-
Notifications
You must be signed in to change notification settings - Fork 14
/
diff.txt
1119 lines (1067 loc) · 57.7 KB
/
diff.txt
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
--- ./stylegan3/train.py 2021-11-24 13:11:24.000000000 -0500
+++ train.py 2022-05-17 17:39:19.470176637 -0400
@@ -75,6 +75,12 @@
print(f'Dataset resolution: {c.training_set_kwargs.resolution}')
print(f'Dataset labels: {c.training_set_kwargs.use_labels}')
print(f'Dataset x-flips: {c.training_set_kwargs.xflip}')
+ if 'patch' in c.G_kwargs.training_mode:
+ print(f'Patches path: {c.patch_kwargs.path}')
+ print(f'Patches size: {c.patch_kwargs.max_size} images')
+ print(f'Patches resolution: {c.patch_kwargs.resolution}')
+ print(f'Patches labels: {c.patch_kwargs.use_labels}')
+ print(f'Patches x-flips: {c.patch_kwargs.xflip}')
print()
# Dry run?
@@ -103,7 +109,11 @@
try:
dataset_kwargs = dnnlib.EasyDict(class_name='training.dataset.ImageFolderDataset', path=data, use_labels=True, max_size=None, xflip=False)
dataset_obj = dnnlib.util.construct_class_by_name(**dataset_kwargs) # Subclass of training.dataset.Dataset.
+ try:
dataset_kwargs.resolution = dataset_obj.resolution # Be explicit about resolution.
+ except AssertionError:
+ print("Cannot determine default dataset resolution, will try to use specified arguments")
+ dataset_kwargs.resolution = None
dataset_kwargs.use_labels = dataset_obj.has_labels # Be explicit about labels.
dataset_kwargs.max_size = len(dataset_obj) # Be explicit about dataset size.
return dataset_kwargs, dataset_obj.name
@@ -161,28 +171,32 @@
@click.option('--workers', help='DataLoader worker processes', metavar='INT', type=click.IntRange(min=1), default=3, show_default=True)
@click.option('-n','--dry-run', help='Print training options and exit', is_flag=True)
-def main(**kwargs):
- """Train a GAN using the techniques described in the paper
- "Alias-Free Generative Adversarial Networks".
+# additional base options
+@click.option('--training_mode', help='generator training mode', type=click.Choice(['global', 'patch', 'global-360']), required=True)
+@click.option('--data_resolution', help='LR dataset resolution (specify if images are not preprocessed to same size and square)', type=click.IntRange(min=0))
+@click.option('--random_crop', help='random crop image on LR dataset (specify if images are not preprocessed to same size and square)', metavar='BOOL', type=bool, default=False, show_default=True)
+@click.option('--data_max_size', help='LR dataset max number of images', type=click.IntRange(min=0))
+@click.option('--g_size', help='size of G (if different from dataset size)', type=click.IntRange(min=0))
+
+# additional options for patch model
+@click.option('--teacher', help='teacher checkpoint', metavar='[PATH|URL]', type=str)
+@click.option('--teacher_lambda', help='teacher regularization weight', metavar='FLOAT', type=click.FloatRange(min=0), default=1.0, show_default=True)
+@click.option('--teacher_mode', help='teacher loss mode', type=click.Choice(['inverse', 'forward']), default='forward', show_default=True)
+@click.option('--scale_anneal', help='scale annealing rate (-1 for no annealing)', metavar='FLOAT', type=click.FloatRange(min=-1), default=-1, show_default=True)
+@click.option('--scale_min', help='minimum sampled scale (leave blank to use image native resolution)', metavar='FLOAT', type=click.FloatRange(min=0))
+@click.option('--scale_max', help='maximum sampled scale', metavar='FLOAT', type=click.FloatRange(min=0), default=1.0, show_default=True)
+@click.option('--base_probability', help='probability to sample from LR dataset with identity transform', metavar='FLOAT', type=click.FloatRange(min=0), default=0.5, show_default=True)
+@click.option('--data_hr', help='HR patch dataset path', metavar='[ZIP|DIR]', type=str)
+@click.option('--patch_crop', help='perform random cropping on non-square images (on patch dataset)', metavar='BOOL', type=bool, default=False, show_default=True)
+@click.option('--data_hr_max_size', help='patch dataset max number of images', type=click.IntRange(min=0))
+@click.option('--scale_mapping_min', help='normalization minimum for scale mapping branch (size = g_size*scale_mapping_min)', type=click.IntRange(min=0))
+@click.option('--scale_mapping_max', help='normalization maximum for scale mapping branch (size = g_size*scale_mapping_max)', type=click.IntRange(min=0))
+@click.option('--scale_mapping_norm', help='normalization type for scale mapping branch', type=click.Choice(['positive', 'zerocentered']), default='positive')
- Examples:
+# additional options for 360 model
+@click.option('--fov', help='fov for one frame in the 360 model', type=click.IntRange(min=0), default=60, show_default=True)
- \b
- # Train StyleGAN3-T for AFHQv2 using 8 GPUs.
- python train.py --outdir=~/training-runs --cfg=stylegan3-t --data=~/datasets/afhqv2-512x512.zip \\
- --gpus=8 --batch=32 --gamma=8.2 --mirror=1
-
- \b
- # Fine-tune StyleGAN3-R for MetFaces-U using 1 GPU, starting from the pre-trained FFHQ-U pickle.
- python train.py --outdir=~/training-runs --cfg=stylegan3-r --data=~/datasets/metfacesu-1024x1024.zip \\
- --gpus=8 --batch=32 --gamma=6.6 --mirror=1 --kimg=5000 --snap=5 \\
- --resume=https://api.ngc.nvidia.com/v2/models/nvidia/research/stylegan3/versions/1/files/stylegan3-r-ffhqu-1024x1024.pkl
-
- \b
- # Train StyleGAN2 for FFHQ at 1024x1024 resolution using 8 GPUs.
- python train.py --outdir=~/training-runs --cfg=stylegan2 --data=~/datasets/ffhq-1024x1024.zip \\
- --gpus=8 --batch=32 --gamma=10 --mirror=1 --aug=noaug
- """
+def main(**kwargs):
# Initialize config.
opts = dnnlib.EasyDict(kwargs) # Command line arguments.
