/
train.py
387 lines (308 loc) · 13.5 KB
/
train.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
import argparse
# import numpy as np
import os
import pandas as pd
import pdb
import random
import soundfile as sf
import time
import torch
import torchaudio
import torch.nn.functional as F
import numpy as np
from collections import defaultdict
from datetime import date, datetime
from model import Voxceleb12_sample_cssl2, AMI_sample_cssl
from model import Wav2VecSpeakerCSSL
from pathlib import Path
from torch.optim import Adam, SGD
from torch.utils.data import DataLoader, Dataset
from torch.nn.utils.rnn import pad_sequence
from transformers import Wav2Vec2Model, Wav2Vec2Config
from transformers.optimization import get_linear_schedule_with_warmup
from utils import EarlyStopping, BalancedBatchSampler
parser = argparse.ArgumentParser(description="Fine-tuning Wav2Vec 2")
parser.add_argument('--base_path', type=str,
default='/data/valencia/voxceleb/data/voxceleb1',
help='base location of the data')
parser.add_argument('--batch_size', type=int, default=16,
help='batch size for training')
parser.add_argument('--model_type', type=str,
default='facebook/wav2vec2-base',
help='pretrained model name or path to trained model')
parser.add_argument('--lr', type=float, default=2e-5,
help='learning rate')
parser.add_argument('--pct_warmup_steps', type=float, default=0.1,
help='percentage of total steps for lr warmup')
parser.add_argument('--lr_scheduler', type=int, default=1,
help='with lr scheduler or without, 1=with, 0=constant')
parser.add_argument('--epochs', type=int, default=5,
help='number of epochs')
parser.add_argument('--num_freeze_steps', type=int, default=0,
help='number of epochs to freeze model layers')
parser.add_argument('--logfile', type=str, default='logs/log_1.log',
help='path to save the final model')
parser.add_argument('--seed', type=int, default=100,
help='random_seed')
parser.add_argument('--optimizer_type', type=str, default="Adam",
help='type of optimizer: Adam or SGD')
parser.add_argument('--custom_embed_size', type=int, default=128,
help='add an extra FC layer and specify embed size/dim to be extracted')
parser.add_argument('--grad_acc_step', type=int, default=4,
help='number of steps to accumulate the gradient for')
parser.add_argument('--save_path', type=str, default="saved_models/test.pt",
help='path to save model')
parser.add_argument('--resume_training', type=int, default=0,
help='resume current training or load huggingace model, 1=resume, 0=load')
parser.add_argument('--data_type', type=str, default="ami",
help='To train with Vox1+2 or ami')
parser.add_argument('--with_relu', type=int, default=0,
help='with or without relu for FC layer and embedding')
parser.add_argument('--dropout_val', type=float, default=0,
help='additional dropout')
parser.add_argument('--refine_matrix', type=int, default=0,
help='whether to have thresholding in CE step of AP loss')
parser.add_argument('--num_classes_in_batch', type=int, default=0,
help='to specify a specific number of speakers in each batch, 0 for random number of speakers')
parser.add_argument('--g_blur', type=float, default=1.,
help='gaussian blur to use with contrastive loss. 0=no blur & use abs threshold')
parser.add_argument('--p_pct', type=int, default=100,
help='threshold for the AP and MSE loss')
parser.add_argument('--mse_fac', type=float, default=0.0,
help='weight of the mse loss')
parser.add_argument('--margin', type=float, default=0.0,
help='P percentile margin')
args = parser.parse_args()
def logging(s, logging_=True, log_=True):
if logging_:
print(s)
if log_:
with open(args.logfile, 'a+') as f_log:
f_log.write(s + '\n')
torch.manual_seed(args.seed)
torch.cuda.manual_seed(args.seed)
torch.cuda.manual_seed_all(args.seed)
np.random.seed(args.seed)
random.seed(args.seed)
torch.backends.cudnn.benchmark = False
torch.backends.cudnn.deterministic = True
logging("\n-------------------")
logging(f"date: {date.today()}")
logging(f"time: {datetime.now().strftime('%H:%M:%S')}")
def collate_pad_batch(batch):
"""
pads data of different length to the same max length
"""
seg_1 = [t[0] for t in batch]
seg_2 = [t[1] for t in batch]
labels = torch.tensor([t[2] for t in batch])
seg_1 = pad_sequence(seg_1, batch_first=True)
seg_2 = pad_sequence(seg_2, batch_first=True)
return seg_1, seg_2, labels
def eval(model, val_loader):
model.eval()
eval_loss = 0
num_samples = 0
with torch.no_grad():
for seg_1, seg_2, label in val_loader:
seg_1 = seg_1.to(device)
seg_2 = seg_2.to(device)
label = label.to(device)
loss, _ = model(seg_1, seg_2, label)
eval_loss += float(loss.item()) * seg_1.size(0)
num_samples += seg_1.size(0)
logging(f"eval loss: {eval_loss/ num_samples}")
return eval_loss
def train(train_loader):
if config["resume_training"]:
logging("------ RESUMING TRAINING -----")
w2v2_model = Wav2Vec2Model(config["w2v2_config"])
model = Wav2VecSpeakerCSSL(w2v2_model, config)
