-
Notifications
You must be signed in to change notification settings - Fork 5
/
Copy pathtrain_aco.py
201 lines (192 loc) · 9.08 KB
/
train_aco.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
import argparse
import musa
from musa.datasets import *
from musa.models import *
from torch.utils.data import DataLoader
from tensorboardX import SummaryWriter
from torch.optim.lr_scheduler import ReduceLROnPlateau
import torch.optim as optim
from musa.core import *
import random
import json
import os
def get_data_loaders(opts):
bsize = opts.batch_size
if opts.no_stateful:
bsize =None
trainset = TCSTAR_aco(opts.cfg_spk, 'train',
opts.aco_dir, opts.lab_dir,
opts.codebooks_dir,
force_gen=opts.force_gen,
parse_workers=opts.parser_workers,
max_seq_len=opts.max_seq_len,
batch_size=bsize,
max_spk_samples=opts.max_samples,
mulout=opts.mulout,
norm_aco=True,
exclude_train_spks=opts.exclude_train_spks)
if opts.mulout:
sampler = MOSampler(trainset.len_by_spk(), trainset, opts.batch_size)
shuffle = False
else:
sampler = None
shuffle = True
train_loader = DataLoader(trainset, batch_size=opts.batch_size, shuffle=shuffle,
num_workers=opts.loader_workers,
sampler=sampler,
collate_fn=varlen_aco_collate)
val_dset = TCSTAR_aco(opts.cfg_spk, 'valid', opts.aco_dir,
opts.lab_dir, opts.codebooks_dir,
norm_aco=True,
exclude_eval_spks=opts.exclude_eval_spks,
max_spk_samples=opts.max_samples,
parse_workers=opts.parser_workers,
max_seq_len=opts.max_seq_len,
batch_size=bsize,
mulout=opts.mulout)
# build validation dataset and loader
if opts.mulout:
va_sampler = MOSampler(val_dset.len_by_spk(), val_dset, opts.batch_size)
else:
va_sampler = None
valid_loader = DataLoader(val_dset, batch_size=opts.batch_size,
shuffle=False,
num_workers=opts.loader_workers,
sampler=va_sampler,
collate_fn=varlen_aco_collate)
return train_loader, valid_loader, trainset
def main(opts):
with open(os.path.join(opts.save_path,
'main.opts'), 'w') as opts_f:
opts_f.write(json.dumps(vars(opts), indent=2))
# Get dataloaders
train_loader, valid_loader, trainset = get_data_loaders(opts)
spk2acostats = trainset.spk2acostats
if opts.mulout:
# select only available speakers to load, not all
model_spks = list(trainset.speakers.keys())
else:
model_spks = list(trainset.all_speakers.keys())
opts.num_inputs = trainset.ling_feats_dim + 2
opts.spks = model_spks
# build a duration model ready to train
aco_model, train_fn_name, eval_fn_name = acoustic_builder(opts.model_type,
opts)
aco_model.describe_model()
#aco_model = acoustic_rnn(num_inputs=trainset.ling_feats_dim + 2,
# emb_size=opts.emb_size,
# rnn_size=opts.rnn_size,
# rnn_layers=opts.rnn_layers,
# dropout=opts.dout,
# speakers=model_spks,
# mulout=opts.mulout,
# cuda=opts.cuda,
# emb_layers=opts.emb_layers,
# emb_activation=opts.emb_activation)
criterion = getattr(nn, opts.loss)(size_average=True)
if opts.optim == 'Noam':
opti = get_std_opt(aco_model)
else:
opti = getattr(optim, opts.optim)(aco_model.parameters(),
lr=opts.lr)
device = 'cpu'
if opts.cuda and torch.cuda.is_available():
device = 'cuda'
if opts.cuda:
aco_model.to(device)
print('*' * 30)
print('Built acoustic model')
print(aco_model)
print('*' * 30)
patience = opts.patience
tr_opts = {'spk2acostats':spk2acostats,
'idx2spk':trainset.idx2spk}
va_opts = {'idx2spk':trainset.idx2spk}
if opts.mulout:
tr_opts['mulout'] = True
va_opts['mulout'] = True
if opts.model_type == 'decsatt':
tr_opts['decoder'] = True
va_opts['decoder'] = True
writer = SummaryWriter(os.path.join(opts.save_path,
'train'))
train_fn = getattr(musa.core, train_fn_name)
eval_fn = getattr(musa.core, eval_fn_name)
train_engine(aco_model, train_loader, opti, opts.log_freq, train_fn,
