forked from mravanelli/pytorch-kaldi
-
Notifications
You must be signed in to change notification settings - Fork 0
/
train_gan.py
1175 lines (873 loc) · 43.2 KB
/
train_gan.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
##########################################################
# pytorch-kaldi-gan
# Walter Heymans
# North West University
# 2020
##########################################################
import sys
import configparser
import os
import time
import numpy
import numpy as np
import random
import torch
import torch.nn.functional as functional
from torch.optim.optimizer import Optimizer
import gan_networks
import itertools
from shutil import copyfile
import math
import matplotlib.pyplot as plt
import weights_and_biases as wandb
import importlib
import warnings
warnings.filterwarnings("ignore", '', UserWarning)
def print_version_info():
print("")
print("".center(40, "#"))
print(" Pytorch-Kaldi-GAN ".center(38, " ").center(40, "#"))
print(" Walter Heymans ".center(38, " ").center(40, "#"))
print(" North West University ".center(38, " ").center(40, "#"))
print(" 2020 ".center(38, " ").center(40, "#"))
print("".center(40, "#"), end="\n\n")
def save_tensor_list_to_png(array, titles=[], fig_name="tensor.png"):
plt.figure(figsize=(8, 6), dpi=300)
for i in range(1, len(array) + 1):
plt.subplot(len(array), 1, i)
if len(array) == 4 and i <= 2:
graph_colour = "b"
elif len(array) == 4:
graph_colour = "r"
elif i == 2:
graph_colour = "r"
else:
graph_colour = "b"
plt.plot(array[i - 1].detach().numpy(), graph_colour)
if len(titles) == len(array):
plt.title(titles[i - 1])
plt.tight_layout()
plt.savefig(fig_name)
plt.close()
def format_time(time_in_seconds):
hours_remaining = math.floor(time_in_seconds / 3600)
minutes_remaining = math.floor(time_in_seconds / 60) - (hours_remaining * 60)
seconds_remaining = math.floor(time_in_seconds) - (minutes_remaining * 60) - (hours_remaining * 3600)
if hours_remaining > 0:
return "{}h {}m {}s ".format(hours_remaining, minutes_remaining, seconds_remaining)
elif minutes_remaining > 0:
return "{}m {}s ".format(minutes_remaining, seconds_remaining)
else:
return "{}s ".format(seconds_remaining)
def get_labels(bs, label):
return torch.ones((bs, 1)) * label
def get_pearson_correlation(tensor1, tensor2):
from scipy.stats import pearsonr
output1 = tensor1.detach().cpu().numpy()
output2 = tensor2.detach().cpu().numpy()
if output1.shape == output2.shape:
# calculate Pearson's correlation
if len(output1.shape) > 1:
correlation = 0
for i in range(output1.shape[0]):
try:
temp_corr, _ = pearsonr(output1[i], output2[i])
except:
temp_corr = 0
correlation += temp_corr
if output1.shape[0] > 0:
correlation = correlation / output1.shape[0]
else:
correlation, _ = pearsonr(output1, output2)
return correlation
else:
return 0
def get_mean_squared_error(tensor1, tensor2):
output1 = tensor1.detach().cpu()
output2 = tensor2.detach().cpu()
if output1.shape == output2.shape:
if len(output1.shape) > 1:
error = 0
for i in range(output1.shape[0]):
error += torch.mean(torch.abs(torch.abs(output1) - torch.abs(output2)))
if output1.shape[0] > 0:
error = error / output1.shape[0]
else:
error = torch.mean(torch.abs(torch.abs(output1) - torch.abs(output2)))
return error.numpy()
else:
return 0
def get_g_performance(clean, noisy, generated):
''' Performance metric using Pearson correlation, mean squared error and L1 loss.
Metric is comparing generator relative to noisy signal.
