-
Notifications
You must be signed in to change notification settings - Fork 7
/
train.py
218 lines (190 loc) · 8.36 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
#!/usr/bin/env python
import os
import json
import time
import argparse
from tqdm import tqdm
from collections import defaultdict
import math
import argparse
import importlib
# torchim:
import torch
from torch import nn
from torch.utils.data import DataLoader, Subset, ConcatDataset
# from tensorboardX import SummaryWriter
import numpy as np
import pytorch_warmup as warmup
# data:
import data
from data.collate import collate_fn, gpu_collate, no_pad_collate
from data.transforms import (
Compose, AddLengths, AudioSqueeze, TextPreprocess,
MaskSpectrogram, ToNumpy, BPEtexts, MelSpectrogram,
ToGpu, Pad, NormalizedMelSpectrogram
)
import youtokentome as yttm
import torchaudio
from audiomentations import (
TimeStretch, PitchShift, AddGaussianNoise
)
from functools import partial
# model:
from model import configs as quartznet_configs
from model.quartznet import QuartzNet
# utils:
import yaml
from easydict import EasyDict as edict
from utils import fix_seeds, remove_from_dict, prepare_bpe
import wandb
from decoder import GreedyDecoder, BeamCTCDecoder
# TODO: wrap to trainer class
def train(config):
fix_seeds(seed=config.train.get('seed', 42))
dataset_module = importlib.import_module(f'.{config.dataset.name}', data.__name__)
bpe = prepare_bpe(config)
transforms_train = Compose([
TextPreprocess(),
ToNumpy(),
BPEtexts(bpe=bpe, dropout_prob=config.bpe.get('dropout_prob', 0.05)),
AudioSqueeze(),
AddGaussianNoise(
min_amplitude=0.001,
max_amplitude=0.015,
p=0.5
),
TimeStretch(
min_rate=0.8,
max_rate=1.25,
p=0.5
),
PitchShift(
min_semitones=-4,
max_semitones=4,
p=0.5
)
# AddLengths()
])
batch_transforms_train = Compose([
ToGpu('cuda' if torch.cuda.is_available() else 'cpu'),
NormalizedMelSpectrogram(
sample_rate=config.dataset.get('sample_rate', 16000),
n_mels=config.model.feat_in,
normalize=config.dataset.get('normalize', None)
).to('cuda' if torch.cuda.is_available() else 'cpu'),
MaskSpectrogram(
probability=0.5,
time_mask_max_percentage=0.05,
frequency_mask_max_percentage=0.15
),
AddLengths(),
Pad()
])
transforms_val = Compose([
TextPreprocess(),
ToNumpy(),
BPEtexts(bpe=bpe),
AudioSqueeze()
])
batch_transforms_val = Compose([
ToGpu('cuda' if torch.cuda.is_available() else 'cpu'),
NormalizedMelSpectrogram(
sample_rate=config.dataset.get('sample_rate', 16000), # for LJspeech
n_mels=config.model.feat_in,
normalize=config.dataset.get('normalize', None)
).to('cuda' if torch.cuda.is_available() else 'cpu'),
AddLengths(),
Pad()
])
# load datasets
train_dataset = dataset_module.get_dataset(config, transforms=transforms_train, part='train')
val_dataset = dataset_module.get_dataset(config, transforms=transforms_val, part='val')
train_dataloader = DataLoader(train_dataset, num_workers=config.train.get('num_workers', 4),
batch_size=config.train.get('batch_size', 1), collate_fn=no_pad_collate)
val_dataloader = DataLoader(val_dataset, num_workers=config.train.get('num_workers', 4),
batch_size=1, collate_fn=no_pad_collate)
model = QuartzNet(
model_config=getattr(quartznet_configs, config.model.name, '_quartznet5x5_config'),
**remove_from_dict(config.model, ['name'])
)
print(model)
optimizer = torch.optim.Adam(model.parameters(), **config.train.get('optimizer', {}))
num_steps = len(train_dataloader) * config.train.get('epochs', 10)
lr_scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(optimizer, T_max=num_steps)
