPI.TRAIN.G_select_worse_view != "none"):
with torch.no_grad():
# NOTE: we use Z from worst-view
worst_views_yaws, worst_views_pitches, z = find_worst_view_per_z(
# device=device,
generator_ddp=generator_ddp,
discriminator_ddp=discriminator_ddp,
mpi_renderer=mpi_renderer,
z=z,
config=config,
metadata=metadata,
stylegan2_mpi_xyz_input=stylegan2_mpi_xyz_input,
enable_mapping_grad=enable_mapping_grad,
enable_syn_feat_net_grad=enable_syn_feat_net_grad,
cur_h_stddev=cur_h_stddev,
cur_v_stddev=cur_v_stddev,
alpha=alpha,
xyz_coords_only_z=xyz_coords_only_z,
n_planes=config.GMPI.MPI.n_gen_planes,
truncation_psi=config.GMPI.MODEL.STYLEGAN2.truncation_psi,
)
else:
worst_views_pitches, worst_views_yaws = None, None
start_batch_idx = 0
for split in range(metadata["batch_split"]):
split_batch_size = split_batch_size_list[split]
subset_z = z[start_batch_idx : (start_batch_idx + split_batch_size)]
if (config.GMPI.TRAIN.n_view_per_z_in_train > 1) and (
config.GMPI.TRAIN.G_select_worse_view != "none"
):
tmp_given_yaws = worst_views_yaws[start_batch_idx : (start_batch_idx + split_batch_size), :]
tmp_given_pitches = worst_views_pitches[
start_batch_idx : (start_batch_idx + split_batch_size), :
]
else:
tmp_given_yaws = None
tmp_given_pitches = None
start_batch_idx = start_batch_idx + split_batch_size
batch_mpi_rgbas = generator_ddp(
z=subset_z,
c=None,
mpi_xyz_coords=stylegan2_mpi_xyz_input,
xyz_coords_only_z=xyz_coords_only_z,
n_planes=config.GMPI.MPI.n_gen_planes,
enable_mapping_grad=enable_mapping_grad,
enable_syn_feat_net_grad=enable_syn_feat_net_grad,
truncation_psi=config.GMPI.MODEL.STYLEGAN2.truncation_psi,
)
if batch_i == 0:
logger.info(f"\n[G] batch_mpi_rgbas: {batch_mpi_rgbas.shape}\n")
if (
config.GMPI.TRAIN.aug_with_lighting
and discriminator.step > config.GMPI.TRAIN.aug_with_lighting_start_iter
):
tmp_mpi_tex_pix_xyz = mpi_tex_pix_xyz[metadata["tex_size"]]
batch_mpi_rgbas = light_renderer.render(
batch_mpi_rgbas, mpi_renderer.static_mpi_plane_dhws, tmp_mpi_tex_pix_xyz
)
tmp_use_edge_aware_loss = (
config.GMPI.TRAIN.use_edge_aware_loss
and discriminator.step >= config.GMPI.TRAIN.edge_aware_loss_start_iter
)
if config.GMPI.TRAIN.use_cano_reconstruct_loss or tmp_use_edge_aware_loss:
raise NotImplementedError
else:
edge_smooth_loss = torch.zeros(1, device=device)
cano_reconstruct_loss = torch.zeros(1, device=device)
if (config.GMPI.TRAIN.n_view_per_z_in_train > 1) and (
config.GMPI.TRAIN.G_select_worse_view != "none"
):
