-
Notifications
You must be signed in to change notification settings - Fork 50
/
train.py
638 lines (591 loc) · 23.4 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
# Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
#
# This source code is licensed under the license found in the
# LICENSE file in the root directory of this source tree.
from PIL import Image, ImageFile
ImageFile.LOAD_TRUNCATED_IMAGES = True
from torch.utils.tensorboard import SummaryWriter
import argparse
import json
import os
import time
import glob
import cv2
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
from dataset import Dataset
from models import DeepAppearanceVAE, WarpFieldVAE
from torch.utils.data import DataLoader
from torch.utils.data.distributed import DistributedSampler
from utils import Renderer, gammaCorrect
def main(args, camera_config, test_segment):
local_rank = torch.distributed.get_rank()
torch.cuda.set_device(local_rank)
device = torch.device("cuda", local_rank)
dataset_train = Dataset(
args.data_dir,
args.krt_dir,
args.framelist_train,
args.tex_size,
camset=None if camera_config is None else camera_config["train"],
exclude_prefix=test_segment,
)
dataset_test = Dataset(
args.data_dir,
args.krt_dir,
args.framelist_test,
args.tex_size,
camset=None if camera_config is None else camera_config["test"],
valid_prefix=test_segment,
)
train_sampler = DistributedSampler(dataset_train)
test_sampler = DistributedSampler(dataset_test)
train_loader = DataLoader(
dataset_train,
args.train_batch_size,
sampler=train_sampler,
num_workers=args.n_worker,
)
test_loader = DataLoader(
dataset_test,
args.val_batch_size,
sampler=test_sampler,
num_workers=args.n_worker,
)
if local_rank == 0:
print("#train samples", len(dataset_train))
print("#test samples", len(dataset_test))
writer = SummaryWriter(log_dir=args.result_path)
n_cams = len(set(dataset_train.cameras).union(set(dataset_test.cameras)))
if args.arch == "base":
model = DeepAppearanceVAE(
args.tex_size, args.mesh_inp_size, n_latent=args.nlatent, n_cams=n_cams
).to(device)
elif args.arch == "res":
model = DeepAppearanceVAE(
args.tex_size,
args.mesh_inp_size,
n_latent=args.nlatent,
res=True,
n_cams=n_cams,
).to(device)
elif args.arch == "warp":
model = WarpFieldVAE(
args.tex_size, args.mesh_inp_size, z_dim=args.nlatent, n_cams=n_cams
).to(device)
elif args.arch == "non":
model = DeepAppearanceVAE(
args.tex_size,
args.mesh_inp_size,
n_latent=args.nlatent,
res=False,
non=True,
n_cams=n_cams,
).to(device)
elif args.arch == "bilinear":
model = DeepAppearanceVAE(
args.tex_size,
args.mesh_inp_size,
n_latent=args.nlatent,
res=False,
non=False,
bilinear=True,
n_cams=n_cams,
).to(device)
else:
raise NotImplementedError
model = torch.nn.parallel.DistributedDataParallel(model, [local_rank], local_rank)
renderer = Renderer()
if args.model_ckpt is not None:
print("loading checkpoint from", args.model_ckpt)
map_location = {"cuda:%d" % 0: "cuda:%d" % local_rank}
model.load_state_dict(torch.load(args.model_ckpt, map_location=map_location))
optimizer = optim.Adam(model.module.get_model_params(), args.lr, (0.9, 0.999))
optimizer_cc = optim.Adam(model.module.get_cc_params(), args.lr, (0.9, 0.999))
mse = nn.MSELoss()
scheduler = optim.lr_scheduler.StepLR(optimizer, step_size=1, gamma=0.95)
texmean = cv2.resize(dataset_train.texmean, (args.tex_size, args.tex_size))
texmin = cv2.resize(dataset_train.texmin, (args.tex_size, args.tex_size))
texmax = cv2.resize(dataset_train.texmax, (args.tex_size, args.tex_size))
texmean = torch.tensor(texmean).permute((2, 0, 1))[None, ...].to(device)
texmin = torch.tensor(texmin).permute((2, 0, 1))[None, ...].to(device)
texmax = torch.tensor(texmax).permute((2, 0, 1))[None, ...].to(device)
texstd = dataset_train.texstd
vertmean = (
torch.tensor(dataset_train.vertmean, dtype=torch.float32)
.view((1, -1, 3))
.to(device)
)
vertstd = dataset_train.vertstd
