forked from lochenchou/MOSNet
/
train_rep.py
334 lines (268 loc) · 11.6 KB
/
train_rep.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
import os
import time
import numpy as np
from tqdm import tqdm
import scipy.stats
import pandas as pd
import matplotlib
# Force matplotlib to not use any Xwindows backend.
matplotlib.use('Agg')
import matplotlib.pyplot as plt
import argparse
import tensorflow as tf
from tensorflow import keras
import model_rep
import utils
import random
from sklearn.metrics import accuracy_score
from sklearn.metrics import confusion_matrix
parser = argparse.ArgumentParser()
parser.add_argument("--model", help="model to train with, CNN")
parser.add_argument("--epoch", type=int, default=100, help="number epochs")
parser.add_argument("--data", help="data: VC, LA")
parser.add_argument("--feats", help="feats: orig, DS-image, xvec_, or CNN")
parser.add_argument("--seed", type=int, default=1984, help="specify a seed")
parser.add_argument("--test_only", type=bool, default=False, help="True for test only")
args = parser.parse_args()
random.seed(args.seed)
if not args.model:
raise ValueError('please specify model to train with, CNN, etc')
print('training with model architecture: {}'.format(args.model))
print('epochs: {}'.format(args.epoch))
print('training with feature type: {}'.format(args.feats))
print('Test only: {}'.format(args.test_only))
# 0 = all messages are logged (default behavior)
# 1 = INFO messages are not printed
# 2 = INFO and WARNING messages are not printed
# 3 = INFO, WARNING, and ERROR messages are not printed
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3' # or any {'0', '1', '2'}
tf.debugging.set_log_device_placement(False)
# set memory growth
gpus = tf.config.experimental.list_physical_devices('GPU')
if gpus:
try:
# Currently, memory growth needs to be the same across GPUs
for gpu in gpus:
tf.config.experimental.set_memory_growth(gpu, True)
logical_gpus = tf.config.experimental.list_logical_devices('GPU')
print(len(gpus), "Physical GPUs,", len(logical_gpus), "Logical GPUs")
except RuntimeError as e:
# Memory growth must be set before GPUs have been initialized
print(e)
def run(l2_val, dr, n, batch_size, bn):
# set dir
DATA_DIR = './data_'+args.data
BIN_DIR = os.path.join(DATA_DIR, args.feats)
OUTPUT_DIR = './results_R3/output_'+args.model+"_"+str(batch_size)+"_"+args.data+"_"+args.feats+"_"+str(l2_val)+"_"+str(dr)+"_"+str(n)+"_"+str(bn)
results_file = OUTPUT_DIR+"/results.pkl"
EPOCHS = args.epoch
BATCH_SIZE = batch_size
if args.data == "VC":
NUM_TRAIN = 13580
NUM_TEST=4000
NUM_VALID=3000
mos_list = utils.read_list(os.path.join(DATA_DIR,'mos_list.txt'))
random.shuffle(mos_list)
train_list= mos_list[0:-(NUM_TEST+NUM_VALID)]
random.shuffle(train_list)
valid_list= mos_list[-(NUM_TEST+NUM_VALID):-NUM_TEST]
test_list= mos_list[-NUM_TEST:]
train_data_feat, train_data_mos = utils.data_rep(train_list, BIN_DIR)
valid_data_feat, valid_data_mos = utils.data_rep(valid_list, BIN_DIR)
if args.data == "LA":
test_list = utils.read_list(os.path.join(DATA_DIR,'test_list.txt'))
train_data_feat = np.load(DATA_DIR+'/'+args.feats+'_X_train.npy')
train_data_mos = np.load(DATA_DIR+'/'+args.feats+'_y_train.npy')
valid_data_feat = np.load(DATA_DIR+'/'+args.feats+'_X_valid.npy')
valid_data_mos = np.load(DATA_DIR+'/'+args.feats+'_y_valid.npy')
NUM_TRAIN = train_data_feat.shape[0]
NUM_TEST=valid_data_feat.shape[0]
NUM_VALID=len(test_list)
if not os.path.exists(OUTPUT_DIR):
os.makedirs(OUTPUT_DIR)
print('{} for training; {} for valid; {} for testing'.format(NUM_TRAIN, NUM_TEST, NUM_VALID))