Higher is better. '''
l1_loss_noisy = torch.nn.functional.l1_loss(clean, noisy).item()
l1_loss_gen = torch.nn.functional.l1_loss(clean, generated).item()
r_clean_noisy = get_pearson_correlation(clean, noisy)
r_clean_gen = get_pearson_correlation(clean, generated)
mse_clean_noisy = get_mean_squared_error(clean, noisy)
mse_clean_gen = get_mean_squared_error(clean, generated)
l1_performance = l1_loss_noisy - l1_loss_gen
r_performance = r_clean_gen - r_clean_noisy
mse_performance = mse_clean_noisy - mse_clean_gen
performance_metric = r_performance + mse_performance + l1_performance
return performance_metric
def compute_gradient_penalty(D, real_samples, fake_samples):
Tensor = torch.cuda.FloatTensor
from torch.autograd import Variable
"""Calculates the gradient penalty loss for WGAN GP"""
# Random weight term for interpolation between real and fake samples
alpha = Tensor(np.random.random((real_samples.size(0), 440)))
# Get random interpolation between real and fake samples
interpolates = (alpha * real_samples + ((1 - alpha) * fake_samples)).requires_grad_(True)
d_interpolates = D(interpolates)
fake = Variable(Tensor(real_samples.shape[0], 1).fill_(1.0), requires_grad=False)
# Get gradient w.r.t. interpolates
gradients = torch.autograd.grad(
outputs=d_interpolates,
inputs=interpolates,
grad_outputs=fake,
create_graph=True,
retain_graph=True,
only_inputs=True,
)[0]
gradients = gradients.view(gradients.size(0), -1)
gradient_penalty = ((gradients.norm(2, dim=1) - 1) ** 2).mean()
return gradient_penalty
script_start_time = time.time()
print_version_info()
# Reading global cfg file (first argument-mandatory file)
cfg_file = sys.argv[1]
if not (os.path.exists(cfg_file)):
sys.stderr.write("ERROR: The config file %s does not exist!\n" % (cfg_file))
sys.exit(0)
else:
config = configparser.ConfigParser()
config.read(cfg_file)
# Output folder creation
out_folder = config["exp"]["out_folder"]
if not os.path.exists(out_folder):
os.makedirs(out_folder)
# Copy the global cfg file into the output folder
cfg_file = out_folder + "/conf.cfg"
with open(cfg_file, "w") as configfile:
config.write(configfile)
# Read hyper-parameters from config file
seed = int(config['hyperparameters']['seed'])
max_epochs = int(config['hyperparameters']['max_epochs'])
batch_size = int(config['hyperparameters']['batch_size'])
lr_g = float(config['hyperparameters']['lr_g'])
lr_d = float(config['hyperparameters']['lr_d'])
try:
d_updates = int(config['hyperparameters']['d_updates'])
except KeyError:
d_updates = 1
real_label = float(config['hyperparameters']['real_label'])
criterion = str(config['hyperparameters']['criterion'])
optimizer = str(config['hyperparameters']['optimizer'])
cycle_consistency_lambda = int(config['hyperparameters']['cycle_consistency_lambda'])
acoustic_model_lambda = float(config['hyperparameters']['acoustic_model_lambda'])
gp_lambda = int(config['hyperparameters']['gp_lambda'])
try:
l1_lambda = int(config['hyperparameters']['l1_lambda'])
l2_lambda = int(config['hyperparameters']['l2_lambda'])
except KeyError:
pass
if config.getboolean("exp", "use_cuda"):
try:
cuda_device = int(config['exp']['cuda_device'])
except ValueError:
cuda_device = 'cpu'
else:
cuda_device = 'cpu'
if config["wandb"]["wandb"] == "True":
wandb_on = True
else:
wandb_on = False
torch.manual_seed(seed = seed)
random.seed(seed)
clean_dataset_path = str(config['datasets']['clean_dataset'])
noisy_dataset_path = str(config['datasets']['noisy_dataset'])
valid_dataset_path = str(config['datasets']['valid_dataset'])
cw_left = int(config['datasets']['cw_left'])
cw_right = int(config['datasets']['cw_right'])
frames_per_sample = cw_left + cw_right + 1
double_features = False
try:
