-
Notifications
You must be signed in to change notification settings - Fork 8
/
train.py
232 lines (207 loc) · 9.38 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
import argparse
from tqdm import tqdm
import torch
import os
from DFCNN import AcousticModel
from readdata import SpeechData
import torch.nn as nn
import logging
from torchsummary import summary
import torchsnooper
import horovod.torch as hvd
from tensorboardX import SummaryWriter as writer
import random
from BeamSearch import ctcBeamSearch
logging.basicConfig(level=logging.INFO, format='%(asctime)s - %(name)s - %(levelname)s - %(message)s')
logger = logging.getLogger(__name__)
parser = argparse.ArgumentParser()
# required args
parser.add_argument('--data_type',
type=str,
required=True,
help='Either aishell or thchs')
parser.add_argument('--model_path',
type=str,
required=True,
help="loading or saving model")
# parameters for features.py and data iterator
parser.add_argument('--config_path',
type=str,
default='data_config',
help='config file')
parser.add_argument('--window_size',
type=int,
default=400,
help="Hamming window size, default value as 400")
parser.add_argument('--numcep',
type=int,
default=26,
help="MFCC feature numbers : number of cepstrum")
parser.add_argument('--time_window',
type=int,
default=25,
help="time window")
parser.add_argument('--step',
type=int,
default=10,
help="window shift step")
parser.add_argument('--batch_size',
type=int,
default=8,)
parser.add_argument('--max_window',
type=int,
default=1600,
help='Maximum of windows')
parser.add_argument('--max_seq_length',
type=int,
default=64,
help='maximum of sequence length')
# training procedures
parser.add_argument('--model_name',
type=str,
default="logs_am.pt")
parser.add_argument('--gpu_rank',
type=int,
default=4,
help="define number of gpus to use")
parser.add_argument('--lr',
type=float,
default=0.0001,
help="learning rate")
parser.add_argument('--epochs',
type=int,
default=5,
help="training epoch")
parser.add_argument('--save_step',
type=int,
default=10)
parser.add_argument('--load_model',
action="store_true")
parser.add_argument('--dev_step',
type=int,
default=5)
args = parser.parse_args()
if __name__ == '__main__':
# initialize horovod
hvd.init()
from torch.utils.data import DataLoader
import torch.optim as optim
torch.cuda.set_device(hvd.local_rank())
if torch.cuda.current_device() == 0:
logger.info("Args are as : %s" % args)
if torch.cuda.is_available():
assert args.gpu_rank <= torch.cuda.device_count(),\
"Should ensure that gpu_rank <= the number of yournvidia devices"
logger.info("%d gpu(s) detected and using %d devices." % (torch.cuda.device_count(), args.gpu_rank))
else:
logger.info("Using cpu will take you quite a long time, not recommended!")
train_data = SpeechData('data_config', dataset=args.data_type)
dev_data = SpeechData('data_config', type='dev', dataset=args.data_type)
train_sampler = torch.utils.data.distributed.DistributedSampler(train_data,
num_replicas=hvd.size(),
rank=hvd.rank())
dev_sampler = torch.utils.data.distributed.DistributedSampler(dev_data,
num_replicas=hvd.size(),
rank=hvd.rank())
vocab_size = train_data.label_nums()
logger.info("Total PNY labels : %d" % vocab_size)
train_loader = DataLoader(train_data, batch_size=args.batch_size, sampler=train_sampler)
dev_loader = DataLoader(dev_data, batch_size=int(args.batch_size/2), sampler=dev_sampler)
model = AcousticModel(vocab_size=vocab_size, input_dimension=200)
model_path = args.model_path + '/' + args.model_name
if os.path.exists(model_path) and args.load_model:
logger.info("Model: %s exists, loading existing model..." % model_path)
model = torch.load(model_path)
else:
if not os.path.exists(args.model_path):
os.mkdir(args.model_path)
logger.info('No existing model in your provided path.')