criterion, opts.epoch, opts.save_path, 'aco_model.ckpt',
tr_opts=tr_opts,
eval_fn=eval_fn, val_dloader=valid_loader,
eval_stats=spk2acostats,
eval_target='total_nosil_aco_mcd',
eval_patience=opts.patience,
cuda=opts.cuda,
va_opts=va_opts,
log_writer=writer)
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument('--cfg_spk', type=str, default='cfg/tcstar.cfg')
parser.add_argument('--spk_id', type=int, default=0)
parser.add_argument('--lab_dir', type=str, default='data/tcstar/lab')
parser.add_argument('--aco_dir', type=str, default='data/tcstar/aco')
parser.add_argument('--synthesize_lab', type=str, default=None,
help='Lab filename to be synthesized')
parser.add_argument('--codebooks_dir', type=str,
default='data/tcstar/codebooks.pkl')
parser.add_argument('--pf', type=float, default=1)
parser.add_argument('--save_path', type=str, default='dur_ckpt')
parser.add_argument('--force-gen', action='store_true',
default=False)
parser.add_argument('--force-dur', action='store_true',
default=False)
parser.add_argument('--max_samples', type=int, default=None,
help='Max samples per speaker in dur loader')
parser.add_argument('--rnn_size', type=int, default=256)
parser.add_argument('--rnn_layers', type=int, default=1)
parser.add_argument('--emb_size', type=int, default=256)
parser.add_argument('--emb_layers', type=int, default=1)
parser.add_argument('--q_classes', type=int, default=None,
help='Num of clusters in dur quantization. '
'If specified, this will triger '
'quantization in dloader and softmax '
'output for the model (Def: None).')
parser.add_argument('--model', type=str, default=None,
help='Trained dur model')
parser.add_argument('--loss', type=str, default='MSELoss',
help='Options: PyTorch losses (Def: MSELoss)')
parser.add_argument('--dout', type=float, default=0.5)
parser.add_argument('--batch_size', type=int, default=50)
parser.add_argument('--epoch', type=int, default=50)
parser.add_argument('--log_freq', type=int, default=25)
parser.add_argument('--patience', type=int, default=5)
parser.add_argument('--seed', type=int, default=1991)
parser.add_argument('--lr', type=float, default=0.001)
parser.add_argument('--optim', type=str, default='Adam',
help='Adam, RMSprop, SGD, (any of pytorch in optim'
'package) and Noam')
parser.add_argument('--emb_activation', type=str, default='Tanh')
parser.add_argument('--out_activation', type=str, default='Sigmoid')
parser.add_argument('--max_seq_len', type=int, default=None)
parser.add_argument('--loader_workers', type=int, default=2)
parser.add_argument('--parser_workers', type=int, default=4)
parser.add_argument('--cuda', default=False, action='store_true')
parser.add_argument('--mulout', default=False, action='store_true')
parser.add_argument('--exclude_train_spks', type=str, default=[], nargs='+')
parser.add_argument('--exclude_eval_spks', type=str, default=[], nargs='+')
parser.add_argument('--model_type', type=str, default='rnn',
help='types: (1) rnn, (2) satt')
parser.add_argument('--N', type=int, default=6)
parser.add_argument('--h', type=int, default=8)
parser.add_argument('--d_ff', type=int, default=2048)
parser.add_argument('--no_stateful', action='store_true', default=False)
parser.add_argument('--no_lnorm', action='store_true', default=False)
parser.add_argument('--conv_out', action='store_true', default=False)
opts = parser.parse_args()
print('Parsed opts: ', json.dumps(vars(opts), indent=2))
if not os.path.exists(opts.save_path):
os.makedirs(opts.save_path)
torch.manual_seed(opts.seed)
np.random.seed(opts.seed)
random.seed(opts.seed)
if opts.cuda:
torch.cuda.manual_seed(opts.seed)
main(opts)