loss_weight_mask = cv2.flip(cv2.imread(args.loss_weight_mask), 0)
loss_weight_mask = loss_weight_mask / loss_weight_mask.max()
loss_weight_mask = (
torch.tensor(loss_weight_mask).permute(2, 0, 1).unsqueeze(0).float().to(device)
)
os.makedirs(args.result_path, exist_ok=True)
def run_net(data):
M = data["M"].cuda()
gt_tex = data["tex"].cuda()
vert_ids = data["vert_ids"].cuda()
uvs = data["uvs"].cuda()
uv_ids = data["uv_ids"].cuda()
avg_tex = data["avg_tex"].cuda()
view = data["view"].cuda()
transf = data["transf"].cuda()
verts = data["aligned_verts"].cuda()
photo = data["photo"].cuda()
mask = data["mask"].cuda()
cams = data["cam"].cuda()
batch, channel, height, width = avg_tex.shape
output = {}
if args.arch == "warp":
pred_tex, pred_verts, unwarped_tex, warp_field, kl = model(
avg_tex, verts, view, cams=cams
)
output["unwarped_tex"] = unwarped_tex
output["warp_field"] = warp_field
else:
pred_tex, pred_verts, kl = model(avg_tex, verts, view, cams=cams)
vert_loss = mse(pred_verts, verts)
pred_verts = pred_verts * vertstd + vertmean
pred_tex = (pred_tex * texstd + texmean) / 255.0
gt_tex = (gt_tex * texstd + texmean) / 255.0
loss_mask = loss_weight_mask.repeat(batch, 1, 1, 1)
tex_loss = mse(pred_tex * mask, gt_tex * mask) * (255**2) / (texstd**2)
if args.lambda_screen > 0:
screen_mask, rast_out = renderer.render(
M, pred_verts, vert_ids, uvs, uv_ids, loss_mask, args.resolution
)
pred_screen, rast_out = renderer.render(
M, pred_verts, vert_ids, uvs, uv_ids, pred_tex, args.resolution
)
screen_loss = (
torch.mean((pred_screen - photo) ** 2 * screen_mask)
* (255**2)
/ (texstd**2)
)
else:
screen_loss, pred_screen = torch.zeros([]), None
total_loss = 0
if args.lambda_verts > 0:
total_loss = total_loss + args.lambda_verts * vert_loss
if args.lambda_tex > 0:
total_loss = total_loss + args.lambda_tex * tex_loss
if args.lambda_screen > 0:
total_loss = total_loss + args.lambda_screen * screen_loss
if args.lambda_kl > 0:
total_loss = total_loss + args.lambda_kl * kl
losses = {
"total_loss": total_loss,
"vert_loss": vert_loss,
"screen_loss": screen_loss,
"tex_loss": tex_loss,
"denorm_tex_loss": tex_loss * (texstd**2),
"kl": kl,
}
output["pred_screen"] = pred_screen
output["pred_verts"] = pred_verts
output["pred_tex"] = pred_tex
return losses, output
def save_img(data, output, tag=""):
gt_screen = data["photo"] * 255
gt_tex = data["tex"].cuda() * texstd + texmean
pred_tex = torch.clamp(output["pred_tex"] * 255, 0, 255)
if output["pred_screen"] is not None:
pred_screen = torch.clamp(output["pred_screen"] * 255, 0, 255)
# apply gamma correction
save_pred_image = pred_screen.detach().cpu().numpy().astype(np.uint8)
save_pred_image = (255 * gammaCorrect(save_pred_image / 255.0)).astype(np.uint8)
Image.fromarray(save_pred_image).save(os.path.join(args.result_path, "pred_%s.png" % tag))
# apply gamma correction
save_gt_image = gt_screen[-1].detach().cpu().numpy().astype(np.uint8)
save_gt_image = (255 * gammaCorrect(save_gt_image / 255.0)).astype(np.uint8)
Image.fromarray(save_gt_image).save(os.path.join(args.result_path, "gt_%s.png" % tag))
# apply gamma correction
save_gt_tex_image = gt_tex[-1].detach().permute((1, 2, 0)).cpu().numpy().astype(np.uint8)
save_gt_tex_image = (255 * gammaCorrect(save_gt_tex_image / 255.0)).astype(np.uint8)
Image.fromarray(save_gt_tex_image).save(os.path.join(args.result_path, "gt_tex_%s.png" % tag))
# apply gamma correction
save_pred_tex_image = pred_tex[-1].detach().permute((1, 2, 0)).cpu().numpy().astype(np.uint8)
save_pred_tex_image = (255 * gammaCorrect(save_pred_tex_image / 255.0)).astype(np.uint8)
Image.fromarray(save_pred_tex_image).save(os.path.join(args.result_path, "pred_tex_%s.png" % tag))
if args.arch == "warp":
warp = output["warp_field"]
grid_img = (
torch.tensor(
np.array(
Image.open("grid.PNG").resize((args.tex_size, args.tex_size)),
dtype=np.float32,
)[None, ...]