if config["hyperparameters"]["double_features"] == "True":
double_features = True
except KeyError:
pass
early_stopping = False
try:
if config["hyperparameters"]["early_stopping"] == "True":
early_stopping = True
except KeyError:
pass
train_d_with_noisy = False
try:
if config["hyperparameters"]["train_d_with_noisy"] == "True":
train_d_with_noisy = True
except KeyError:
pass
print("@ Progress: Reading config complete\n")
def print_settings():
print_width = 64
print(" Hyper-parameters ".center(print_width, "="))
print("# Seed:\t\t\t", seed)
print("# Epochs:\t\t", max_epochs)
print("# Batch size:\t\t", batch_size)
print("# Learning rate G:\t", lr_g)
print("# Learning rate D:\t", lr_d)
print("# Acoustic model lambda:", acoustic_model_lambda)
print("# Gradient penalty lambda:", gp_lambda)
print("# Real label:\t\t", real_label)
print("# Criterion:\t\t", criterion)
print("# Optimizer:\t\t", optimizer)
print("# Cuda device:\t\t", cuda_device)
print("# Weights and Biases:\t", wandb_on)
print("# Output directory:\t", out_folder)
print("# Double features:\t", double_features)
print("# Early stopping:\t", early_stopping)
print("=".center(print_width, "="), end = "\n\n")
print_settings()
class Dataset(torch.utils.data.Dataset):
'Characterizes a dataset for PyTorch'
def __init__(self, dataset_path_clean, dataset_path_noisy, chunk):
self.validation_set = False
if not os.path.exists(dataset_path_noisy) or dataset_path_noisy == "":
self.validation_set = True
'Initialization'
self.dataset_path_clean = dataset_path_clean
files = sorted(os.listdir(self.dataset_path_clean))
if chunk <= len(files):
self.dataset_object = torch.load(os.path.join(self.dataset_path_clean, ("chunk_" + str(chunk) + ".pt")), map_location = 'cpu')
clean_length = int(math.floor(self.dataset_object.shape[0] / frames_per_sample))
if not self.validation_set:
# Noisy dataset
self.dataset_path_noisy = dataset_path_noisy
files = sorted(os.listdir(self.dataset_path_noisy))
if chunk <= len(files):
self.dataset_object_noisy = torch.load(os.path.join(self.dataset_path_noisy, ("chunk_" + str(chunk) + ".pt")), map_location = 'cpu')
noisy_lenght = int(math.floor(self.dataset_object_noisy.shape[0] / frames_per_sample))
self.dataset_len = min([clean_length, noisy_lenght])
else:
self.dataset_len = clean_length
def __len__(self):
'Denotes the total number of samples'
return self.dataset_len
def __getitem__(self, index):
'Generates one sample of data'
for frame in range(frames_per_sample):
label = self.dataset_object[index,-1]
if frame == 0:
clean = self.dataset_object[index + frame, :40]
else:
clean = torch.cat((clean, self.dataset_object[index + frame, :40]), dim = 0)
if not self.validation_set:
for frame in range(frames_per_sample):
label_noisy = self.dataset_object_noisy[index, -1]
if frame == 0:
noisy = self.dataset_object_noisy[index + frame, :40]
else:
noisy = torch.cat((noisy, self.dataset_object_noisy[index + frame, :40]), dim = 0)
return clean, noisy, label, label_noisy
else:
return clean, label
def getbatch(self, index, batch_size):
clean, noisy, _, _ = self.__getitem__(index)
clean = torch.unsqueeze(clean, dim = 0)
noisy = torch.unsqueeze(noisy, dim = 0)
for bs in range(batch_size-1):
tempclean, tempnoisy, _, _ = self.__getitem__(index+bs+1)
tempclean = torch.unsqueeze(tempclean, dim = 0)
tempnoisy = torch.unsqueeze(tempnoisy, dim = 0)
clean = torch.cat((clean, tempclean), dim = 0)
noisy = torch.cat((noisy, tempnoisy), dim = 0)
return clean, noisy
number_of_chunks = len(os.listdir(clean_dataset_path))
train_set = Dataset(clean_dataset_path, noisy_dataset_path, 1)
train_loader = torch.utils.data.DataLoader(train_set,
batch_size = batch_size,
shuffle = True,
num_workers = 4)