# NOTE: we have already chosen the worst view, no need to render multiple views.
tmp_n_view_per_z_in_train = 1
else:
tmp_n_view_per_z_in_train = config.GMPI.TRAIN.n_view_per_z_in_train
if tmp_n_view_per_z_in_train > 1:
# [B, #planes, 4, H, W]
tmp_bs, tmp_n_planes, _, tmp_h, tmp_w = batch_mpi_rgbas.shape
batch_mpi_rgbas = batch_mpi_rgbas.unsqueeze(1).expand(
-1, tmp_n_view_per_z_in_train, -1, -1, -1, -1
)
batch_mpi_rgbas = batch_mpi_rgbas.reshape(
(tmp_n_view_per_z_in_train * tmp_bs, tmp_n_planes, 4, tmp_h, tmp_w)
)
gen_imgs, _, gen_c2w_mats, gen_positions = mpi_renderer.render(
batch_mpi_rgbas,
metadata["img_size"],
metadata["img_size"],
horizontal_std=cur_h_stddev,
vertical_std=cur_v_stddev,
given_yaws=tmp_given_yaws,
given_pitches=tmp_given_pitches,
)
if config.GMPI.TRAIN.D_cond_pose_dim == 9:
gen_w2c_mats = torch.inverse(gen_c2w_mats[:, :3, :3])
elif config.GMPI.TRAIN.D_cond_pose_dim == 16:
gen_w2c_mats = torch.inverse(gen_c2w_mats)
else:
raise ValueError
flat_gen_w2c_mats = gen_w2c_mats.reshape((split_batch_size * tmp_n_view_per_z_in_train, -1))
if config.GMPI.TRAIN.D_cond_on_pose:
g_preds, g_pred_latent, g_pred_position = discriminator_ddp(
gen_imgs, alpha, flat_gen_w2c_mats, **metadata
)
else:
g_preds, g_pred_latent, g_pred_position = discriminator_ddp(gen_imgs, alpha, None, **metadata)
topk_num = int(g_preds.shape[0])
# NOTE: -1 * log sigmoid(f(x)) = log (1 + exp(-f(x))) = softplus(-f(x))
# Therefore, large g_preds/f(x) indicates small G_loss
g_preds = torch.topk(g_preds, topk_num, dim=0, sorted=False).values
g_gan_loss = torch.nn.functional.softplus(-g_preds).mean()
g_loss = g_gan_loss + edge_smooth_loss + cano_reconstruct_loss
generator_losses.append(g_loss.item())
# NOTE: to make sure loss are same as using a whole batch
g_loss = g_loss / (metadata["batch_split"] + 1e-8)
g_loss.backward()
torch.nn.utils.clip_grad_norm_(generator_ddp.parameters(), metadata.get("grad_clip", 0.3))
optimizer_G.step()
ema.update(generator_ddp.parameters())
ema2.update(generator_ddp.parameters())
if rank == 0:
if discriminator.step % 10 == 0:
tqdm.write(
f"[Experiment: {opt.output_dir}] "
# f"[Epoch: {discriminator.epoch}/{opt.n_epochs}] "
f"[D loss: {d_loss.item()}] [G loss: {g_loss.item()}] "
f"[Step: {discriminator.step}] "
f"[Img Size: {metadata['img_size']}] [Batch Size: {metadata['batch_size']}] "
f"[TopK: {topk_num}] "
)
# fmt: off
writer.add_scalar(f"loss_d", d_gan_loss.item(), global_step=int(discriminator.step))
writer.add_scalar(f"loss_d_real", d_gan_loss_real.item(), global_step=int(discriminator.step))
writer.add_scalar(f"loss_d_fake", d_gan_loss_fake.item(), global_step=int(discriminator.step))
writer.add_scalar(f"loss_g_fake", g_gan_loss.item(), global_step=int(discriminator.step))
writer.add_scalar(f"grad_penalty", grad_penalty.item(), global_step=int(discriminator.step))
writer.add_scalar(f"edge_smooth_loss", edge_smooth_loss.item(), global_step=int(discriminator.step))
writer.add_scalar(f"cano_reconstruct_loss", cano_reconstruct_loss.item(), global_step=int(discriminator.step))
writer.add_scalar(f"d_fade_in_alpha", alpha, global_step=int(discriminator.step))
writer.add_scalar(f"cur_h_stddev", cur_h_stddev, global_step=int(discriminator.step))
writer.add_scalar(f"cur_v_stddev", cur_v_stddev, global_step=int(discriminator.step))
if config.GMPI.TRAIN.aug_with_lighting:
writer.add_scalar(f"cur_ka", light_renderer.cur_ka, global_step=int(discriminator.step))
writer.add_scalar(f"cur_kd", light_renderer.cur_kd, global_step=int(discriminator.step))
# fmt: on
# fmt: off
if (rank == 0):
# if (discriminator.step > 0) and (discriminator.step % opt.sample_interval == 0):
if (discriminator.step % opt.sample_interval == 0):