# warmup_scheduler = warmup.UntunedLinearWarmup(optimizer)
if config.train.get('from_checkpoint', None) is not None:
model.load_weights(config.train.from_checkpoint)
if torch.cuda.is_available():
model = model.cuda()
criterion = nn.CTCLoss(blank=0, reduction='mean', zero_infinity=True)
# criterion = nn.CTCLoss(blank=config.model.vocab_size)
decoder = GreedyDecoder(bpe=bpe)
prev_wer = 1000
wandb.init(project=config.wandb.project, config=config)
wandb.watch(model, log="all", log_freq=config.wandb.get('log_interval', 5000))
for epoch_idx in tqdm(range(config.train.get('epochs', 10))):
# train:
model.train()
for batch_idx, batch in enumerate(train_dataloader):
batch = batch_transforms_train(batch)
optimizer.zero_grad()
logits = model(batch['audio'])
output_length = torch.ceil(batch['input_lengths'].float() / model.stride).int()
loss = criterion(logits.permute(2, 0, 1).log_softmax(dim=2), batch['text'], output_length, batch['target_lengths'])
loss.backward()
torch.nn.utils.clip_grad_norm_(model.parameters(), config.train.get('clip_grad_norm', 15))
optimizer.step()
lr_scheduler.step()
# warmup_scheduler.dampen()
if batch_idx % config.wandb.get('log_interval', 5000) == 0:
target_strings = decoder.convert_to_strings(batch['text'])
decoded_output = decoder.decode(logits.permute(0, 2, 1).softmax(dim=2))
wer = np.mean([decoder.wer(true, pred) for true, pred in zip(target_strings, decoded_output)])
cer = np.mean([decoder.cer(true, pred) for true, pred in zip(target_strings, decoded_output)])
step = epoch_idx * len(train_dataloader) * train_dataloader.batch_size + batch_idx * train_dataloader.batch_size
wandb.log({
"train_loss": loss.item(),
"train_wer": wer,
"train_cer": cer,
"train_samples": wandb.Table(
columns=['gt_text', 'pred_text'],
data=zip(target_strings, decoded_output)
)
}, step=step)
# validate:
model.eval()
val_stats = defaultdict(list)
for batch_idx, batch in enumerate(val_dataloader):
batch = batch_transforms_val(batch)
with torch.no_grad():
logits = model(batch['audio'])
output_length = torch.ceil(batch['input_lengths'].float() / model.stride).int()
loss = criterion(logits.permute(2, 0, 1).log_softmax(dim=2), batch['text'], output_length, batch['target_lengths'])
target_strings = decoder.convert_to_strings(batch['text'])
decoded_output = decoder.decode(logits.permute(0, 2, 1).softmax(dim=2))
wer = np.mean([decoder.wer(true, pred) for true, pred in zip(target_strings, decoded_output)])
cer = np.mean([decoder.cer(true, pred) for true, pred in zip(target_strings, decoded_output)])
val_stats['val_loss'].append(loss.item())
val_stats['wer'].append(wer)
val_stats['cer'].append(cer)
for k, v in val_stats.items():
val_stats[k] = np.mean(v)
val_stats['val_samples'] = wandb.Table(columns=['gt_text', 'pred_text'], data=zip(target_strings, decoded_output))
wandb.log(val_stats, step=step)
# save model, TODO: save optimizer:
if val_stats['wer'] < prev_wer:
os.makedirs(config.train.get('checkpoint_path', 'checkpoints'), exist_ok=True)
prev_wer = val_stats['wer']
torch.save(
model.state_dict(),
os.path.join(config.train.get('checkpoint_path', 'checkpoints'), f'model_{epoch_idx}_{prev_wer}.pth')
)
wandb.save(os.path.join(config.train.get('checkpoint_path', 'checkpoints'), f'model_{epoch_idx}_{prev_wer}.pth'))
if __name__ == '__main__':
parser = argparse.ArgumentParser(description='Training model.')
parser.add_argument('--config', default='configs/train_LJSpeech.yml',
help='path to config file')
args = parser.parse_args()
with open(args.config, 'r') as f:
config = edict(yaml.safe_load(f))
train(config)