@@ -200,6 +214,61 @@
raise click.ClickException('--cond=True requires labels specified in dataset.json')
c.training_set_kwargs.use_labels = opts.cond
c.training_set_kwargs.xflip = opts.mirror
+ if opts.data_max_size:
+ c.training_set_kwargs.max_size = opts.max_size
+ if opts.data_resolution:
+ if c.training_set_kwargs.resolution != opts.data_resolution:
+ print("using specified data resolution %d rather than default" % (opts.data_resolution))
+ c.training_set_kwargs.resolution = opts.data_resolution
+ c.training_set_kwargs.crop_image = opts.random_crop
+ # by this point, resolution should be determined
+ # either from init_dataset function or opts.data_resolution
+ assert(c.training_set_kwargs.resolution is not None)
+
+ # set up training mode
+ training_mode = c.G_kwargs.training_mode = opts.training_mode
+ # set up generator size
+ if opts.g_size is not None:
+ assert(opts.g_size == c.training_set_kwargs.resolution)
+ else:
+ opts.g_size = c.training_set_kwargs.resolution
+ if 'patch' in training_mode:
+ # patch dataset kwargs
+ patch_kwargs = dnnlib.EasyDict(
+ class_name='training.dataset.ImagePatchDataset',
+ path=opts.data_hr, resolution=opts.g_size,
+ scale_min=opts.scale_min, scale_max=opts.scale_max,
+ scale_anneal=opts.scale_anneal, random_crop=opts.patch_crop,
+ use_labels=True, max_size=None, xflip=False)
+ patch_obj = dnnlib.util.construct_class_by_name(**patch_kwargs) # gets initial args
+ patch_name = patch_obj.name
+ patch_kwargs.resolution = patch_obj.resolution # Be explicit about resolution.
+ patch_kwargs.use_labels = patch_obj.has_labels # Be explicit about labels.
+ patch_kwargs.max_size = len(patch_obj) # Be explicit about dataset size.
+ c.patch_kwargs = patch_kwargs
+ c.patch_kwargs.use_labels = opts.cond
+ c.patch_kwargs.xflip = opts.mirror
+ if opts.data_hr_max_size:
+ c.patch_kwargs.max_size = opts.data_hr_max_size
+ # added G_kwargs
+ c.G_kwargs.scale_mapping_kwargs = dnnlib.EasyDict(
+ scale_mapping_min = opts.scale_mapping_min,
+ scale_mapping_max = opts.scale_mapping_max,
+ scale_mapping_norm = opts.scale_mapping_norm
+ )
+ # added training options
+ c.added_kwargs = dnnlib.EasyDict(
+ img_size=opts.g_size,
+ teacher=opts.teacher,
+ teacher_lambda=opts.teacher_lambda,
+ teacher_mode=opts.teacher_mode,
+ scale_min=opts.scale_min,
+ scale_max=opts.scale_max,
+ scale_anneal=opts.scale_anneal,
+ base_probability=opts.base_probability,
+ )
+ elif '360' in training_mode:
+ c.G_kwargs.fov = opts.fov
# Hyperparameters & settings.
c.num_gpus = opts.gpus
@@ -218,6 +287,8 @@
c.kimg_per_tick = opts.tick
c.image_snapshot_ticks = c.network_snapshot_ticks = opts.snap
c.random_seed = c.training_set_kwargs.random_seed = opts.seed
+ if 'patch' in training_mode:
+ c.patch_kwargs.random_seed = opts.seed
c.data_loader_kwargs.num_workers = opts.workers
# Sanity checks.
@@ -261,6 +332,9 @@
# Resume.
if opts.resume is not None:
c.resume_pkl = opts.resume
+
+ if opts.teacher is not None or opts.resume is not None:
+ # disable rampups for finetuning or resuming models
c.ada_kimg = 100 # Make ADA react faster at the beginning.
c.ema_rampup = None # Disable EMA rampup.
c.loss_kwargs.blur_init_sigma = 0 # Disable blur rampup.
diff -bur ./stylegan3/training/dataset.py training/dataset.py
--- ./stylegan3/training/dataset.py 2021-11-24 13:11:24.000000000 -0500
+++ training/dataset.py 2022-06-17 23:26:59.794824950 -0400
@@ -16,10 +16,12 @@
import torch
import dnnlib
-try:
- import pyspng
-except ImportError:
- pyspng = None
+pyspng = None # disable pyspng for image resizing on load
+# https://stackoverflow.com/questions/51152059/pillow-in-python-wont-let-me-open-image-exceeds-limit
+PIL.Image.MAX_IMAGE_PIXELS = 933120000
+from util import patch_util
+import random
+
#----------------------------------------------------------------------------
@@ -85,6 +87,7 @@
return self._raw_idx.size
def __getitem__(self, idx):
+ if not self.is_patch: # full image loader
image = self._load_raw_image(self._raw_idx[idx])
assert isinstance(image, np.ndarray)
assert list(image.shape) == self.image_shape
@@ -93,6 +96,14 @@
assert image.ndim == 3 # CHW
image = image[:, :, ::-1]
return image.copy(), self.get_label(idx)
+ else: # image patch loader
+ # handle xflips when loading the image
+ data = self._load_raw_image(self._raw_idx[idx], self._xflip[idx])
+ assert isinstance(data, dict)
+ assert list(data['image'].shape) == self.image_shape
+ assert data['image'].dtype == np.uint8
+ data['image'] = data['image'].copy()
+ return data, self.get_label(idx)
def get_label(self, idx):
label = self._get_raw_labels()[self._raw_idx[idx]]
@@ -153,18 +164,30 @@
#----------------------------------------------------------------------------
-class ImageFolderDataset(Dataset):
+class BaseImageDataset(Dataset):
def __init__(self,
path, # Path to directory or zip.
resolution = None, # Ensure specific resolution, None = highest available.
**super_kwargs, # Additional arguments for the Dataset base class.
):
+
self._path = path
self._zipfile = None
if os.path.isdir(self._path):
self._type = 'dir'
- self._all_fnames = {os.path.relpath(os.path.join(root, fname), start=self._path) for root, _dirs, files in os.walk(self._path) for fname in files}
+ if os.path.isfile(self._path + '_cache.txt'):
+ # use cache file if it exists
+ with open(self._path + '_cache.txt') as cache:
+ self._all_fnames = set([line.strip() for line in cache])
+ else:
+ print("Walking dataset...")