mpi_tex_pix_xyz, mpi_tex_pix_normalized_xyz = mpi_renderer.get_xyz(metadata["tex_size"], metadata["tex_size"], ret_single_res=mpi_return_single_res_xyz)
if config.GMPI.TRAIN.use_normalized_xyz:
stylegan2_mpi_xyz_input = mpi_tex_pix_normalized_xyz
else:
stylegan2_mpi_xyz_input = mpi_tex_pix_xyz
minibatch = 1
n_minibatches = int(np.ceil(fixed_z.shape[0] / minibatch))
generator_ddp.eval()
with torch.no_grad():
copied_metadata = copy.deepcopy(metadata)
copied_metadata['h_stddev'] = copied_metadata['v_stddev'] = 0
copied_metadata['img_size'] = copied_metadata['eval_img_size']
gen_imgs = []
for i in range(n_minibatches):
mb_mpi_rgbas = generator_ddp.module.forward(
z=fixed_z[(i * minibatch):((i + 1) * minibatch)].to(device),
c=None,
mpi_xyz_coords=stylegan2_mpi_xyz_input,
xyz_coords_only_z=xyz_coords_only_z,
n_planes=config.GMPI.MPI.n_gen_planes,
truncation_psi=config.GMPI.MODEL.STYLEGAN2.truncation_psi,
)
mb_gen_imgs, _, _, _ = mpi_renderer.render(
mb_mpi_rgbas, copied_metadata['img_size'], copied_metadata['img_size'],
horizontal_mean=copied_metadata['h_mean'],
horizontal_std=copied_metadata['h_stddev'],
vertical_mean=copied_metadata['v_mean'],
vertical_std=copied_metadata['v_stddev'],
)
gen_imgs.append(mb_gen_imgs)
gen_imgs = torch.cat(gen_imgs, dim=0)
save_image(gen_imgs[:25], os.path.join(opt.output_dir, f"{discriminator.step}_fixed.png"), nrow=5, normalize=True)
with torch.no_grad():
if opt.curriculum == "FFHQ":
tmp_h_angle = 0.5
elif opt.curriculum == "AFHQCat":
tmp_h_angle = 0.3
elif opt.curriculum == "MetFaces":
tmp_h_angle = 0.5
else:
raise ValueError
copied_metadata = copy.deepcopy(metadata)
copied_metadata['h_stddev'] = copied_metadata['v_stddev'] = 0
copied_metadata['h_mean'] += tmp_h_angle
copied_metadata['img_size'] = copied_metadata['eval_img_size']
if config.GMPI.MPI.CAM_SETUP.cam_pose_n_truncated_stds * metadata["h_stddev"] > tmp_h_angle:
gen_imgs = []
for i in range(n_minibatches):
mb_mpi_rgbas = generator_ddp.module.forward(
z=fixed_z[(i * minibatch):((i + 1) * minibatch)].to(device),
c=None,
mpi_xyz_coords=stylegan2_mpi_xyz_input,
xyz_coords_only_z=xyz_coords_only_z,
n_planes=config.GMPI.MPI.n_gen_planes,
truncation_psi=config.GMPI.MODEL.STYLEGAN2.truncation_psi,
)
mb_gen_imgs, _, _, _ = mpi_renderer.render(
mb_mpi_rgbas, copied_metadata['img_size'], copied_metadata['img_size'],
horizontal_mean=copied_metadata['h_mean'],
horizontal_std=copied_metadata['h_stddev'],
vertical_mean=copied_metadata['v_mean'],
vertical_std=copied_metadata['v_stddev'],
)
gen_imgs.append(mb_gen_imgs)
gen_imgs = torch.cat(gen_imgs, dim=0)
else:
gen_imgs = torch.zeros((25, 3, copied_metadata['img_size'], copied_metadata['img_size']))
save_image(gen_imgs[:25], os.path.join(opt.output_dir, f"{discriminator.step}_tilted.png"), nrow=5, normalize=True)
ema.store(generator_ddp.parameters())
ema.copy_to(generator_ddp.parameters())
generator_ddp.eval()
with torch.no_grad():
copied_metadata = copy.deepcopy(metadata)
copied_metadata['h_stddev'] = copied_metadata['v_stddev'] = 0
copied_metadata['img_size'] = copied_metadata['eval_img_size']
gen_imgs = []
for i in range(n_minibatches):
mb_mpi_rgbas = generator_ddp.module.forward(
z=fixed_z[(i * minibatch):((i + 1) * minibatch)].to(device),
c=None,
mpi_xyz_coords=stylegan2_mpi_xyz_input,
xyz_coords_only_z=xyz_coords_only_z,
n_planes=config.GMPI.MPI.n_gen_planes,
truncation_psi=config.GMPI.MODEL.STYLEGAN2.truncation_psi,
)
mb_gen_imgs, _, _, _ = mpi_renderer.render(
mb_mpi_rgbas, copied_metadata['img_size'], copied_metadata['img_size'],
horizontal_mean=copied_metadata['h_mean'],
horizontal_std=copied_metadata['h_stddev'],
vertical_mean=copied_metadata['v_mean'],
vertical_std=copied_metadata['v_stddev'],
)
gen_imgs.append(mb_gen_imgs)