)
.permute(0, 3, 1, 2)
.to(warp.device)
)
grid_img = F.grid_sample(grid_img, warp[-1:])
Image.fromarray(
grid_img[-1].detach().permute((1, 2, 0)).cpu().numpy().astype(np.uint8)
).save(os.path.join(args.result_path, "warp_grid_%s.png" % tag))
prev_loss = 1e8
prev_vert_loss = 1e8
prev_kl = 1e8
batch_idx, val_idx = 0, 0
best_screen_loss = 1e8
best_tex_loss = 1e8
best_vert_loss = 1e8
model.train()
train_screen_losses = []
train_tex_losses = []
train_vert_losses = []
window = 20
begin_time = time.time()
for epoch in range(args.epochs):
for i, data in enumerate(train_loader):
losses, output = run_net(data)
if batch_idx % args.val_every == 0:
if local_rank == 0:
torch.save(
model.state_dict(), os.path.join(args.result_path, "model.pth")
)
print(
"model.pth saved [Epoch {} Batch Index {}]".format(
epoch, batch_idx
)
)
if (
(losses["total_loss"].item() > args.pass_thres * prev_loss)
or (losses["vert_loss"].item() > args.pass_thres * prev_vert_loss)
or (losses["kl"].item() > args.pass_thres * prev_kl)
):
print("throw away batch")
continue
if local_rank == 0:
writer.add_scalar('train/loss_tex',losses['tex_loss'].item(), batch_idx)
writer.add_scalar('train/loss_verts', losses['vert_loss'].item(), batch_idx)
writer.add_scalar('train/loss_screen', losses['screen_loss'].item(), batch_idx)
writer.add_scalar('train/loss_kl', losses['kl'].item(), batch_idx)
prev_loss = losses["total_loss"].item()
prev_vert_loss = losses["vert_loss"].item()
prev_kl = losses["kl"].item()
train_screen_losses.append(losses["screen_loss"].item())
train_tex_losses.append(losses["tex_loss"].item())
train_vert_losses.append(losses["vert_loss"].item())
if len(train_screen_losses) > window:
del train_screen_losses[0]
del train_tex_losses[0]
del train_vert_losses[0]
optimizer.zero_grad()
optimizer_cc.zero_grad()
losses["total_loss"].backward()
torch.nn.utils.clip_grad_norm_(model.parameters(), 10)
optimizer.step()
optimizer_cc.step()
if batch_idx % args.log_every == 0:
print(
"%d loss %.3f vert %.3f tex %.3f screen %.3f kl %.3f"
% (
batch_idx,
losses["total_loss"].item(),
losses["vert_loss"].item(),
losses["tex_loss"].item(),
losses["screen_loss"].item(),
losses["kl"].item(),
)
)
if local_rank == 0:
save_img(data, output, "train_%d" % batch_idx)
if batch_idx % args.val_every == 0:
model.eval()
total, vert, tex, screen, kl = [], [], [], [], []
for i, data in enumerate(test_loader):
optimizer_cc.zero_grad()
losses, output = run_net(data)
losses["total_loss"].backward()
optimizer_cc.step()
if i == args.val_num:
break
for i, data in enumerate(test_loader):
with torch.no_grad():
losses, output = run_net(data)
total.append(losses["total_loss"].item())
vert.append(losses["vert_loss"].item())
tex.append(losses["tex_loss"].item())
screen.append(losses["screen_loss"].item())
kl.append(losses["kl"].item())
if i == args.val_num:
break
tex_loss = np.array(tex).mean()
vert_loss = np.array(vert).mean()
screen_loss = np.array(screen).mean()
kl = np.array(kl).mean()
if local_rank == 0:
writer.add_scalar('val/loss_tex',tex_loss, val_idx)
writer.add_scalar('val/loss_verts', vert_loss, val_idx)
writer.add_scalar('val/loss_screen', screen_loss, val_idx)
writer.add_scalar('val/loss_kl', kl, val_idx)
save_img(data, output, "val_%d" % val_idx)
val_idx += 1
print(
"val %d vert %.3f tex %.3f screen %.3f kl %.3f"
% (val_idx, vert_loss, tex_loss, screen_loss, kl)
)
best_screen_loss = min(best_screen_loss, screen_loss)
best_tex_loss = min(best_tex_loss, tex_loss)
best_vert_loss = min(best_vert_loss, vert_loss)