+ self._all_fnames = [os.path.relpath(os.path.join(root, fname), start=self._path)
+ for root, _dirs, files in os.walk(self._path, followlinks=True) for fname in files]
+ with open(self._path + '_cache.txt', 'w') as cache:
+ [cache.write("%s\n" % fname) for fname in self._all_fnames]
+ self._all_fnames = set(self._all_fnames)
+ print("Done walking")
elif self._file_ext(self._path) == '.zip':
self._type = 'zip'
self._all_fnames = set(self._get_zipfile().namelist())
@@ -177,9 +200,12 @@
raise IOError('No image files found in the specified path')
name = os.path.splitext(os.path.basename(self._path))[0]
- raw_shape = [len(self._image_fnames)] + list(self._load_raw_image(0).shape)
- if resolution is not None and (raw_shape[2] != resolution or raw_shape[3] != resolution):
- raise IOError('Image files do not match the specified resolution')
+ if resolution is not None:
+ raw_shape = [len(self._image_fnames)] + [3, resolution, resolution]
+ else:
+ # do not resize it to determine initial shape (will fail if images not square)
+ raw_shape = [len(self._image_fnames)] + list(self._load_raw_image(0, resize=False).shape)
+
super().__init__(name=name, raw_shape=raw_shape, **super_kwargs)
@staticmethod
@@ -209,18 +235,6 @@
def __getstate__(self):
return dict(super().__getstate__(), _zipfile=None)
- def _load_raw_image(self, raw_idx):
- fname = self._image_fnames[raw_idx]
- with self._open_file(fname) as f:
- if pyspng is not None and self._file_ext(fname) == '.png':
- image = pyspng.load(f.read())
- else:
- image = np.array(PIL.Image.open(f))
- if image.ndim == 2:
- image = image[:, :, np.newaxis] # HW => HWC
- image = image.transpose(2, 0, 1) # HWC => CHW
- return image
-
def _load_raw_labels(self):
fname = 'dataset.json'
if fname not in self._all_fnames:
@@ -235,4 +249,116 @@
labels = labels.astype({1: np.int64, 2: np.float32}[labels.ndim])
return labels
-#----------------------------------------------------------------------------
+
+class ImageFolderDataset(BaseImageDataset):
+ def __init__(self,
+ path, # Path to directory or zip.
+ resolution = None, # Ensure specific resolution, None = highest available.
+ crop_image = False, # default: assumes inputs are square images, if True it will perform a random crop
+ **super_kwargs, # Additional arguments for the Dataset base class.
+ ):
+ self.crop_image = crop_image
+ self.is_patch = False
+ super().__init__(path=path, resolution=resolution, **super_kwargs)
+
+ def _load_raw_image(self, raw_idx, resize=True):
+ fname = self._image_fnames[raw_idx]
+ with self._open_file(fname) as f:
+ if pyspng is not None and self._file_ext(fname) == '.png':
+ image = pyspng.load(f.read())
+ else:
+ image = PIL.Image.open(f).convert('RGB')
+ w, h = image.size
+ if self.crop_image and w != h:
+ # perform random crop if needed
+ min_size = min(w, h)
+ if w == min_size:
+ x_start = 0
+ y_start = random.randint(0, h - min_size)
+ else:
+ x_start = random.randint(0, w - min_size)
+ y_start = 0
+ image = image.crop((x_start, y_start, x_start+min_size, y_start+min_size))
+ if resize:
+ # at this point it should be square
+ assert(image.size[0] == image.size[1])
+ target_size = tuple(self.image_shape[1:])
+ if image.size != target_size:
+ # it should only downsize, but there are a small number
+ # of images in the datasets that are a few pixels
+ # smaller than 256, so allow a small leeway
+ assert(target_size[-1] < image.size[-1] + 10)
+ image = image.resize(target_size, PIL.Image.ANTIALIAS)
+ image = np.array(image)
+ if image.ndim == 2:
+ image = image[:, :, np.newaxis] # HW => HWC
+ image = image.transpose(2, 0, 1) # HWC => CHW
+ return image
+
+class ImagePatchDataset(BaseImageDataset):
+ def __init__(self,
+ path, # Path to directory or zip.
+ resolution, # patch size
+ scale_min, # minimum scale of the patches (largest image size)
+ scale_max, # maximum scale of the patches (smallest image size)
+ scale_anneal=-1, # annealing rate
+ random_crop=True, # add random crop for non-square images
+ **super_kwargs, # Additional arguments for the Dataset base class.
+ ):
+ assert(resolution is not None) # patch resolution must be specified
+
+ # annealing not implemented, need to update iteration counter and
+ # adjust counter when resuming training
+ assert(scale_anneal == -1)
+
+ # crop sampler
+ self.patch_size = resolution
+ self.random_crop = random_crop
+ self.sampler = patch_util.PatchSampler(
+ patch_size=self.patch_size, scale_anneal=scale_anneal,
+ min_scale=scale_min, max_scale=scale_max)
+ self.is_patch = True
+
+ super().__init__(path=path, resolution=resolution, **super_kwargs)
+
+ def _load_raw_image(self, raw_idx, is_flipped):
+ fname = self._image_fnames[raw_idx]
+ with self._open_file(fname) as f:
+ if pyspng is not None and self._file_ext(fname) == '.png':
+ image = pyspng.load(f.read())
+ else:
+ image = PIL.Image.open(f).convert('RGB')
+
+ # first, flip image if necessary
+ if is_flipped:
+ image = image.transpose(PIL.Image.FLIP_LEFT_RIGHT)
+
+ # add random crop if necessary
+ if self.random_crop:
+ w, h = image.size
+ min_size = min(w, h)
+ x_start = random.randint(0, max(0, w - min_size))
+ y_start = random.randint(0, max(0, h - min_size))