gen_imgs = torch.cat(gen_imgs, dim=0)
save_image(gen_imgs[:25], os.path.join(opt.output_dir, f"{discriminator.step}_fixed_ema.png"), nrow=5, normalize=True)
with torch.no_grad():
if opt.curriculum == "FFHQ":
tmp_h_angle = 0.5
elif opt.curriculum == "AFHQCat":
tmp_h_angle = 0.3
elif opt.curriculum == "MetFaces":
tmp_h_angle = 0.5
else:
raise ValueError
copied_metadata = copy.deepcopy(metadata)
copied_metadata['h_stddev'] = copied_metadata['v_stddev'] = 0
copied_metadata['h_mean'] += tmp_h_angle
copied_metadata['img_size'] = copied_metadata['eval_img_size']
if config.GMPI.MPI.CAM_SETUP.cam_pose_n_truncated_stds * metadata["h_stddev"] > tmp_h_angle:
gen_imgs = []
for i in range(n_minibatches):
mb_mpi_rgbas = generator_ddp.module.forward(
z=fixed_z[(i * minibatch):((i + 1) * minibatch)].to(device),
c=None,
mpi_xyz_coords=stylegan2_mpi_xyz_input,
xyz_coords_only_z=xyz_coords_only_z,
n_planes=config.GMPI.MPI.n_gen_planes,
truncation_psi=config.GMPI.MODEL.STYLEGAN2.truncation_psi,
)
mb_gen_imgs, _, _, _ = mpi_renderer.render(
mb_mpi_rgbas, copied_metadata['img_size'], copied_metadata['img_size'],
horizontal_mean=copied_metadata['h_mean'],
horizontal_std=copied_metadata['h_stddev'],
vertical_mean=copied_metadata['v_mean'],
vertical_std=copied_metadata['v_stddev'],
)
gen_imgs.append(mb_gen_imgs)
gen_imgs = torch.cat(gen_imgs, dim=0)
else:
gen_imgs = torch.zeros((25, 3, copied_metadata['img_size'], copied_metadata['img_size']))
save_image(gen_imgs[:25], os.path.join(opt.output_dir, f"{discriminator.step}_tilted_ema.png"), nrow=5, normalize=True)
with torch.no_grad():
copied_metadata = copy.deepcopy(metadata)
copied_metadata['img_size'] = copied_metadata['eval_img_size']
copied_metadata['h_stddev'] = copied_metadata['v_stddev'] = 0
copied_metadata['psi'] = 0.7
rnd_z = torch.randn_like(fixed_z).to(device)
gen_imgs = []
for i in range(n_minibatches):
# [B, #planes, 4, tex_h, tex_w]
mb_mpi_rgbas = generator_ddp.module.forward(
z=rnd_z[(i * minibatch):((i + 1) * minibatch)].to(device),
c=None,
mpi_xyz_coords=stylegan2_mpi_xyz_input,
xyz_coords_only_z=xyz_coords_only_z,
n_planes=config.GMPI.MPI.n_gen_planes,
truncation_psi=config.GMPI.MODEL.STYLEGAN2.truncation_psi,
)
mb_gen_imgs, _, _, _ = mpi_renderer.render(
mb_mpi_rgbas, copied_metadata['img_size'], copied_metadata['img_size'],
horizontal_mean=copied_metadata['h_mean'],
horizontal_std=copied_metadata['h_stddev'],
vertical_mean=copied_metadata['v_mean'],
vertical_std=copied_metadata['v_stddev'],
)
gen_imgs.append(mb_gen_imgs)
gen_imgs = torch.cat(gen_imgs, dim=0)
save_image(gen_imgs[:25], os.path.join(opt.output_dir, f"{discriminator.step}_random.png"), nrow=5, normalize=True)
if config.GMPI.TRAIN.aug_with_lighting and discriminator.step > config.GMPI.TRAIN.aug_with_lighting_start_iter:
tmp_mpi_tex_pix_xyz = mpi_tex_pix_xyz[metadata["tex_size"]]
mb_mpi_rgbas = light_renderer.render(mb_mpi_rgbas, mpi_renderer.static_mpi_plane_dhws, tmp_mpi_tex_pix_xyz)
tmp_i = int(np.random.choice(mb_mpi_rgbas.shape[0], size=1)[0])
mpi_rgb = mb_mpi_rgbas[tmp_i, :, :3, ...]
mpi_alpha = mb_mpi_rgbas[tmp_i, :, 3:, ...]
save_image(mpi_rgb, os.path.join(opt.output_dir, f"{discriminator.step}_random_mpi_rgb.png"), nrow=8, normalize=False)
save_image(mpi_alpha, os.path.join(opt.output_dir, f"{discriminator.step}_random_mpi_alpha.png"), nrow=8, normalize=True)
ema.restore(generator_ddp.parameters())
torch.cuda.empty_cache()
if discriminator.step % opt.sample_interval == 0:
torch.save(ema.state_dict(), os.path.join(config.CHECKPOINT_FOLDER, 'ema.pth'))
torch.save(ema2.state_dict(), os.path.join(config.CHECKPOINT_FOLDER, 'ema2.pth'))
torch.save(generator_ddp.module.state_dict(), os.path.join(config.CHECKPOINT_FOLDER, 'generator.pth'))