-
Notifications
You must be signed in to change notification settings - Fork 5
/
Copy pathevaluate_experiment_pcgan.py
417 lines (326 loc) · 19.9 KB
/
evaluate_experiment_pcgan.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
import dataset
import librosa
from torch.utils.data import DataLoader, random_split
import torch
import torch.nn.functional as F
from utils import *
import torchvision.models as models
from torch.optim.lr_scheduler import StepLR
from torch.utils.tensorboard import SummaryWriter
import argparse
import yaml
from pathlib import Path
import pandas as pd
import time
#from models import *
#from filter import *
from networks import *
from torch.autograd import Variable
import glob
from mel2wav.modules import MelGAN_Generator, Audio2Mel
LongTensor = torch.cuda.LongTensor
FloatTensor = torch.cuda.FloatTensor
def parse_args():
parser = argparse.ArgumentParser()
parser.add_argument("--device", type = str, default = '0')
parser.add_argument("--experiment_path", type = str, default = None)
parser.add_argument("--save_specs", type = bool, default = False)
parser.add_argument("--save_audio", type = bool, default = False)
parser.add_argument("--filter_receptive_field", type = int, default = 3)
parser.add_argument("--n_mel_channels", type = int, default = 80)
parser.add_argument("--ngf", type = int, default = 32)
parser.add_argument("--n_residual_layers", type = int, default = 3)
# Data parameters
parser.add_argument("--sampling_rate", type = int, default = 8000)
parser.add_argument("--segment_length", type = int, default = 8192)
parser.add_argument("--batch_size", type = int, default = 8)
args = parser.parse_args()
return args
def main():
args = parse_args()
runs = sorted(listdirs(args.experiment_path))
num_runs = len(runs)
# Some hyper parameters
num_genders = 2
num_digits = 10
device = 'cuda:' + args.device
print(device)
# Meta data and list of data files
annotation_file = '/home/edvinli/thesis_spring_2020/audio_mnist/audioMNIST_meta.json'
train_file_index = librosa.util.find_files('/home/edvinli/thesis_spring_2020/audio_mnist/')
split_ratio = 5
# Build indices for the data
file_index, annotation_index_gender, annotation_index_digit, annotation_index_speaker_id = dataset.build_annotation_index(
train_file_index, annotation_file, balanced_genders = False)
test_annotation_index, train_annotation_index, test_ids, train_ids = dataset.balanced_annotation_split(file_index, annotation_index_gender, annotation_index_digit, annotation_index_speaker_id, split_ratio)
# Create the dataset
test_data = dataset.AnnotatedAudioDataset(
test_annotation_index, args.sampling_rate, args.segment_length
)
n_test = test_data.__len__()
if args.save_audio:
test_loader = DataLoader(test_data, batch_size = 25, num_workers = 1)
else:
test_loader = DataLoader(test_data, batch_size = args.batch_size, num_workers = 1)
# Set up models that are not trained
fft = Audio2Mel(sampling_rate = args.sampling_rate)
Mel2Audio = MelGAN_Generator(args.n_mel_channels, args.ngf, args.n_residual_layers).to(device)
utility_netD = load_modified_AlexNet(num_digits).to(device)
fixed_netD = load_modified_AlexNet(num_genders).to(device)
utility_audio_net = AudioNet(num_digits).to(device)
privacy_audio_net = AudioNet(num_genders).to(device)
spec_FID_net = FID_AlexNet(num_digits).to(device)
audio_FID_net = AudioNet(num_digits).to(device)
# Pretrained Mel spectrogram inversion and digit classification
Mel2Audio.load_state_dict(torch.load('mel2wav/best_netG_epoch_2120.pt'))