validation_set = Dataset(valid_dataset_path, "", 1)
valid_loader = torch.utils.data.DataLoader(validation_set,
batch_size = batch_size,
shuffle = True,
num_workers = 4)
print("@ Progress: Dataset loaded")
if cuda_device != 'cpu':
torch.cuda.set_device(cuda_device)
print("@ Progress: Cuda device set to", cuda_device)
# Create acoustic model
acoustic_model_path = str(config["acoustic_model"]["pretrained_file"])
train_with_am = False
use_external_model = False
try:
if str(config["acoustic_model"]["use_external_model"]) == "True":
use_external_model = True
except KeyError:
pass
if use_external_model:
if os.path.exists(acoustic_model_path):
def get_number_hidden_layers(dictionary_keys):
layer_count = 0
for key in dictionary_keys:
if 'wx' in key:
layer_count += 1
layer_count /= 2
return int(layer_count)
def get_n_out_dim(dictionary):
num_layers = get_number_hidden_layers(dictionary.keys())
last_layer_key = 'wx.' + str(num_layers - 1) + '.weight'
for key in dictionary.keys():
if last_layer_key == key:
return dictionary[key].shape[0]
return 0
try:
if cuda_device != 'cpu':
if int(config["exp"]["cuda_device"]) == 0:
checkpoint_load = torch.load(acoustic_model_path, map_location="cuda:0")
elif int(config["exp"]["cuda_device"]) == 1:
checkpoint_load = torch.load(acoustic_model_path, map_location="cuda:1")
else:
checkpoint_load = torch.load(acoustic_model_path, map_location="cpu")
N_out_lab_cd = get_n_out_dim(checkpoint_load["model_par"])
# import the class
module = importlib.import_module(config["acoustic_model"]["arch_library"])
nn_class = getattr(module, config["acoustic_model"]["arch_class"])
config["acoustic_model"]["dnn_lay"] = config["acoustic_model"]["dnn_lay"].replace('N_out_lab_cd', str(N_out_lab_cd))
if double_features:
acoustic_model = nn_class(config["acoustic_model"], int(2 * frames_per_sample * 40))
else:
acoustic_model = nn_class(config["acoustic_model"], int(frames_per_sample * 40))
acoustic_model.load_state_dict(checkpoint_load["model_par"])
acoustic_model = acoustic_model.cuda()
except RuntimeError:
print("Error loading acoustic model! Check that models in config file match.")
else:
if os.path.exists(acoustic_model_path):
def get_number_hidden_layers(dictionary_keys):
layer_count = 0
for key in dictionary_keys:
if 'wx' in key:
layer_count += 1
layer_count /= 2
return int(layer_count)
def get_n_out_dim(dictionary):
num_layers = get_number_hidden_layers(dictionary.keys())
last_layer_key = 'wx.' + str(num_layers - 1) + '.weight'
for key in dictionary.keys():
if last_layer_key == key:
return dictionary[key].shape[0]
return 0
try:
if cuda_device != 'cpu':
if int(config["exp"]["cuda_device"]) == 0:
checkpoint_load = torch.load(acoustic_model_path, map_location="cuda:0")
elif int(config["exp"]["cuda_device"]) == 1:
checkpoint_load = torch.load(acoustic_model_path, map_location="cuda:1")
else:
checkpoint_load = torch.load(acoustic_model_path, map_location="cpu")
N_out_lab_cd = get_n_out_dim(checkpoint_load["model_par"])
# import the class
module = importlib.import_module(config["acoustic_model"]["arch_library"])
nn_class = getattr(module, config["acoustic_model"]["arch_class"])
config["acoustic_model"]["dnn_lay"] = config["acoustic_model"]["dnn_lay"].replace('N_out_lab_cd', str(N_out_lab_cd))
if double_features:
acoustic_model = nn_class(config["acoustic_model"], int(2 * frames_per_sample * 40))
else:
acoustic_model = nn_class(config["acoustic_model"], int(frames_per_sample * 40))
acoustic_model.load_state_dict(checkpoint_load["model_par"])
acoustic_model = acoustic_model.cuda()
train_with_am = True
except RuntimeError:
print("Error loading acoustic model! Check that models in config file match.")
else:
print("Acoustic model path doesnt exist!")