# CNN-LDA has 100, and CNN-PCA has 512 ??
rep_dims = {'DS-image':4096, 'CNN':100, 'xvec_0':512, 'xvec_1':512, 'xvec_2':512, 'xvec_3':512, 'xvec_4':512, 'xvec_5':512}
# init model
if args.model == 'CNN':
dim = rep_dims[args.feats]
MOSNet = model_rep.CNN(dim, l2_val, dr, n, bn)
# elif args.model == 'FFN':
# dim = rep_dims[args.feats]
# MOSNet = model_rep.FFN(dim, dr, n, bn)
else:
raise ValueError('please specify model to train with, CNN, FFN')
sys.exit()
model = MOSNet.build()
model.compile(
optimizer=tf.keras.optimizers.Adam(1e-4),metrics=["mean_absolute_error"],
loss="mse")
CALLBACKS = [
keras.callbacks.ModelCheckpoint(
filepath=os.path.join(OUTPUT_DIR,'mosnet.h5'),
save_best_only=True,
monitor='val_loss',
verbose=1),
keras.callbacks.EarlyStopping(
monitor='val_loss',
mode='min',
min_delta=0,
patience=5,
verbose=1)
]
train_data_feat = np.expand_dims(train_data_feat, axis=3)
valid_data_feat = np.expand_dims(valid_data_feat, axis=3)
print(train_data_feat.shape)
print(train_data_mos.shape)
# start fitting model
hist = model.fit(x=train_data_feat, y=train_data_mos,
epochs=EPOCHS,
callbacks=CALLBACKS,
shuffle=True,
batch_size=BATCH_SIZE,
validation_data=(valid_data_feat, valid_data_mos),
verbose=1)
# plot testing result
model.load_weights(os.path.join(OUTPUT_DIR,'mosnet.h5'),) # Load the best model
print('testing...')
MOS_Predict=np.zeros([len(test_list),])
MOS_true =np.zeros([len(test_list),])
df = pd.DataFrame(columns=['audio', 'true_mos','predict_mos','system_ID','speaker_ID'])
for i in tqdm(range(len(test_list))):
if args.data == "VC":
filepath=test_list[i].split(',')
filename=filepath[0].split('.')[0]
sysid = ""
speakerid = ""
mos=float(filepath[1])
elif args.data == "LA":
filepath=test_list[i].split(',')
filename=filepath[2].split('.')[0]
sysid = filepath[1]
speakerid = filepath[0]
mos=float(filepath[3])
_DS = utils.read_rep(os.path.join(BIN_DIR,filename+'.npy'))
_DS = np.expand_dims(_DS, axis=3)
Average_score=model.predict(_DS, verbose=0, batch_size=1)
MOS_Predict[i]=Average_score
MOS_true[i] =mos
df = df.append({'audio': filepath[0],
'true_mos': MOS_true[i],
'predict_mos': MOS_Predict[i],
'system_ID': sysid,
'speaker_ID': speakerid},
ignore_index=True)
df.to_pickle(results_file)
plt.style.use('seaborn-deep')
x = df['true_mos']
y = df['predict_mos']
bins = np.linspace(1, 5, 40)
plt.figure(2)
plt.hist([x, y], bins, label=['true_mos', 'predict_mos'])
plt.legend(loc='upper right')
plt.xlabel('MOS')
plt.ylabel('number')
plt.savefig('./'+OUTPUT_DIR+'/MOSNet_distribution.png', dpi=150)
LCC=np.corrcoef(MOS_true, MOS_Predict)
print('[UTTERANCE] Linear correlation coefficient= %f' % LCC[0][1])
SRCC=scipy.stats.spearmanr(MOS_true.T, MOS_Predict.T)
print('[UTTERANCE] Spearman rank correlation coefficient= %f' % SRCC[0])
MSE=np.mean((MOS_true-MOS_Predict)**2)
print('[UTTERANCE] Test error= %f' % MSE)
# Plotting scatter plot
M=np.max([np.max(MOS_Predict),5])
plt.figure(3)
plt.scatter(MOS_true, MOS_Predict, s =15, color='b', marker='o', edgecolors='b', alpha=.20)
plt.xlim([0.5,M])
plt.ylim([0.5,M])
plt.xlabel('True MOS')
plt.ylabel('Predicted MOS')
plt.title('Utterance-Level')
plt.savefig('./'+OUTPUT_DIR+'/MOSNet_scatter_plot.png', dpi=150)
if args.data == "VC":