utility_netD.load_state_dict(torch.load('fixed_classifier_checkpoints/best_digit_alexnet_spectrograms_epoch_26.pt'))
fixed_netD.load_state_dict(torch.load('fixed_classifier_checkpoints/best_gender_alexnet_epoch_29.pt'))
utility_audio_net.load_state_dict(torch.load('fixed_classifier_checkpoints/audio_digit_net_early_stop_epoch_26.pt'))
privacy_audio_net.load_state_dict(torch.load('fixed_classifier_checkpoints/audio_gender_net_early_stop_epoch_36.pt'))
utility_netD.eval()
fixed_netD.eval()
utility_audio_net.eval()
privacy_audio_net.eval()
# Pretrained FID loading
spec_FID_net.load_state_dict(torch.load('fixed_classifier_checkpoints/best_digit_alexnet_spectrograms_epoch_26.pt'))
audio_FID_net.load_state_dict(torch.load('fixed_classifier_checkpoints/audio_digit_net_early_stop_epoch_26.pt'))
# Initialize arrays for accuracies and fid scores
spec_digit_accuracy_F = []
spec_original_gender_accuracy_F = []
spec_digit_accuracy_G = []
spec_original_gender_accuracy_G = []
spec_sampled_gender_accuracy_G = []
audio_digit_accuracy_F = []
audio_original_gender_accuracy_F = []
audio_digit_accuracy_G = []
audio_original_gender_accuracy_G = []
audio_sampled_gender_accuracy_G = []
fid_spec_F = []
fid_spec_G = []
fid_audio_F = []
fid_audio_G = []
fid_inverted_audio = []
for i in range(num_runs):
run_path = os.path.join(args.experiment_path, runs[i])
result_dir = os.path.join(run_path, 'results')
audio_result_dir = os.path.join(result_dir, 'audio')
spec_result_dir = os.path.join(result_dir, 'spectrograms')
checkpoint_dir = os.path.join(run_path, 'checkpoints')
# os.mkdir(result_dir)
# os.mkdir(audio_result_dir)
# os.mkdir(spec_result_dir)
# Set up and load trained model
netF = UNetFilter(1, 1, chs=[8, 16, 32, 64, 128], kernel_size = args.filter_receptive_field, image_width=32, image_height=80, noise_dim=10, nb_classes=2, embedding_dim=16, use_cond = False).to(device)
netG = UNetFilter(1, 1, chs=[8, 16, 32, 64, 128], kernel_size = args.filter_receptive_field, image_width=32, image_height=80, noise_dim=10, nb_classes=2, embedding_dim=16, use_cond = True).to(device)
netF.load_state_dict(torch.load(os.path.join(checkpoint_dir, 'netF_latest_epoch_1000.pt')))
netG.load_state_dict(torch.load(os.path.join(checkpoint_dir, 'netG_latest_epoch_1000.pt')))
spec_correct_digit_F = 0
spec_correct_original_gender_F = 0
spec_correct_digit_G = 0
spec_correct_original_gender_G = 0
spec_correct_sampled_gender_G = 0
audio_correct_digit_F = 0
audio_correct_original_gender_F = 0
audio_correct_digit_G = 0
audio_correct_original_gender_G = 0
audio_correct_sampled_gender_G = 0
acts_real_spec = []
acts_fake_spec_F = []
acts_fake_spec_G = []
acts_real_audio = []
acts_fake_audio_F = []
acts_fake_audio_G = []
acts_inverted_audio = []
for j, (x, gender, digit, speaker_id) in enumerate(test_loader):
x = torch.unsqueeze(x,1)
spectrograms = fft(x).detach()
spectrograms, means, stds = preprocess_spectrograms(spectrograms)
spectrograms = torch.unsqueeze(spectrograms,1).to(device)
gender = gender.to(device)
digit = digit.to(device)
# --------------------------
# Spectrogram calculations
# --------------------------
z1 = torch.randn(spectrograms.shape[0], 10).to(device)
filter_mel = netF(spectrograms,z1, gender.long())
z2 = torch.randn(filter_mel.shape[0], 10).to(device)
gen_secret = Variable(LongTensor(np.random.choice([0.0, 1.0], filter_mel.shape[0]))).to(device)
gen_mel = netG(filter_mel,z2,gen_secret)
spec_pred_digit_F = utility_netD(filter_mel)
spec_pred_gender_F = fixed_netD(filter_mel)
spec_pred_digit_G = utility_netD(gen_mel)
spec_pred_gender_G = fixed_netD(gen_mel)