+ image = image.crop((x_start, y_start, x_start+min_size, y_start+min_size))
+ else:
+ # otherwise, center crop
+ w, h = image.size
+ min_size = min(w, h)
+ if w != h:
+ if w == min_size:
+ x_start = 0
+ y_start = (h - min_size) // 2
+ else:
+ x_start = (w - min_size) // 2
+ y_start = 0
+ image = image.crop((x_start, y_start, x_start+min_size, y_start+min_size))
+
+ # sample the resize and crop parameters
+ crop, params = self.sampler.sample_patch(image)
+ image = np.asarray(crop)
+ image = image.transpose(2, 0, 1) # HWC => CHW
+ data = {
+ 'image': image,
+ 'params': params,
+ }
+ return data
diff -bur ./stylegan3/training/loss.py training/loss.py
--- ./stylegan3/training/loss.py 2021-11-24 13:11:24.000000000 -0500
+++ training/loss.py 2022-06-17 23:33:28.910631833 -0400
@@ -14,6 +14,11 @@
from torch_utils.ops import conv2d_gradfix
from torch_utils.ops import upfirdn2d
+# added imports
+from metrics import equivariance
+from util import losses, util, patch_util
+import random
+
#----------------------------------------------------------------------------
class Loss:
@@ -22,8 +27,23 @@
#----------------------------------------------------------------------------
+def apply_affine_batch(img, transform):
+ # hacky .. apply affine transformation with cuda kernel in batch form
+ crops = []
+ masks = []
+ for i, t in zip(img, transform):
+ crop, mask = equivariance.apply_affine_transformation(
+ i[None], t.inverse())
+ crops.append(crop)
+ masks.append(mask)
+ crops = torch.cat(crops, dim=0)
+ masks = torch.cat(masks, dim=0)
+ return crops, masks
+
class StyleGAN2Loss(Loss):
- def __init__(self, device, G, D, augment_pipe=None, r1_gamma=10, style_mixing_prob=0, pl_weight=0, pl_batch_shrink=2, pl_decay=0.01, pl_no_weight_grad=False, blur_init_sigma=0, blur_fade_kimg=0):
+ def __init__(self, device, G, D, augment_pipe=None, r1_gamma=10, style_mixing_prob=0, pl_weight=0, pl_batch_shrink=2,
+ pl_decay=0.01, pl_no_weight_grad=False, blur_init_sigma=0,
+ blur_fade_kimg=0, teacher=None, added_kwargs=None):
super().__init__()
self.device = device
self.G = G
@@ -39,14 +59,51 @@
self.blur_init_sigma = blur_init_sigma
self.blur_fade_kimg = blur_fade_kimg
- def run_G(self, z, c, update_emas=False):
- ws = self.G.mapping(z, c, update_emas=update_emas)
+ self.teacher = teacher
+ self.added_kwargs = added_kwargs
+ self.training_mode = self.G.training_mode
+ if self.teacher is not None:
+ self.loss_l1 = losses.Masked_L1_Loss().to(device)
+ self.loss_lpips = losses.Masked_LPIPS_Loss(net='alex', device=device)
+ util.set_requires_grad(False, self.loss_lpips)
+ util.set_requires_grad(False, self.teacher)
+
+ def style_mix(self, z, c, ws):
if self.style_mixing_prob > 0:
with torch.autograd.profiler.record_function('style_mixing'):
cutoff = torch.empty([], dtype=torch.int64, device=ws.device).random_(1, ws.shape[1])
cutoff = torch.where(torch.rand([], device=ws.device) < self.style_mixing_prob, cutoff, torch.full_like(cutoff, ws.shape[1]))
ws[:, cutoff:] = self.G.mapping(torch.randn_like(z), c, update_emas=False)[:, cutoff:]
- img = self.G.synthesis(ws, update_emas=update_emas)
+ return ws
+
+ def run_G(self, z, c, transform, update_emas=False):
+ mapped_scale = None
+ crop_fn = None
+ if 'patch' in self.training_mode:
+ ws = self.G.mapping(z, c, update_emas=update_emas)
+ scale, mapped_scale = patch_util.compute_scale_inputs(self.G, ws, transform)
+ ws = self.style_mix(z, c, ws)
+ img = self.G.synthesis(ws, mapped_scale=mapped_scale, transform=transform, update_emas=update_emas)
+ elif '360' in self.training_mode:
+ ws = self.G.mapping(z, c, update_emas=update_emas)
+ ws = self.style_mix(z, c, ws)
+ input_layer = self.G.synthesis.input
+ crop_start = random.randint(0, 360 // input_layer.fov * input_layer.frame_size[0] - 1)
+ crop_fn = lambda grid : grid[:, :, crop_start:crop_start+input_layer.size[0], :]
+ img_base = self.G.synthesis(ws, crop_fn=crop_fn, update_emas=update_emas)
+ crop_shift = crop_start + input_layer.frame_size[0]
+ # generate shifted frame for cross-frame discriminator
+ crop_fn_shift = lambda grid : grid[:, :, crop_shift:crop_shift+input_layer.size[0], :]
+ img_shifted = self.G.synthesis(ws, crop_fn=crop_fn_shift, update_emas=update_emas)
+ img_splice = torch.cat([img_base, img_shifted], dim=3)
+ img_size = img_base.shape[-1]
+ splice_start = random.randint(0, img_size)
+ img = img_splice[:, :, :, splice_start:splice_start+img_size]
+ elif 'global' in self.training_mode:
+ ws = self.G.mapping(z, c, update_emas=update_emas)
+ ws = self.style_mix(z, c, ws)
+ assert(transform is None)
+ img = self.G.synthesis(ws, transform=transform, update_emas=update_emas)
return img, ws
def run_D(self, img, c, blur_sigma=0, update_emas=False):
@@ -60,7 +117,8 @@
logits = self.D(img, c, update_emas=update_emas)
return logits
- def accumulate_gradients(self, phase, real_img, real_c, gen_z, gen_c, gain, cur_nimg):
+ def accumulate_gradients(self, phase, real_img, real_c, transform, gen_z,
+ gen_c, gain, cur_nimg, min_scale, max_scale):
assert phase in ['Gmain', 'Greg', 'Gboth', 'Dmain', 'Dreg', 'Dboth']
if self.pl_weight == 0:
phase = {'Greg': 'none', 'Gboth': 'Gmain'}.get(phase, phase)