# load vcc2018_system
sys_df = pd.read_csv(os.path.join(DATA_DIR,'vcc2018_system.csv'))
df['system_ID'] = df['audio'].str.split('_').str[-1].str.split('.').str[0] + '_' + df['audio'].str.split('_').str[0]
elif args.data == "LA":
# load LA 2019 system
sys_df = pd.read_csv(os.path.join(DATA_DIR,'LA_mos_system.csv'))
sys_result_mean = df[['system_ID', 'predict_mos']].groupby(['system_ID']).mean()
sys_mer_df = pd.merge(sys_result_mean, sys_df, on='system_ID')
sys_true = sys_mer_df['mean']
sys_predicted = sys_mer_df['predict_mos']
print(sys_true)
print(sys_predicted)
print(sys_true.shape)
print(sys_predicted.shape)
LCC=np.corrcoef(sys_true, sys_predicted)
print('[SYSTEM] Linear correlation coefficient= %f' % LCC[0][1])
SRCC=scipy.stats.spearmanr(sys_true.T, sys_predicted.T)
print('[SYSTEM] Spearman rank correlation coefficient= %f' % SRCC[0])
MSE=np.mean((sys_true-sys_predicted)**2)
print('[SYSTEM] Test error= %f' % MSE)
# Plotting scatter plot
M=np.max([np.max(sys_predicted),5])
# m=np.max([np.min(sys_predicted)-1,0.5])
plt.figure(4)
plt.scatter(sys_true, sys_predicted, s =25, color='b', marker='o', edgecolors='b')
plt.xlim([1,M])
plt.ylim([1,M])
plt.xlabel('True MOS')
plt.ylabel('Predicted MOS')
plt.title('System-Level')
# # add system id
# for i in range(len(sys_mer_df)):
# sys_ID = mer_df['system_ID'][i]
# x = mer_df['mean'][i]
# y = mer_df['predict_mos'][i]
# plt.text(x-0.05, y+0.1, sys_ID, fontsize=8)
plt.savefig('./'+OUTPUT_DIR+'/MOSNet_system_scatter_plot.png', dpi=150)
if args.data == "LA":
spk_df = pd.read_csv(os.path.join(DATA_DIR,'LA_mos_speaker.csv'))
spk_result_mean = df[['speaker_ID', 'predict_mos']].groupby(['speaker_ID']).mean()
spk_mer_df = pd.merge(spk_result_mean, spk_df, on='speaker_ID')
spk_result_mean = df[['speaker_ID', 'predict_mos']].groupby(['speaker_ID']).mean()
spk_mer_df = pd.merge(spk_result_mean, spk_df, on='speaker_ID')
spk_true = spk_mer_df['mean']
spk_predicted = spk_mer_df['predict_mos']
LCC=np.corrcoef(spk_true, spk_predicted)
print('[SPEAKER] Linear correlation coefficient= %f' % LCC[0][1])
SRCC=scipy.stats.spearmanr(spk_true.T, spk_predicted.T)
print('[SPEAKER] Spearman rank correlation coefficient= %f' % SRCC[0])
MSE=np.mean((spk_true-spk_predicted)**2)
print('[SPEAKER] Test error= %f' % MSE)
# Plotting scatter plot
M=np.max([np.max(spk_predicted),5])
# m=np.max([np.min(spk_predicted)-1,0.5])
plt.figure(4)
plt.scatter(spk_true, spk_predicted, s =25, color='b', marker='o', edgecolors='b')
plt.xlim([1,M])
plt.ylim([1,M])
plt.xlabel('True MOS')
plt.ylabel('Predicted MOS')
plt.title('Speaker-Level')
# # add system id
# for i in range(len(spk_mer_df)):
# spk_ID = mer_df['speaker_ID'][i]
# x = mer_df['mean'][i]
# y = mer_df['predict_mos'][i]
# plt.text(x-0.05, y+0.1, spk_ID, fontsize=8)
plt.savefig('./'+OUTPUT_DIR+'/MOSNet_speaker_scatter_plot.png', dpi=150)
##############################################################################
# sweep these vals
L2_VALS = [0.0001, 0.001, 0.01, 0.1]
DRS = [0.1, 0.2, 0.3]
N = [16, 32, 64, 128]
BATCH_SIZES = [16, 64, 128]
BN = [True, False]
# vals for testing
#L2_VALS = [0.001]
#DRS = [0.1]
#N = [16]
#BATCH_SIZES = [16]
for l2_val in L2_VALS:
for dr in DRS:
for n in N:
for bn in BN:
for batch_size in BATCH_SIZES:
run(l2_val, dr, n, batch_size, bn)