if local_rank == 0:
if (args.lambda_screen > 0 and best_screen_loss == screen_loss) or (
args.lambda_screen == 0 and best_tex_loss == tex_loss
):
torch.save(
model.state_dict(),
os.path.join(args.result_path, "best_model.pth"),
)
model.train()
if batch_idx >= args.max_iter:
print(
"best screen loss %f, best tex loss %f best vert loss %f"
% (best_screen_loss, best_tex_loss, best_vert_loss)
)
if local_rank == 0:
torch.save(
model.state_dict(), os.path.join(args.result_path, "model.pth")
)
train_screen_loss = np.mean(np.array(train_screen_losses))
train_tex_loss = np.mean(np.array(train_tex_losses))
train_vert_loss = np.mean(np.array(train_vert_losses))
end_time = time.time()
print("Training takes %f seconds" % (end_time - begin_time))
return (
best_screen_loss,
best_tex_loss,
best_vert_loss,
screen_loss,
tex_loss,
vert_loss,
train_screen_loss,
train_tex_loss,
train_vert_loss,
)
batch_idx += 1
scheduler.step()
print(
"best screen loss %f, best tex loss %f best vert loss %f"
% (best_screen_loss, best_tex_loss, best_vert_loss)
)
if local_rank == 0:
torch.save(model.state_dict(), os.path.join(args.result_path, "model.pth"))
train_screen_loss = np.mean(np.array(train_screen_losses))
train_tex_loss = np.mean(np.array(train_tex_losses))
train_vert_loss = np.mean(np.array(train_vert_losses))
end_time = time.time()
print("Training takes %f seconds" % (end_time - begin_time))
return (
best_screen_loss,
best_tex_loss,
best_vert_loss,
screen_loss,
tex_loss,
vert_loss,
train_screen_loss,
train_tex_loss,
train_vert_loss,
)
if __name__ == "__main__":
torch.distributed.init_process_group(backend="nccl")
parser = argparse.ArgumentParser(description="Process some integers.")
parser.add_argument(
"--local_rank", type=int, default=0, help="Local rank for distributed run"
)
parser.add_argument(
"--train_batch_size", type=int, default=16, help="Training batch size"
)
parser.add_argument(
"--val_batch_size", type=int, default=8, help="Validation batch size"
)
parser.add_argument(
"--arch",
type=str,
default="base",
help="Model architecture - base|warp|res|non|bilinear",
)
parser.add_argument(
"--nlatent", type=int, default=256, help="Latent code dimension - 128|256"
)
parser.add_argument(
"--lr", type=float, default=3e-4, help="Learning rate for training"
)
parser.add_argument(
"--resolution",
default=[2048, 1334],
nargs=2,
type=int,
help="Rendering resolution",
)
parser.add_argument("--tex_size", type=int, default=1024, help="Texture resolution")
parser.add_argument(
"--mesh_inp_size", type=int, default=21918, help="Input mesh dimension"
)
parser.add_argument(
"--epochs", type=int, default=50, help="Number of training epochs"
)
parser.add_argument(
"--data_dir",
type=str,
default="/mnt/captures/zhengningyuan/m--20180226--0000--6674443--GHS",
help="Directory to dataset root",
)
parser.add_argument(
"--krt_dir",
type=str,
default="/mnt/captures/zhengningyuan/m--20180226--0000--6674443--GHS/KRT",
help="Directory to KRT file",
)
parser.add_argument(
"--loss_weight_mask",
type=str,
default="./loss_weight_mask.png",
help="Mask for weighted loss of face",
)
parser.add_argument(
"--framelist_train",
type=str,
default="/mnt/captures/zhengningyuan/m--20180226--0000--6674443--GHS/frame_list.txt",
help="Frame list for training",
)
parser.add_argument(
"--framelist_test",
type=str,
default="/mnt/captures/zhengningyuan/m--20180226--0000--6674443--GHS/frame_list.txt",
help="Frame list for testing",
)
parser.add_argument(
"--test_segment_config",
type=str,
default=None,