# Create networks and optimizers
# Create Generator
input_dim = train_set.__getitem__(0)[0].shape[0]
generator_class = getattr(gan_networks, config["generator"]["arch_name"])
generator = generator_class(input_dim,
input_dim,
config["generator"])
if config["hyperparameters"]["criterion"] == "cycle":
generator_f = generator_class(input_dim,
input_dim,
config["generator"])
# Create Discriminator
discriminator_class = getattr(gan_networks, config["discriminator"]["arch_name"])
discriminator = discriminator_class(input_dim, config["discriminator"])
if config["hyperparameters"]["criterion"] == "cycle":
discriminator_h = discriminator_class(input_dim, config["discriminator"])
generator = generator.cuda()
discriminator = discriminator.cuda()
if config["hyperparameters"]["criterion"] == "cycle":
generator_f = generator_f.cuda()
discriminator_h = discriminator_h.cuda()
# Creating directories
directory_g = os.path.join(out_folder, config["gan"]["output_path_g"])
directory_d = os.path.join(out_folder, config["gan"]["output_path_d"])
gan_dir = os.path.dirname(directory_g)
if not os.path.exists(gan_dir):
os.mkdir(gan_dir)
if not os.path.exists(gan_dir + "/images"):
os.mkdir(gan_dir + "/images")
# Copy pretrained models into directory if it is set
try:
if str(config["generator"]["pretrained_file"]) != "none":
if os.path.exists(str(config["generator"]["pretrained_file"])):
copyfile(str(config["generator"]["pretrained_file"]), directory_g)
print("Loaded pretrained G.")
except KeyError:
pass
try:
if str(config["discriminator"]["pretrained_file"]) != "none":
if os.path.exists(str(config["discriminator"]["pretrained_file"])):
copyfile(str(config["discriminator"]["pretrained_file"]), directory_d)
print("Loaded pretrained D.")
except KeyError:
pass
# Load pretrained models
if os.path.exists(directory_g):
try:
generator.load_state_dict(torch.load(directory_g))
if criterion == "cycle":
generator_f.load_state_dict(torch.load(os.path.dirname(directory_g) + "/generator_f.pt"))
except RuntimeError:
print("Load error loading G, network will be recreated.")
if os.path.exists(directory_d):
try:
discriminator.load_state_dict(torch.load(directory_d))
if criterion == "cycle":
discriminator_h.load_state_dict(torch.load(os.path.dirname(directory_d) + "/discriminator_h.pt"))
except RuntimeError:
print("Load error loading D, network will be recreated.")
# Optimizer initialization
if config["hyperparameters"]["optimizer"] == "adam":
if criterion == "cycle":
optimizer_g = torch.optim.Adam(itertools.chain(generator.parameters(), generator_f.parameters()), lr = lr_g)
optimizer_d = torch.optim.Adam(discriminator.parameters(), lr = lr_d)
optimizer_h = torch.optim.Adam(discriminator_h.parameters(), lr = lr_d)
else:
optimizer_g = torch.optim.Adam(generator.parameters(), lr = lr_g, betas = (0.5, 0.999), weight_decay = 0.001)
optimizer_d = torch.optim.Adam(discriminator.parameters(), lr = lr_d, betas = (0.5, 0.999), weight_decay = 0.001)
elif config["hyperparameters"]["optimizer"] == "rmsprop":
if criterion == "cycle":
optimizer_g = torch.optim.RMSprop(itertools.chain(generator.parameters(), generator_f.parameters()), lr = lr_g)
optimizer_d = torch.optim.RMSprop(discriminator.parameters(), lr = lr_d)
optimizer_h = torch.optim.RMSprop(discriminator_h.parameters(), lr = lr_d)
else:
optimizer_g = torch.optim.RMSprop(generator.parameters(), lr = lr_g)
optimizer_d = torch.optim.RMSprop(discriminator.parameters(), lr = lr_d)
elif config["hyperparameters"]["optimizer"] == "sgd":
if criterion == "cycle":