# Calculate digit accuracy for fixed net on F and G outputs
pred_digit_F = torch.argmax(spec_pred_digit_F.data,1)
spec_correct_digit_F += (pred_digit_F == digit).sum().item()
pred_digit_G = torch.argmax(spec_pred_digit_G.data,1)
spec_correct_digit_G += (pred_digit_G == digit).sum().item()
# Calculate gender accuracy for fixed net on F and G outputs
pred_gender_F = torch.argmax(spec_pred_gender_F.data,1)
spec_correct_original_gender_F += (pred_gender_F == gender.long()).sum().item()
pred_gender_G = torch.argmax(spec_pred_gender_G.data,1)
spec_correct_original_gender_G += (pred_gender_G == gender.long()).sum().item()
spec_correct_sampled_gender_G += (pred_gender_G == gen_secret).sum().item()
# Calculate FID on transformed spectrograms
# acts1_tmp = spec_FID_net(spectrograms)
# acts2_tmp = spec_FID_net(filter_mel)
# acts3_tmp = spec_FID_net(gen_mel)
#
# acts_real_spec.append(np.squeeze(acts1_tmp.detach().cpu().numpy()))
# acts_fake_spec_F.append(np.squeeze(acts2_tmp.detach().cpu().numpy()))
# acts_fake_spec_G.append(np.squeeze(acts3_tmp.detach().cpu().numpy()))
# Save spectrogram samples
if args.save_specs:
z1 = torch.randn(spectrograms.shape[0], 10).to(device)
filtered = netF(spectrograms,z1,gender.long()).detach()
z2 = torch.randn(spectrograms.shape[0], 10).to(device)
male = Variable(LongTensor(spectrograms.size(0)).fill_(1.0), requires_grad=False).to(device)
female = Variable(LongTensor(spectrograms.size(0)).fill_(0.0), requires_grad=False).to(device)
generated_male = netG(filtered, z2, male).detach()
generated_female = netG(filtered, z2, female).detach()
filtered = torch.squeeze(filtered,1).to(device) * 3 * stds.to(device) + means.to(device)
generated_male = generated_male.to(device) * 3 * stds.to(device) + means.to(device)
generated_female = generated_female.to(device) * 3 * stds.to(device) + means.to(device)
spectrograms = spectrograms.to(device) * 3 * stds.to(device) + means.to(device)
if gender == 0:
gender_title = 'female'
else:
gender_title = 'male'
orig_title = 'Original spectrogram - Gender: {} - Digit: {}'.format(gender_title, digit.item())
male_title = 'Sampled gender: male '
female_title = 'Sampled gender: female'
filtered_title = 'Filtered spectrogram'
f_name_original = os.path.join(spec_result_dir, 'speaker_{}_digit_{}_original.png'.format(
speaker_id.item(), digit.item()
))
f_name_male = os.path.join(spec_result_dir, 'speaker_{}_digit_{}_male.png'.format(
speaker_id.item(), digit.item()
))
f_name_female = os.path.join(spec_result_dir, 'speaker_{}_digit_{}_female.png'.format(
speaker_id.item(), digit.item()
))
f_name_filtered = os.path.join(spec_result_dir, 'speaker_{}_digit_{}_filtered.png'.format(speaker_id.item(), digit.item()
))
save_spec_plot(f_name_original, spectrograms, orig_title)
save_spec_plot(f_name_male, generated_male, male_title)
save_spec_plot(f_name_female, generated_female, female_title)
save_spec_plot(f_name_filtered, filtered, filtered_title)
# --------------------------
# Audio calculations
# --------------------------
# Denormalize spectrograms before inversion
filter_mel = torch.squeeze(filter_mel,1).to(device) * 3 * stds.to(device) + means.to(device)
gen_mel = torch.squeeze(gen_mel,1).to(device) * 3 * stds.to(device) + means.to(device)
spectrograms = torch.squeeze(spectrograms,1).to(device) * 3 * stds.to(device) + means.to(device)
# Invert spectrograms using MelGAN
original_audio = Mel2Audio(spectrograms)
filter_audio = Mel2Audio(filter_mel)
gen_audio = Mel2Audio(gen_mel)
# Classify transformed audio
audio_pred_digit_F, _ = utility_audio_net(filter_audio)