# for case where there might be mismatching keys due to diff dim etc.
finetuned_dict = torch.load(config["w2v2_model"], map_location=device)
model_dict = model.state_dict()
# 1. filter out unnecessary keys
finetuned_dict = {k: v for k, v in finetuned_dict.items() if
(k in model_dict) and (model_dict[k].shape == finetuned_dict[k].shape)}
# 2. overwrite entries in the existing state dict
model_dict.update(finetuned_dict)
# 3. load the new state dict
model.load_state_dict(model_dict)
else:
w2v2_model = Wav2Vec2Model.from_pretrained(config["w2v2_model"])
model = Wav2VecSpeakerCSSL(w2v2_model, config)
model.to(device)
if args.optimizer_type == "Adam":
optimizer = Adam(filter(lambda p: p.requires_grad, model.parameters()), lr=config["lr"])
elif args.optimizer_type == "SGD":
optimizer = SGD(filter(lambda p: p.requires_grad, model.parameters()), lr=config["lr"])
else:
raise ValueError('Optimizer must be one of Adam or SGD')
t_total = len(train_loader) // config["grad_acc_step"] * config["epochs"]
num_warmup_steps = int(t_total * config["pct_warmup_steps"])
logging(f"total steps: {t_total}")
if config["lr_scheduler"]:
lr_scheduler = get_linear_schedule_with_warmup(
optimizer,
num_warmup_steps,
t_total)
logging("Training model...")
min_val_loss = float('inf')
estop = EarlyStopping(best_loss=min_val_loss, patience=1, min_delta=0)
num_steps = 0
for epoch in range(config["epochs"]):
logging(f"epoch {epoch}")
start = time.time()
running_loss = 0.0
samples = 0
curr_epoch_steps = 0
model.train()
for i, (seg_1, seg_2, label) in enumerate(dataloader):
seg_1 = seg_1.to(device)
seg_2 = seg_2.to(device)
label = label.to(device)
# freeze base
if num_steps < config["num_freeze_steps"]:
model.freeze_base()
else:
model.unfreeze_base()
loss, _ = model(seg_1, seg_2, label)
loss = loss / config["grad_acc_step"]
# backward
loss.backward()
# grad accumulation
if (i+1) % config["grad_acc_step"] == 0:
optimizer.step()
optimizer.zero_grad()
if config["lr_scheduler"]:
lr_scheduler.step()
num_steps += 1
curr_epoch_steps += 1
running_loss += float(loss.item()) * seg_1.size(0) * config['grad_acc_step']
samples += seg_1.size(0)
if curr_epoch_steps > 0 and curr_epoch_steps % 50 == 0 and (i+1) % curr_epoch_steps == 0:
logging(f"epoch: {epoch} | new step: {curr_epoch_steps} | loss: {running_loss/samples} | time taken: {time.time() - start}")
print("eval ---")
# evaluate model
eval_loss = eval(model, val_dataloader)
logging(f"End of epoch {epoch} | time taken: {time.time() - start}")
# save best model
# but continue training till end
estop(eval_loss)
if eval_loss <= estop.best_loss:
logging("Saving model...")
torch.save(model.state_dict(), config["save_path"])
else:
logging("loss up")
if config['epochs'] == 1:
logging("Saving model coz only one epoch...")
torch.save(model.state_dict(), config["save_path"])
logging("Finished training!")