@@ -71,12 +129,49 @@
# Gmain: Maximize logits for generated images.
if phase in ['Gmain', 'Gboth']:
with torch.autograd.profiler.record_function('Gmain_forward'):
- gen_img, _gen_ws = self.run_G(gen_z, gen_c)
+ gen_img, _gen_ws = self.run_G(gen_z, gen_c, transform)
+ # vutils.save_image(gen_img, 'out_fake_patch.png', range=(-1, 1),
+ # normalize=True, nrow=4)
gen_logits = self.run_D(gen_img, gen_c, blur_sigma=blur_sigma)
training_stats.report('Loss/scores/fake', gen_logits)
training_stats.report('Loss/signs/fake', gen_logits.sign())
loss_Gmain = torch.nn.functional.softplus(-gen_logits) # -log(sigmoid(gen_logits))
training_stats.report('Loss/G/loss', loss_Gmain)
+ training_stats.report('Scale/G/min_scale', min_scale)
+ training_stats.report('Scale/G/max_scale', max_scale)
+ if self.teacher is not None and self.added_kwargs.teacher_lambda > 0:
+ teacher_img = self.teacher(gen_z, gen_c)
+ if self.added_kwargs.teacher_mode == 'forward':
+ teacher_crop, teacher_mask = apply_affine_batch(teacher_img, transform)
+ # removes the border around the above mask
+ # (mask should be all ones bc zooming in)
+ teacher_mask = torch.ones_like(teacher_mask)
+ l1_loss = self.loss_l1(gen_img, teacher_crop,
+ teacher_mask[:, :1])
+ lpips_loss = self.loss_lpips(
+ losses.adaptive_downsample256(gen_img),
+ losses.adaptive_downsample256(teacher_crop),
+ losses.adaptive_downsample256(teacher_mask[:, :1],
+ mode='nearest')
+ )
+ elif self.added_kwargs.teacher_mode == 'inverse':
+ out_crop, out_mask = apply_affine_batch(gen_img, transform.inverse())
+ l1_loss = self.loss_l1(out_crop, teacher_img,
+ out_mask[:, :1])
+ lpips_loss = self.loss_lpips(
+ losses.adaptive_downsample256(out_crop),
+ losses.adaptive_downsample256(teacher_img),
+ losses.adaptive_downsample256(out_mask[:, :1],
+ mode='nearest')
+ )
+ else:
+ assert(False)
+ teacher_loss = (l1_loss + lpips_loss)[:, None]
+ loss_Gmain = (loss_Gmain + self.added_kwargs.teacher_lambda
+ * teacher_loss)
+ training_stats.report('Loss/G/loss_teacher_l1', l1_loss)
+ training_stats.report('Loss/G/loss_teacher_lpips', lpips_loss)
+ training_stats.report('Loss/G/loss_total', loss_Gmain)
with torch.autograd.profiler.record_function('Gmain_backward'):
loss_Gmain.mean().mul(gain).backward()
@@ -84,7 +179,9 @@
if phase in ['Greg', 'Gboth']:
with torch.autograd.profiler.record_function('Gpl_forward'):
batch_size = gen_z.shape[0] // self.pl_batch_shrink
- gen_img, gen_ws = self.run_G(gen_z[:batch_size], gen_c[:batch_size])
+ gen_img, gen_ws = self.run_G(gen_z[:batch_size],
+ gen_c[:batch_size],
+ transform[:batch_size])
pl_noise = torch.randn_like(gen_img) / np.sqrt(gen_img.shape[2] * gen_img.shape[3])
with torch.autograd.profiler.record_function('pl_grads'), conv2d_gradfix.no_weight_gradients(self.pl_no_weight_grad):
pl_grads = torch.autograd.grad(outputs=[(gen_img * pl_noise).sum()], inputs=[gen_ws], create_graph=True, only_inputs=True)[0]
@@ -102,7 +199,7 @@
loss_Dgen = 0
if phase in ['Dmain', 'Dboth']:
with torch.autograd.profiler.record_function('Dgen_forward'):
- gen_img, _gen_ws = self.run_G(gen_z, gen_c, update_emas=True)
+ gen_img, _gen_ws = self.run_G(gen_z, gen_c, transform, update_emas=True)
gen_logits = self.run_D(gen_img, gen_c, blur_sigma=blur_sigma, update_emas=True)
training_stats.report('Loss/scores/fake', gen_logits)
training_stats.report('Loss/signs/fake', gen_logits.sign())
diff -bur ./stylegan3/training/networks_stylegan3.py training/networks_stylegan3.py
--- ./stylegan3/training/networks_stylegan3.py 2021-11-24 13:11:24.000000000 -0500
+++ training/networks_stylegan3.py 2022-06-14 22:42:13.642624353 -0400
@@ -18,6 +18,8 @@
from torch_utils.ops import conv2d_gradfix
from torch_utils.ops import filtered_lrelu
from torch_utils.ops import bias_act
+import math
+import random
#----------------------------------------------------------------------------
@@ -173,6 +175,7 @@
size, # Output spatial size: int or [width, height].
sampling_rate, # Output sampling rate.
bandwidth, # Output bandwidth.
+ margin_size, # Extra margin on input.
):
super().__init__()
self.w_dim = w_dim
@@ -180,6 +183,7 @@
self.size = np.broadcast_to(np.asarray(size), [2])
self.sampling_rate = sampling_rate
self.bandwidth = bandwidth
+ self.margin_size = margin_size
# Draw random frequencies from uniform 2D disc.
freqs = torch.randn([self.channels, 2])
@@ -195,9 +199,12 @@
self.register_buffer('freqs', freqs)
self.register_buffer('phases', phases)
- def forward(self, w):
+ def forward(self, w, transform=None, **kwargs):
# Introduce batch dimension.
- transforms = self.transform.unsqueeze(0) # [batch, row, col]
+ if transform is None:
+ # sanity check; should not modify transform from identity
+ assert(torch.equal(self.transform, torch.eye(3, 3).to(self.transform.device)))
+ transform = self.transform.unsqueeze(0) # [batch, row, col]
freqs = self.freqs.unsqueeze(0) # [batch, channel, xy]
phases = self.phases.unsqueeze(0) # [batch, channel]
@@ -212,7 +219,7 @@
m_t = torch.eye(3, device=w.device).unsqueeze(0).repeat([w.shape[0], 1, 1]) # Inverse translation wrt. resulting image.
m_t[:, 0, 2] = -t[:, 2] # t'_x
m_t[:, 1, 2] = -t[:, 3] # t'_y
- transforms = m_r @ m_t @ transforms # First rotate resulting image, then translate, and finally apply user-specified transform.
+ transforms = m_r @ m_t @ transform # First rotate resulting image, then translate, and finally apply user-specified transform.
# Transform frequencies.
phases = phases + (freqs @ transforms[:, :2, 2:]).squeeze(2)
@@ -247,6 +254,120 @@
f'w_dim={self.w_dim:d}, channels={self.channels:d}, size={list(self.size)},',
f'sampling_rate={self.sampling_rate:g}, bandwidth={self.bandwidth:g}'])
+
+@persistence.persistent_class
+class SynthesisInput360(torch.nn.Module):
+ def __init__(self,
+ w_dim, # Intermediate latent (W) dimensionality.
+ channels, # Number of output channels.
+ size, # Output spatial size: int or [width, height].
+ sampling_rate, # Output sampling rate.
+ bandwidth, # Output bandwidth.
+ margin_size, # Extra margin on input.
+ fov, # panorama FOV.