optimizer_g = torch.optim.SGD(itertools.chain(generator.parameters(), generator_f.parameters()), lr = lr_g)
optimizer_d = torch.optim.SGD(discriminator.parameters(), lr = lr_d)
optimizer_h = torch.optim.SGD(discriminator_h.parameters(), lr = lr_d)
else:
optimizer_g = torch.optim.SGD(generator.parameters(), lr = lr_g)
optimizer_d = torch.optim.SGD(discriminator.parameters(), lr = lr_d)
# Start training
print("\n@ Progress: Starting training")
train_start_time = time.time()
number_of_batches = len(train_loader)
if str(config["wandb"]["wandb"]) == "True":
wandb_cfg = wandb.load_cfg_dict_from_yaml(str(config["wandb"]["config"]))
# UPDATE config file if Weights and Biases file is different
wandb_cfg["max_epochs"] = max_epochs
wandb_cfg["seed"] = seed
wandb_cfg["batch_size"] = batch_size
wandb_cfg["lr_g"] = lr_g
wandb_cfg["lr_d"] = lr_d
wandb_cfg["criterion"] = criterion
wandb_cfg["optimizer"] = optimizer
wandb_cfg["generator"] = str(config["generator"]["arch_name"])
wandb_cfg["discriminator"] = str(config["discriminator"]["arch_name"])
wandb_cfg["dataset"] = str(config["exp"]["dataset_name"])
wandb_cfg["acoustic_model_lambda"] = acoustic_model_lambda
wandb_cfg["cycle_consistency_lambda"] = cycle_consistency_lambda
wandb_cfg["gp_lambda"] = gp_lambda
wandb_details = os.path.join(out_folder, "wandb_details.txt")
if not os.path.exists(wandb_details):
wandb_details_file = open(wandb_details, "w")
wandb.initialize_wandb(project = str(config["wandb"]["project"]),
config = wandb_cfg,
directory = out_folder,
resume = False)
try:
wandb_details_file.write(wandb.get_run_id() + '\n')
wandb_details_file.write(wandb.get_run_name())
except TypeError:
pass
wandb_details_file.close()
else:
wandb_details_file = open(wandb_details, "r")
try:
file_content = wandb_details_file.read().splitlines()
wandb_run_id = file_content[0]
wandb_run_name = file_content[1]
except IndexError:
pass
wandb_details_file.close()
try:
wandb.initialize_wandb(project = str(config["wandb"]["project"]),
config = wandb_cfg,
directory = out_folder,
resume = True,
identity = wandb_run_id,
name = wandb_run_name)
except NameError:
wandb.initialize_wandb(project = str(config["wandb"]["project"]),
config = wandb_cfg,
directory = out_folder,
resume = True)
def create_log_file():
if not os.path.exists(os.path.join(out_folder, 'log.log')):
log_file = open(os.path.join(out_folder, 'log.log'), "w")
log_file.close()
def update_log_file(text_str):
log_file = open(os.path.join(out_folder, 'log.log'), "a")
log_file.write(text_str + "\n")
log_file.close()
def get_last_trained_epoch():
log_file = open(os.path.join(out_folder, 'log.log'), "r")
file_lines = log_file.readlines()
log_file.close()
if len(file_lines) > 0:
epoch_last, chunk_last = (file_lines[-1].replace("epoch_", "")).split("_")
return int(epoch_last), int(chunk_last)
else:
return 0, 0
def validate_generator_results():
with torch.no_grad():
number_of_valid_batches = len(valid_loader)
validation_loss = 0
correct = 0
total_samples = 0
for valid_batch, valid_label_batch in valid_loader:
valid_batch = valid_batch.cuda()
valid_label_batch = valid_label_batch.cuda()
if criterion == "am-gan":
gen_output, _ = generator(valid_batch)
else:
gen_output = generator(valid_batch)
if g_output.shape[0] > 1:
if double_features:
am_evaluation = acoustic_model(torch.cat((valid_batch, gen_output), dim = 1))
else:
am_evaluation = acoustic_model(gen_output)
validation_loss += functional.nll_loss(am_evaluation, valid_label_batch.long()).item()
pred = am_evaluation.data.max(1, keepdim = True)[1]
correct += torch.sum(pred.eq(valid_label_batch.data.view_as(pred))).item()