model = model.cuda()
if torch.cuda.current_device() == 0:
print(model)
criterion = nn.CTCLoss()
optimizer = optim.Adam(model.parameters(), lr=args.lr * hvd.size()/2)
hvd.broadcast_parameters(model.state_dict(), root_rank=0)
hvd.broadcast_optimizer_state(optimizer, root_rank=0)
optimizer = hvd.DistributedOptimizer(optimizer, named_parameters=model.named_parameters())
if torch.cuda.current_device() == 0:
writer = writer()
tr_run = 0
for epoch in tqdm(range(args.epochs)):
if (epoch+1) % args.dev_step == 0:
model.eval()
model.cuda()
eval_loss = 0.0
if torch.cuda.current_device() == 0:
logger.info("Starting evaluating!")
for batch_idx, samples in enumerate(dev_loader):
# optimizer.zero_grad()
X, y, input_lengths, label_lengths, transcripts = samples
X = X.type(torch.FloatTensor).cuda()
y = y.type(torch.LongTensor).cuda()
input_lengths = model.convert(input_lengths)
label_lengths = label_lengths.cuda()
input_lengths = input_lengths.cuda()
with torch.no_grad():
outputs = model(X)
# print(outputs.shape, y.shape)
# outputs = outputs.cpu()
loss = criterion(outputs,
y,
input_lengths,
label_lengths)
eval_loss += loss
# loss.backward()
# optimizer.step()
if (batch_idx + 1) % int(len(dev_data)/(args.gpu_rank*int(args.batch_size/2))) == 0:
# print(outputs)
logger.info("Evaluating step %d, step loss : %.5f , total_mean_loss : %.5f"
% (batch_idx + 1, loss, eval_loss / (batch_idx + 1)))
# logger.info("mean_eval_loss: %.5f" % eval_loss/len(dev_data))
else:
model.train()
model.cuda()
train_sampler.set_epoch(epoch)
epoch_loss = 0.0
logger.info("Running epoch %d/%d =>" % (epoch+1, args.epochs))
for batch_idx, samples in enumerate(train_loader):
tr_run += 1
optimizer.zero_grad()
X, y, input_lengths, label_lengths, transcripts = samples
X = X.type(torch.FloatTensor).cuda()
# y = y.reshape((args.batch_size*64,))
y = y.type(torch.LongTensor).cuda()
input_lengths = model.convert(input_lengths)
label_lengths = label_lengths.cuda()
input_lengths = input_lengths.cuda()
outputs = model(X)
# print(outputs.shape, y.shape)
# outputs = outputs.cpu()
loss = criterion(outputs,
y,
input_lengths,
label_lengths)
if torch.cuda.current_device() == 0:
writer.add_scalar('tr_loss', loss.data.item(), tr_run)
epoch_loss += loss
loss.backward()
optimizer.step()
if (batch_idx+1) % 10 == 0 and torch.cuda.current_device() == 0:
# print(outputs)
logger.info("Training step %d, progress %d/%d, step loss : %.5f , total_mean_loss : %.5f"
% (batch_idx+1, batch_idx+1, int(len(train_data)/(args.batch_size * args.gpu_rank)),
loss, epoch_loss/(batch_idx+1)))
# if loss < 10:
# # select_index = random.randint(0, args.batch_size-1)
# result = outputs.detach().transpose(0, 1).cpu().numpy()
# # result = np.transpose(result, [1, 0, 2])
# # print(result.shape)
# input_length = input_lengths.cpu().numpy()
# # result = np.hstack((result[:, 1:], result[:, 0].reshape((-1, 1))))
# # classes = list(train_data.i2w.values())[1:]
# decode = K.ctc_decode(result, input_length, greedy=False)
# print(K.get_value(decode[0][0]))
if (epoch+1) % args.save_step == 0 and torch.cuda.current_device() == 0:
logger.info("Saving model parameters into %s" % model_path)
torch.save(model.cpu(), model_path)
writer.close()