+ ):
+ super().__init__()
+ self.w_dim = w_dim
+ self.channels = channels
+ self.fov = fov
+ self.size = np.broadcast_to(np.asarray(size), [2])
+ self.sampling_rate = sampling_rate
+ self.bandwidth = bandwidth
+ self.margin_size = margin_size
+ self.frame_size = self.size - 2 * self.margin_size
+
+ # Draw random frequencies from uniform 2D disc.
+ freqs = torch.randn([self.channels, 2])
+ radii = freqs.square().sum(dim=1, keepdim=True).sqrt()
+ freqs /= radii * radii.square().exp().pow(0.25)
+ freqs *= bandwidth
+ phases = torch.rand([self.channels]) - 0.5
+
+ # Setup parameters and buffers.
+ self.weight = torch.nn.Parameter(torch.randn([self.channels, self.channels]))
+ self.affine = FullyConnectedLayer(w_dim, 4, weight_init=0, bias_init=[1,0,0,0])
+ self.register_buffer('transform', torch.eye(3, 3)) # User-specified inverse transform wrt. resulting image.
+ self.register_buffer('freqs', freqs)
+ self.register_buffer('phases', phases)
+
+ def forward(self, w, transform=None, crop_fn=None):
+ # Introduce batch dimension.
+ if transform is None:
+ transforms = self.transform.unsqueeze(0) # [batch, row, col]
+ else:
+ transforms = transform
+ freqs = self.freqs.unsqueeze(0) # [batch, channel, xy]
+ phases = self.phases.unsqueeze(0) # [batch, channel]
+
+ # does not add learned rotation for 360 model
+ transforms = transforms.expand(w.shape[0], -1, -1)
+
+ # Dampen out-of-band frequencies that may occur due to the user-specified transform.
+ amplitudes = (1 - (freqs.norm(dim=2) - self.bandwidth) / (self.sampling_rate / 2 - self.bandwidth)).clamp(0, 1)
+
+ # Construct sampling grid.
+ theta = torch.eye(2, 3, device=w.device)
+ theta[0, 0] = 0.5 * self.size[0] / self.sampling_rate # tx
+ theta[1, 1] = 0.5 * self.size[1] / self.sampling_rate # ty
+ grid_width = self.frame_size[0] * 360 // self.fov + 2 * self.margin_size
+ grids = torch.nn.functional.affine_grid(theta.unsqueeze(0),
+ [1, 1, self.size[1], grid_width],
+ align_corners=False)
+ # extended grid to ensure that the x coordinate completes a full circle without padding
+ base_width = grid_width - 2*self.margin_size
+ corrected_x = torch.arange(-self.margin_size, base_width*2+self.margin_size, device=grids.device) / base_width * 2 - 1
+ corrected_y = grids[0, :, 0, 1]
+ corrected_grids = torch.cat([corrected_x.view(1, 1, -1, 1).repeat(1, self.size[1], 1, 1),
+ corrected_y.view(1, -1, 1, 1).repeat(1, 1, grid_width+base_width, 1)], dim=3)
+ grids = corrected_grids
+
+ if crop_fn is None:
+ crop_start = random.randint(0, base_width - 1)
+ grids = grids[:, :, crop_start:crop_start+self.size[1], :]
+ else:
+ grids = crop_fn(grids)
+
+ # apply transformation first
+ rotation = transforms[:, :2, :2]
+ translation = transforms[:, :2, 2:].squeeze(2)
+ # normalize grid x s.t. transformations can operate on square affine ratio
+ grids_normalized = grids.clone()
+ min_bound = torch.min(grids_normalized[:, :, :, 0])
+ max_bound = torch.max(grids_normalized[:, :, :, 0])
+ target_range = torch.max(grids_normalized[:, :, :, 1]) - torch.min(grids_normalized[:, :, :, 1])
+ grids_normalized[:, :, :, 0] = (grids_normalized[:, :, :, 0] - min_bound) / (max_bound - min_bound)
+ grids_normalized[:, :, :, 0] = grids_normalized[:, :, :, 0] * target_range - target_range / 2
+ # # xT @ RT = (Rx)T --> it is transposed
+ grids_transformed = (grids_normalized.unsqueeze(3) @ rotation.permute(0, 2, 1).unsqueeze(1).unsqueeze(2)).squeeze(3)
+ grids_transformed = grids_transformed + translation.unsqueeze(1).unsqueeze(2)
+ grids_transformed[:, :, :, 0] = (grids_transformed[:, :, :, 0] + target_range / 2) / target_range * (max_bound - min_bound) + min_bound
+
+ # map discontinuous x-angle to continuous cylindrical coordinate
+ grids_transformed_sin = grids_transformed.clone()
+ grids_transformed_cos = grids_transformed.clone()
+ grids_transformed_sin[:, :, :, 0] = torch.sin(grids_transformed_sin[:, :, :, 0] * torch.tensor(math.pi))
+ grids_transformed_cos[:, :, :, 0] = torch.cos(grids_transformed_cos[:, :, :, 0] * torch.tensor(math.pi))
+
+ x_sin = (grids_transformed_sin.unsqueeze(3) @ freqs[:, :self.channels//2, :].permute(0, 2, 1).unsqueeze(1).unsqueeze(2)).squeeze(3) # [batch, height, width, channel]
+ x_sin = x_sin + phases[:, :self.channels//2].unsqueeze(1).unsqueeze(2)
+ x_sin = torch.sin(x_sin * (np.pi * 2))
+ x_sin = x_sin * amplitudes[:, :self.channels//2].unsqueeze(1).unsqueeze(2)
+ x_cos = (grids_transformed_cos.unsqueeze(3) @ freqs[:, self.channels//2:, :].permute(0, 2, 1).unsqueeze(1).unsqueeze(2)).squeeze(3) # [batch, height, width, channel]
+ x_cos = x_cos + phases[:, self.channels//2:].unsqueeze(1).unsqueeze(2)
+ x_cos = torch.sin(x_cos * (np.pi * 2))
+ x_cos = x_cos * amplitudes[:, self.channels//2:].unsqueeze(1).unsqueeze(2)
+ x = torch.cat([x_sin, x_cos], dim=-1)
+
+ # Apply trainable mapping.
+ weight = self.weight / np.sqrt(self.channels)
+ x = x @ weight.t()
+
+ # Ensure correct shape.
+ x = x.permute(0, 3, 1, 2) # [batch, channel, height, width]
+ misc.assert_shape(x, [w.shape[0], self.channels, int(self.size[1]), int(self.size[0])])
+ return x
+
+
#----------------------------------------------------------------------------
@persistence.persistent_class
@@ -276,6 +397,9 @@
use_radial_filters = False, # Use radially symmetric downsampling filter? Ignored for critically sampled layers.
conv_clamp = 256, # Clamp the output to [-X, +X], None = disable clamping.
magnitude_ema_beta = 0.999, # Decay rate for the moving average of input magnitudes.