total_samples += valid_label_batch.shape[0]
validation_loss = validation_loss / number_of_valid_batches
validation_error = 1 - (correct / total_samples)
return validation_loss, validation_error
def check_discriminator_classification():
with torch.no_grad():
v_set = Dataset(clean_dataset_path, noisy_dataset_path, chunk)
v_loader = torch.utils.data.DataLoader(v_set,
batch_size=batch_size,
shuffle=True,
num_workers=4)
nob = len(v_loader)
validation_loss = 0
correct = 0
total_samples = 0
for v_clean_batch, v_noisy_batch, _, _ in v_loader:
nob += 1
v_clean_batch, v_noisy_batch = v_clean_batch.cuda(), v_noisy_batch.cuda()
v_input = torch.cat((v_clean_batch, v_noisy_batch), dim=0)
v_target = torch.cat((torch.ones(v_clean_batch.shape[0]).long(), torch.zeros(v_noisy_batch.shape[0]).long()), dim=0).to(cuda_device)
v_output = discriminator(v_input)
validation_loss += functional.cross_entropy(v_output, v_target).item()
pred = v_output.data.max(1, keepdim=True)[1]
correct += torch.sum(pred.eq(v_target.data.view_as(pred))).item()
total_samples += v_target.shape[0]
validation_loss = validation_loss / nob
validation_error = 1 - (correct / total_samples)
return validation_loss, validation_error
create_log_file()
if wandb_on:
wandb.quick_log("status", "training", commit = False)
epochs_skipped = 0
lowest_valid_error = 1
early_stopping_epoch = 0
file_loss = open(os.path.join(out_folder, "losses"), "w")
file_loss.close()
for epoch in range(1, max_epochs+1):
# Check if epoch has been processed
last_ep, last_ch = get_last_trained_epoch()
if (epoch < last_ep) or (last_ep == epoch and last_ch == number_of_chunks):
print("")
print(" Previously completed epoch: {} ".format(epoch).center(64, "#"))
epochs_skipped += 1
continue
if wandb_on:
wandb.quick_log("epoch", epoch, commit = False)
# Training
epoch_start_time = time.time()
print("")
print(" Optimizing epoch: {}/{} ".format(epoch, max_epochs).center(64, "#"))
for chunk in range(1, number_of_chunks + 1):
# Check if chunk has been processed
if (last_ep == epoch) and (chunk <= last_ch):
continue
if wandb_on:
wandb.quick_log("chunk", chunk, commit = True)
current_batch = 0
g_loss = 0
d_loss = 0
tot_am_loss = 0
train_set = Dataset(clean_dataset_path, noisy_dataset_path, chunk)
train_loader = torch.utils.data.DataLoader(train_set,
batch_size = batch_size,
shuffle = True,
num_workers = 6)
number_of_batches = len(train_loader)
print(" Chunk: {}/{} ".format(chunk, number_of_chunks).center(64, "-"))
for clean_batch, noisy_batch, label_batch, label_noisy_batch in train_loader:
current_batch += 1
# Transfer to GPU
clean_batch, noisy_batch = clean_batch.cuda(), noisy_batch.cuda()
label_batch, label_noisy_batch = label_batch.cuda(), label_noisy_batch.cuda()
d_clean_batch = clean_batch
d_noisy_batch = noisy_batch
if d_clean_batch.shape[0] > 1 and d_noisy_batch.shape[0] > 1:
for k in range(d_updates):
# TRAIN DISCRIMINATOR
optimizer_d.zero_grad()
d_output_clean = discriminator(d_clean_batch)
if criterion == "am-gan":
g_output, am_gan_output = generator(d_noisy_batch)
g_output = g_output.detach()
else:
g_output = generator(d_noisy_batch).detach()
d_output_g = discriminator(g_output)
real_labels = get_labels(d_clean_batch.shape[0], real_label).cuda()
fake_labels = get_labels(d_clean_batch.shape[0], 0).cuda()
if criterion == "bce" or criterion == "bce-l1" or criterion == "bce-l2" or criterion == "bce-all" or criterion == "am-gan":
loss_clean = functional.binary_cross_entropy(d_output_clean, real_labels)
loss_noisy = functional.binary_cross_entropy(d_output_g, fake_labels)