audio_pred_gender_F, _ = privacy_audio_net(filter_audio)
audio_pred_digit_G, _ = utility_audio_net(gen_audio)
audio_pred_gender_G, _ = privacy_audio_net(gen_audio)
pred_digit_F = torch.argmax(audio_pred_digit_F.data, 1)
pred_gender_F = torch.argmax(audio_pred_gender_F.data, 1)
pred_digit_G = torch.argmax(audio_pred_digit_G.data, 1)
pred_gender_G = torch.argmax(audio_pred_gender_G.data, 1)
audio_correct_digit_F += (pred_digit_F == digit).sum().item()
audio_correct_original_gender_F += (pred_gender_F == gender.long()).sum().item()
audio_correct_digit_G += (pred_digit_G == digit).sum().item()
audio_correct_sampled_gender_G += (pred_gender_G == gen_secret).sum().item()
audio_correct_original_gender_G += (pred_gender_G == gender.long()).sum().item()
# Compute activations for FID calculations for the transformed audio
_, acts1_tmp = audio_FID_net(x.to(device))
_, acts2_tmp = audio_FID_net(filter_audio.to(device))
_, acts3_tmp = audio_FID_net(gen_audio.to(device))
_, acts4_tmp = audio_FID_net(original_audio.to(device))
acts1_tmp = torch.flatten(acts1_tmp,1)
acts2_tmp = torch.flatten(acts2_tmp,1)
acts3_tmp = torch.flatten(acts3_tmp,1)
acts4_tmp = torch.flatten(acts4_tmp,1)
acts_real_audio.append(np.squeeze(acts1_tmp.detach().cpu().numpy()))
acts_fake_audio_F.append(np.squeeze(acts2_tmp.detach().cpu().numpy()))
acts_fake_audio_G.append(np.squeeze(acts3_tmp.detach().cpu().numpy()))
acts_inverted_audio.append(np.squeeze(acts4_tmp.detach().cpu().numpy()))
# --------------------------
# Save audio sample
# --------------------------
if args.save_audio:
if j % 2 == 0:
original_audio_sample = torch.squeeze(original_audio[0]).detach().cpu()
gen_audio_sample = torch.squeeze(gen_audio[0]).detach().cpu()
speaker_id_sample = speaker_id[0].detach().cpu()
digit_sample = digit[0].detach().cpu()
gender_sample = gender[0].detach().cpu()
gen_secret_sample = gen_secret[0].detach().cpu()
f_name_orig_audio = os.path.join(audio_result_dir,'speaker_{}_digit_{}_original_inverted.wav'.format(speaker_id_sample, digit_sample, gender_sample, gen_secret_sample))
f_name_gen_audio = os.path.join(audio_result_dir,'speaker_{}_digit_{}_gender_orig_{}_sampled_{}.wav'.format(speaker_id_sample, digit_sample, gender_sample, gen_secret_sample))
save_sample(f_name_orig_audio, args.sampling_rate, original_audio_sample)
save_sample(f_name_gen_audio, args.sampling_rate, gen_audio_sample)
# Calcuate accuracies
spec_digit_accuracy_F.append(100*spec_correct_digit_F / n_test)
spec_original_gender_accuracy_F.append(100*spec_correct_original_gender_F / n_test)
spec_digit_accuracy_G.append(100*spec_correct_digit_G / n_test)
spec_original_gender_accuracy_G.append(100*spec_correct_original_gender_G / n_test)
spec_sampled_gender_accuracy_G.append(100*spec_correct_sampled_gender_G / n_test)
audio_digit_accuracy_F.append(100*audio_correct_digit_F / n_test)
audio_original_gender_accuracy_F.append(100*audio_correct_original_gender_F / n_test)
audio_digit_accuracy_G.append(100*audio_correct_digit_G / n_test)
audio_original_gender_accuracy_G.append(100*audio_correct_original_gender_G / n_test)
audio_sampled_gender_accuracy_G.append(100*audio_correct_sampled_gender_G / n_test)
# Concatenate batch activations into single array
# acts_real_spec = np.concatenate(acts_real_spec, axis = 0)
# acts_fake_spec_F = np.concatenate(acts_fake_spec_F, axis = 0)
# acts_fake_spec_G = np.concatenate(acts_fake_spec_G, axis = 0)
acts_real_audio = np.concatenate(acts_real_audio, axis = 0)