#### LOADING DATA #############
base_path = '/data/valencia/voxceleb/data/voxceleb1'
dev_path = Path(os.path.join(base_path, 'dev/wav'))
dev_path2 = Path("/data/valencia/voxceleb/data/voxceleb2/dev/aac")
if args.data_type == "vox_12":
logging(f"using positive samples from other utterance Vox 12")
train_file1 = '/home/mifs/epcl2/project/w2v2_sv/data/voxceleb1_train.csv'
train_file2 = '/home/mifs/epcl2/project/w2v2_sv/data/voxceleb2_train.csv'
val_file1 = '/home/mifs/epcl2/project/w2v2_sv/data/voxceleb1_val.csv'
val_file2 = '/home/mifs/epcl2/project/w2v2_sv/data/voxceleb2_val.csv'
train_df1 = pd.read_csv(train_file1)
train_df2 = pd.read_csv(train_file2)
val_df1 = pd.read_csv(val_file1)
val_df2 = pd.read_csv(val_file2)
train_data = Voxceleb12_sample_cssl2(dev_path, dev_path2, train_df1, train_df2)
val_data = Voxceleb12_sample_cssl2(dev_path, dev_path2, val_df1, val_df2)
train_df = pd.concat([train_df1, train_df2]).reset_index(drop=True)
val_df = pd.concat([val_df1, val_df2]).reset_index(drop=True)
label_column = 'speaker_id'
logging(f"num classes: {train_df[label_column].nunique()}")
elif args.data_type == 'ami':
logging(f"Using ami with positive samples from other utterances AMI")
base_path = '/home/dawna/cz277/ami/transcription/data/wave'
train_file = '/home/mifs/epcl2/project/ami/data/random_split/ami_train.csv'
val_file = '/home/mifs/epcl2/project/ami/data/random_split/ami_val.csv'
train_df = pd.read_csv(train_file)
val_df = pd.read_csv(val_file)
train_data = AMI_sample_cssl(base_path, train_df)
val_data = AMI_sample_cssl(base_path, val_df)
label_column = 'label'
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
print(f"device: {device}")
# ensures that there are num_classes_in_batch diff speakers in each batch
if args.num_classes_in_batch:
train_batch_sampler = BalancedBatchSampler(data_df=train_df,
n_classes=args.num_classes_in_batch,
n_samples=int(args.batch_size//args.num_classes_in_batch),
label_column_name=label_column)
dataloader = DataLoader(
train_data,
batch_sampler=train_batch_sampler,
collate_fn=collate_pad_batch,
num_workers=2
)
val_batch_sampler = BalancedBatchSampler(data_df=val_df,
n_classes=args.num_classes_in_batch,
n_samples=int(args.batch_size//args.num_classes_in_batch),
label_column_name=label_column)
val_dataloader = DataLoader(
val_data,
batch_sampler=val_batch_sampler,
collate_fn=collate_pad_batch,
num_workers=2
)
else:
dataloader = DataLoader(
train_data,
batch_size=args.batch_size,
shuffle=True,
collate_fn=collate_pad_batch,
num_workers=2
)
val_dataloader = DataLoader(
val_data,
batch_size=args.batch_size,
shuffle=False,
collate_fn=collate_pad_batch,
num_workers=2
)
if args.dropout_val:
w2v2_config = Wav2Vec2Config(hidden_dropout = args.dropout_val,
activation_dropout = args.dropout_val,
attention_dropout = args.dropout_val,
feat_proj_dropout = args.dropout_val,
feat_quantizer_dropout = args.dropout_val,
final_dropout = args.dropout_val,
layerdrop = args.dropout_val)
else:
w2v2_config = Wav2Vec2Config()
config = {
"base_path": args.base_path,
"batch_size": args.batch_size,
"num_classes_in_batch": args.num_classes_in_batch,
"w2v2_model": args.model_type,
"device": device,
"lr": args.lr,
"pct_warmup_steps": args.pct_warmup_steps,
"num_freeze_steps": args.num_freeze_steps,
"epochs": args.epochs,
"optimizer_type": args.optimizer_type,
"grad_acc_step": args.grad_acc_step,
"save_path": args.save_path,
"resume_training": args.resume_training,
"lr_scheduler":args.lr_scheduler,
"data_type": args.data_type,
"custom_embed_size": args.custom_embed_size,
"with_relu": args.with_relu,
"dropout_val": args.dropout_val,
"refine_matrix": args.refine_matrix,
"g_blur": args.g_blur,
"p_pct": args.p_pct,
"mse_fac": args.mse_fac,
"margin": args.margin,
"w2v2_config": w2v2_config,
}
for item in config:
logging(f"{item}: {config[item]}")
train(dataloader)