loss_discriminator = loss_clean + loss_noisy
loss_discriminator.backward()
optimizer_d.step()
d_loss += loss_discriminator.item()
file_loss = open(os.path.join(out_folder, "losses"), "a")
file_loss.write(str(epoch) + "," + str(chunk) + "," + str(loss_discriminator.item()) + ",")
file_loss.close()
elif criterion == "wgan":
if gp_lambda > 0:
gp_loss = compute_gradient_penalty(discriminator, d_clean_batch, g_output)
else:
gp_loss = 0
if train_d_with_noisy:
loss_discriminator = - torch.mean(d_output_clean) + torch.mean(d_output_g) + (0.1*torch.mean(discriminator(d_noisy_batch))) + (gp_lambda * gp_loss)
else:
loss_discriminator = - torch.mean(d_output_clean) + torch.mean(d_output_g) + (gp_lambda * gp_loss)
loss_discriminator.backward()
optimizer_d.step()
d_loss += loss_discriminator.item()
temp_d_loss = - torch.mean(d_output_clean) + torch.mean(d_output_g)
if gp_lambda > 0:
file_loss = open(os.path.join(out_folder, "losses"), "a")
file_loss.write(str(epoch) + "," + str(chunk) + "," + str(temp_d_loss.item()) + "," + str(gp_loss.item()) + ",")
file_loss.close()
elif criterion == "cycle":
optimizer_h.zero_grad()
h_output_noisy = discriminator_h(d_noisy_batch) # H_f output of noisy signal
h_output_f = discriminator_h(generator_f(d_clean_batch)) # H_f output of F
criterion_GAN = torch.nn.MSELoss()
criterion_GAN = criterion_GAN.cuda()
# TRAIN Discriminator D
# Real loss
loss_real = criterion_GAN(d_output_clean, real_labels)
# Fake loss
loss_fake = criterion_GAN(d_output_g, fake_labels)
# Total loss
loss_d = loss_real + loss_fake
loss_d.backward()
optimizer_d.step()
# TRAIN Discriminator H
# Real loss
loss_real = criterion_GAN(h_output_noisy, real_labels)
# Fake loss
loss_fake = criterion_GAN(h_output_f, fake_labels)
# Total loss
loss_h = loss_real + loss_fake
loss_h.backward()
optimizer_h.step()
d_loss += (loss_d.item() + loss_h.item()) / 2
if k < (d_updates - 1):
d_clean_batch, d_noisy_batch = train_set.getbatch(random.randint(0, train_set.__len__() - batch_size - 1), batch_size)
d_clean_batch = d_clean_batch.to(cuda_device)
d_noisy_batch = d_noisy_batch.to(cuda_device)
# TRAIN GENERATOR
optimizer_g.zero_grad()
if criterion == "am-gan":
g_output, am_gan_output = generator(noisy_batch)
else:
g_output = generator(noisy_batch)
d_verdict = discriminator(g_output)
am_loss = 0
if train_with_am:
if g_output.shape[0] > 1:
if double_features:
am_output = acoustic_model(torch.cat((noisy_batch, g_output), dim = 1))
else:
am_output = acoustic_model(g_output)
am_loss = functional.nll_loss(am_output, label_noisy_batch.long())
f = open(os.path.join(out_folder, "am_loss.txt"), 'a')
f.writelines(str(am_loss))
f.close()
tot_am_loss += am_loss.item()
else:
am_loss = 0
elif use_external_model:
if g_output.shape[0] > 1:
am_output = acoustic_model(g_output)
numpy_output = am_output.detach().cpu().numpy()
imported_am_output = torch.from_numpy(numpy_output).cuda()
am_loss = functional.nll_loss(imported_am_output, label_noisy_batch.long())
f = open(os.path.join(out_folder, "am_loss.txt"), 'a')
f.writelines(str(am_loss))
f.close()
tot_am_loss += am_loss.item()
else:
am_loss = 0
if criterion == "bce":
gen_labels = get_labels(clean_batch.shape[0], real_label).cuda()
bce_loss = functional.binary_cross_entropy(d_verdict, gen_labels)
loss_generator = bce_loss + (acoustic_model_lambda * am_loss)
loss_generator.backward()
optimizer_g.step()
if am_loss > 0:
g_loss += loss_generator.item() - (acoustic_model_lambda * am_loss.item())
file_loss = open(os.path.join(out_folder, "losses"), "a")