+
+ # added
+ use_scale_affine = False
):
super().__init__()
self.w_dim = w_dim
@@ -326,7 +450,12 @@
pad_hi = pad_total - pad_lo
self.padding = [int(pad_lo[0]), int(pad_hi[0]), int(pad_lo[1]), int(pad_hi[1])]
- def forward(self, x, w, noise_mode='random', force_fp32=False, update_emas=False):
+ # added
+ self.use_scale_affine = use_scale_affine
+ if self.use_scale_affine:
+ self.scale_affine = FullyConnectedLayer(self.w_dim, self.in_channels, bias_init=0)
+
+ def forward(self, x, w, scale=None, noise_mode='random', force_fp32=False, update_emas=False):
assert noise_mode in ['random', 'const', 'none'] # unused
misc.assert_shape(x, [None, self.in_channels, int(self.in_size[1]), int(self.in_size[0])])
misc.assert_shape(w, [x.shape[0], self.w_dim])
@@ -340,6 +469,12 @@
# Execute affine layer.
styles = self.affine(w)
+ # added here
+ if self.use_scale_affine:
+ assert(scale is not None)
+ styles_scale = self.scale_affine(scale)
+ styles = styles + styles_scale # equivalent to concatenation
+
if self.is_torgb:
weight_gain = 1 / np.sqrt(self.in_channels * (self.conv_kernel ** 2))
styles = styles * weight_gain
@@ -411,6 +546,8 @@
margin_size = 10, # Number of additional pixels outside the image.
output_scale = 0.25, # Scale factor for the output image.
num_fp16_res = 4, # Use FP16 for the N highest resolutions.
+ training_mode = 'global', # training mode for input layer
+ fov = None, # Specify FOV for 360 model
**layer_kwargs, # Arguments for SynthesisLayer.
):
super().__init__()
@@ -440,9 +577,17 @@
channels[-1] = self.img_channels
# Construct layers.
+ if '360' not in training_mode:
self.input = SynthesisInput(
w_dim=self.w_dim, channels=int(channels[0]), size=int(sizes[0]),
- sampling_rate=sampling_rates[0], bandwidth=cutoffs[0])
+ sampling_rate=sampling_rates[0], bandwidth=cutoffs[0],
+ margin_size=margin_size)
+ else:
+ assert(fov is not None)
+ self.input = SynthesisInput360(
+ w_dim=self.w_dim, channels=int(channels[0]), size=int(sizes[0]),
+ sampling_rate=sampling_rates[0], bandwidth=cutoffs[0],
+ margin_size=margin_size, fov=fov)
self.layer_names = []
for idx in range(self.num_layers + 1):
prev = max(idx - 1, 0)
@@ -461,14 +606,19 @@
setattr(self, name, layer)
self.layer_names.append(name)
- def forward(self, ws, **layer_kwargs):
+ def forward(self, ws, mapped_scale=None, transform=None, crop_fn=None, **layer_kwargs):
misc.assert_shape(ws, [None, self.num_ws, self.w_dim])
ws = ws.to(torch.float32).unbind(dim=1)
+ if mapped_scale is not None:
+ scale = mapped_scale.to(torch.float32).unbind(dim=1)
+ else:
+ scale = [None] * self.num_ws
+ # ws is a list of ws for every layer
# Execute layers.
- x = self.input(ws[0])
- for name, w in zip(self.layer_names, ws[1:]):
- x = getattr(self, name)(x, w, **layer_kwargs)
+ x = self.input(ws[0], transform=transform, crop_fn=crop_fn)
+ for name, w , sc in zip(self.layer_names, ws[1:], scale[1:]):
+ x = getattr(self, name)(x, w, sc, **layer_kwargs)
if self.output_scale != 1:
x = x * self.output_scale
@@ -495,6 +645,8 @@
img_resolution, # Output resolution.
img_channels, # Number of output color channels.
mapping_kwargs = {}, # Arguments for MappingNetwork.
+ training_mode = 'global',
+ scale_mapping_kwargs = {}, # Arguments for Scale Mapping Network
**synthesis_kwargs, # Arguments for SynthesisNetwork.
):
super().__init__()
@@ -503,13 +655,65 @@
self.w_dim = w_dim
self.img_resolution = img_resolution
self.img_channels = img_channels
- self.synthesis = SynthesisNetwork(w_dim=w_dim, img_resolution=img_resolution, img_channels=img_channels, **synthesis_kwargs)
+ self.training_mode = training_mode
+ self.scale_mapping_kwargs = scale_mapping_kwargs
+ use_scale_affine = True if 'patch' in self.training_mode else False # add affine layer on style input
+ self.synthesis = SynthesisNetwork(w_dim=w_dim, img_resolution=img_resolution, img_channels=img_channels,
+ training_mode=training_mode, use_scale_affine=use_scale_affine,
+ **synthesis_kwargs)
self.num_ws = self.synthesis.num_ws
self.mapping = MappingNetwork(z_dim=z_dim, c_dim=c_dim, w_dim=w_dim, num_ws=self.num_ws, **mapping_kwargs)
+ if 'patch' in self.training_mode:
+ self.scale_mapping_kwargs = scale_mapping_kwargs
+ scale_mapping_norm = scale_mapping_kwargs.scale_mapping_norm
+ scale_mapping_min = scale_mapping_kwargs.scale_mapping_min
+ scale_mapping_max = scale_mapping_kwargs.scale_mapping_max
+ if scale_mapping_norm == 'zerocentered':
+ self.scale_norm = ScaleNormalizeZeroCentered(scale_mapping_min, scale_mapping_max)
+ scale_in_dim = 1
+ elif scale_mapping_norm == 'positive':
+ self.scale_norm = ScaleNormalizePositive(scale_mapping_min, scale_mapping_max)
+ scale_in_dim = 1
+ else:
+ assert(False)
+ self.scale_mapping = MappingNetwork(z_dim=scale_in_dim, c_dim=c_dim, w_dim=w_dim, num_ws=self.num_ws, **mapping_kwargs)
+
- def forward(self, z, c, truncation_psi=1, truncation_cutoff=None, update_emas=False, **synthesis_kwargs):
+ def forward(self, z, c, transform=None, truncation_psi=1, truncation_cutoff=None, update_emas=False, **synthesis_kwargs):
ws = self.mapping(z, c, truncation_psi=truncation_psi, truncation_cutoff=truncation_cutoff, update_emas=update_emas)
- img = self.synthesis(ws, update_emas=update_emas, **synthesis_kwargs)
+ if transform is None:
+ scale = torch.ones(z.shape[0], 1).to(z.device)
+ else:
+ scale = 1/transform[:, [0], 0]
+ if self.scale_mapping_kwargs:
+ scale = self.scale_norm(scale)
+ mapped_scale = self.scale_mapping(scale, c, update_emas=update_emas)