acts_fake_audio_F = np.concatenate(acts_fake_audio_F, axis = 0)
acts_fake_audio_G = np.concatenate(acts_fake_audio_G, axis = 0)
acts_inverted_audio = np.concatenate(acts_inverted_audio, axis = 0)
# Calculate FID scores
# fid_spec_tmp_F = compute_frechet_inception_distance(acts_real_spec, acts_fake_spec_F)
# fid_spec_tmp_G = compute_frechet_inception_distance(acts_real_spec, acts_fake_spec_G)
#fid_audio_tmp_F = compute_frechet_inception_distance(acts_real_audio, acts_fake_audio_F)
#fid_audio_tmp_G = compute_frechet_inception_distance(acts_real_audio, acts_fake_audio_G)
#if i == 0:
# fid_inverted_audio_tmp = compute_frechet_inception_distance(acts_real_audio, acts_inverted_audio)
# fid_spec_F.append(fid_spec_tmp_F)
#fid_audio_F.append(fid_audio_tmp_F)
# fid_spec_G.append(fid_spec_tmp_G)
#fid_audio_G.append(fid_audio_tmp_G)
#fid_inverted_audio.append(fid_inverted_audio_tmp)
print("Computed accuracies and FID for run {}.".format(i))
# Create data frames with the accuracies
spec_digit_acc_F_df = pd.DataFrame(spec_digit_accuracy_F)
spec_orig_gender_acc_F_df = pd.DataFrame(spec_original_gender_accuracy_F)
spec_digit_acc_G_df = pd.DataFrame(spec_digit_accuracy_G)
spec_orig_gender_acc_G_df = pd.DataFrame(spec_original_gender_accuracy_G)
spec_sampled_gender_acc_G_df = pd.DataFrame(spec_sampled_gender_accuracy_G)
audio_digit_acc_F_df = pd.DataFrame(audio_digit_accuracy_F)
audio_orig_gender_acc_F_df = pd.DataFrame(audio_original_gender_accuracy_F)
audio_digit_acc_G_df = pd.DataFrame(audio_digit_accuracy_G)
audio_orig_gender_acc_G_df = pd.DataFrame(audio_original_gender_accuracy_G)
audio_sampled_gender_acc_G_df = pd.DataFrame(audio_sampled_gender_accuracy_G)
fid_spec_F_df = pd.DataFrame(fid_spec_F)
fid_audio_F_df = pd.DataFrame(fid_audio_F)
fid_spec_G_df = pd.DataFrame(fid_spec_G)
fid_audio_G_df = pd.DataFrame(fid_audio_G)
fid_inverted_audio_df = pd.DataFrame(fid_inverted_audio)
# Save accuracies and FID scores
spec_digit_acc_F_df.to_csv(os.path.join(args.experiment_path, 'spec_digit_acc_F.csv'), index = False)
spec_orig_gender_acc_F_df.to_csv(os.path.join(args.experiment_path, 'spec_orig_gender_acc_F.csv'), index = False)
spec_digit_acc_G_df.to_csv(os.path.join(args.experiment_path, 'spec_digit_acc_G.csv'), index = False)
spec_orig_gender_acc_G_df.to_csv(os.path.join(args.experiment_path, 'spec_orig_gender_acc_G.csv'), index = False)
spec_sampled_gender_acc_G_df.to_csv(os.path.join(args.experiment_path, 'spec_sampled_gender_acc_G.csv'), index = False)
audio_digit_acc_F_df.to_csv(os.path.join(args.experiment_path, 'audio_digit_acc_F.csv'), index = False)
audio_orig_gender_acc_F_df.to_csv(os.path.join(args.experiment_path, 'audio_orig_gender_acc_F.csv'), index = False)
audio_digit_acc_G_df.to_csv(os.path.join(args.experiment_path, 'audio_digit_acc_G.csv'), index = False)
audio_orig_gender_acc_G_df.to_csv(os.path.join(args.experiment_path, 'audio_orig_gender_acc_G.csv'), index = False)
audio_sampled_gender_acc_G_df.to_csv(os.path.join(args.experiment_path, 'audio_sampled_gender_acc_G.csv'), index = False)
fid_spec_F_df.to_csv(os.path.join(args.experiment_path, 'fid_spec_F.csv'), index = False)
fid_audio_F_df.to_csv(os.path.join(args.experiment_path, 'fid_audio_F.csv'), index = False)
fid_spec_G_df.to_csv(os.path.join(args.experiment_path, 'fid_spec_G.csv'), index = False)
fid_audio_G_df.to_csv(os.path.join(args.experiment_path, 'fid_audio_G.csv'), index = False)
fid_inverted_audio_df.to_csv(os.path.join(args.experiment_path, 'fid_inverted_audio_9_june_G.csv'), index = False)
if __name__ == "__main__":
main()