help="Directory of expression segments for testing (exclude from training)",
)
parser.add_argument(
"--lambda_verts", type=float, default=1, help="Multiplier of vertex loss"
)
parser.add_argument(
"--lambda_screen", type=float, default=0, help="Multiplier of screen loss"
)
parser.add_argument(
"--lambda_tex", type=float, default=1, help="Multiplier of texture loss"
)
parser.add_argument(
"--lambda_kl", type=float, default=1e-2, help="Multiplier of KL divergence"
)
parser.add_argument(
"--max_iter",
type=int,
default=200000,
help="Maximum number of training iterations, overrides epoch",
)
parser.add_argument(
"--log_every", type=int, default=1000, help="Interval of printing training loss"
)
parser.add_argument(
"--val_every", type=int, default=5000, help="Interval of validating on test set"
)
parser.add_argument(
"--val_num", type=int, default=500, help="Number of iterations for validation"
)
parser.add_argument(
"--n_worker", type=int, default=8, help="Number of workers loading dataset"
)
parser.add_argument(
"--pass_thres",
type=int,
default=50,
help="If loss is x times higher than the previous batch, discard this batch",
)
parser.add_argument(
"--result_path",
type=str,
default="./runs/experiment",
help="Directory to output files",
)
parser.add_argument(
"--model_ckpt", type=str, default=None, help="Model checkpoint path"
)
experiment_args = parser.parse_args()
print(experiment_args)
# load camera config
subject_id = experiment_args.data_dir.split("--")[-2]
camera_config_path = f"camera_configs/camera-split-config_{subject_id}.json"
if os.path.exists(camera_config_path):
print(f"camera config file for {subject_id} exists, loading...")
f = open(camera_config_path, "r")
camera_config = json.load(f)
f.close()
else:
print(f"camera config file for {subject_id} NOT exists, generating...")
# generate camera config based on downloaded data if not existed
segments = [os.path.basename(x) for x in glob.glob(f"{experiment_args.data_dir}/unwrapped_uv_1024/*")]
assert len(segments) > 0
# select a segment to check available camera ids
camera_ids = [os.path.basename(x) for x in glob.glob(f"{experiment_args.data_dir}/unwrapped_uv_1024/{segments[0]}/*")]
camera_ids.remove('average')
camera_config = {
"full": {
"train": camera_ids,
"test": camera_ids,
"visual": camera_ids[:2]
}
}
# save the config for future use
os.makedirs("camera_configs", exist_ok=True)
with open(camera_config_path, 'w') as f:
json.dump(camera_config, f)
camera_set = camera_config["full"]
if experiment_args.test_segment_config is not None:
f = open(experiment_args.test_segment_config, "r")
test_segment_config = json.load(f)
f.close()
test_segment = test_segment_config["segment"]
else:
test_segment = ["EXP_ROM", "EXP_free_face"]
(
best_screen_loss,
best_tex_loss,
best_vert_loss,
screen_loss,
tex_loss,
vert_loss,
train_screen_loss,
train_tex_loss,
train_vert_loss,
) = main(experiment_args, camera_set, test_segment)
if torch.distributed.get_rank() == 0:
print(
best_screen_loss,
best_tex_loss,
best_vert_loss,
screen_loss,
tex_loss,
vert_loss,
train_screen_loss,
train_tex_loss,
train_vert_loss,
)
f = open(os.path.join(experiment_args.result_path, "result.txt"), "a")
f.write("\n")
f.write(
"Best screen loss %f, best tex loss %f, best vert loss %f, screen loss %f, tex loss %f, vert_loss %f, train screen loss %f, train tex loss %f, train vert loss %f"
% (
best_screen_loss,
best_tex_loss,
best_vert_loss,
screen_loss,
tex_loss,
vert_loss,
train_screen_loss,
train_tex_loss,
train_vert_loss,
)
)
f.close()