+ else:
+ mapped_scale = None
+ img = self.synthesis(ws, mapped_scale=mapped_scale, transform=transform, update_emas=update_emas, **synthesis_kwargs)
return img
#----------------------------------------------------------------------------
+@persistence.persistent_class
+class ScaleNormalizeZeroCentered(torch.nn.Module):
+ def __init__(self, scale_mapping_min, scale_mapping_max):
+ super().__init__()
+ self.scale_mapping_min = scale_mapping_min
+ self.scale_mapping_max = scale_mapping_max
+
+ def forward(self, scale):
+ # remaps scale to (-1, 1)
+ scale = (scale - self.scale_mapping_min) / (self.scale_mapping_max - self.scale_mapping_min)
+ return 2 * scale - 1
+
+@persistence.persistent_class
+class ScaleNormalizePositive(torch.nn.Module):
+ def __init__(self, scale_mapping_min, scale_mapping_max):
+ super().__init__()
+ self.scale_mapping_min = scale_mapping_min
+ self.scale_mapping_max = scale_mapping_max
+
+ def forward(self, scale):
+ # add a small offset to avoid zero point: [0.1 to 1.1]
+ scale = (scale - self.scale_mapping_min) / (self.scale_mapping_max - self.scale_mapping_min)
+ return scale + 0.1
Only in training: __pycache__
diff -bur ./stylegan3/training/training_loop.py training/training_loop.py
--- ./stylegan3/training/training_loop.py 2021-11-24 13:11:24.000000000 -0500
+++ training/training_loop.py 2022-05-17 18:02:20.264047277 -0400
@@ -26,6 +26,10 @@
import legacy
from metrics import metric_main
+from util import util
+import random
+from metrics import equivariance
+
#----------------------------------------------------------------------------
def setup_snapshot_image_grid(training_set, random_seed=0):
@@ -89,7 +93,8 @@
def training_loop(
run_dir = '.', # Output directory.
- training_set_kwargs = {}, # Options for training set.
+ training_set_kwargs = {}, # Options for base training set.
+ patch_kwargs = {}, # Options for patch dataset.
data_loader_kwargs = {}, # Options for torch.utils.data.DataLoader.
G_kwargs = {}, # Options for generator network.
D_kwargs = {}, # Options for discriminator network.
@@ -120,6 +125,7 @@
cudnn_benchmark = True, # Enable torch.backends.cudnn.benchmark?
abort_fn = None, # Callback function for determining whether to abort training. Must return consistent results across ranks.
progress_fn = None, # Callback function for updating training progress. Called for all ranks.
+ added_kwargs = {}, # added
):
# Initialize.
start_time = time.time()
@@ -132,12 +138,17 @@
conv2d_gradfix.enabled = True # Improves training speed.
grid_sample_gradfix.enabled = True # Avoids errors with the augmentation pipe.
+ # ADDED: to prevent data_loader pin_memory to load to device 0 for every process
+ torch.cuda.set_device(device)
+ training_mode = G_kwargs.training_mode
+
# Load training set.
if rank == 0:
print('Loading training set...')
training_set = dnnlib.util.construct_class_by_name(**training_set_kwargs) # subclass of training.dataset.Dataset
training_set_sampler = misc.InfiniteSampler(dataset=training_set, rank=rank, num_replicas=num_gpus, seed=random_seed)
training_set_iterator = iter(torch.utils.data.DataLoader(dataset=training_set, sampler=training_set_sampler, batch_size=batch_size//num_gpus, **data_loader_kwargs))
+
if rank == 0:
print()
print('Num images: ', len(training_set))
@@ -145,14 +156,56 @@
print('Label shape:', training_set.label_shape)
print()
+ # Load patch dataset
+ if 'patch' in training_mode:
+ if rank == 0:
+ print('Loading patch dataset...')
+ patch_dset = dnnlib.util.construct_class_by_name(**patch_kwargs) # subclass of training.dataset.Dataset
+ patch_dset_sampler = misc.InfiniteSampler(dataset=patch_dset, rank=rank, num_replicas=num_gpus, seed=random_seed)
+ patch_dset_iterator = iter(torch.utils.data.DataLoader(dataset=patch_dset, sampler=patch_dset_sampler, batch_size=batch_size//num_gpus, **data_loader_kwargs))
+ if rank == 0:
+ print()
+ print('Patch Num images: ', len(patch_dset))
+ print('Patch Image shape:', patch_dset.image_shape)
+ print('Patch Label shape:', patch_dset.label_shape)
+ print()
+
# Construct networks.
if rank == 0:
print('Constructing networks...')
- common_kwargs = dict(c_dim=training_set.label_dim, img_resolution=training_set.resolution, img_channels=training_set.num_channels)
+
+ # modified: use specified img_resolution
+ img_resolution = training_set.resolution
+ if 'patch' in training_mode and added_kwargs.img_size is not None:
+ img_resolution = added_kwargs.img_size
+ if rank == 0:
+ print("Using specified img resolution: %d" % img_resolution)
+ assert(added_kwargs.img_size == training_set.resolution)
+ common_kwargs = dict(c_dim=training_set.label_dim, img_resolution=img_resolution, img_channels=training_set.num_channels)
G = dnnlib.util.construct_class_by_name(**G_kwargs, **common_kwargs).train().requires_grad_(False).to(device) # subclass of torch.nn.Module
D = dnnlib.util.construct_class_by_name(**D_kwargs, **common_kwargs).train().requires_grad_(False).to(device) # subclass of torch.nn.Module
G_ema = copy.deepcopy(G).eval()
+
+ # copy G for teacher network: copy teacher G_ema to G_ema:,
+ # uses G state dict for the generator to align with D
+ if 'patch' in training_mode and added_kwargs.teacher is not None:
+ teacher = copy.deepcopy(G).to(device).eval()
+ # deactivate scale affine adding in teacher model; so it matches original model
+ for layer_name in teacher.synthesis.layer_names:
+ layer = getattr(teacher.synthesis, layer_name)