-
Notifications
You must be signed in to change notification settings - Fork 1
/
train_nar.py
327 lines (266 loc) · 12.9 KB
/
train_nar.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
import os
import shutil
import argparse
import datetime
from pathlib import Path
from typing import Optional
import numpy as np
import torch
import torch.nn as nn
from torch.utils.data import DataLoader
from torch.utils.tensorboard import SummaryWriter
from torch.distributed import init_process_group
from accelerate import Accelerator
from accelerate.utils import DistributedDataParallelKwargs
from torch.optim.lr_scheduler import LambdaLR
from tqdm import tqdm
from model.nar import VoxInstructNAR
from utils.optimizer import get_optimizer, ScheduledOptim
from utils.dataset import VoxDataset
from utils.utils import (
get_config_from_file, sequence_mask, compute_loss, accum_log,
calculate_model_params, cycle, to_device
)
class NARModelTrainer(nn.Module):
"""
NARModelTrainer class for training the VoxInstructNAR model.
"""
def __init__(
self,
config_path: str,
save_path: str,
restore_path: Optional[str] = None,
log_type='tensorboard',
logging_dir='logs',
accelerate_kwargs: dict = dict(),
):
"""
Initialize the NARModelTrainer class.
Args:
config_path (str): Path to the configuration file.
save_path (str): Directory to save outputs.
restore_path (Optional[str]): Path to restore model checkpoint.
log_type (str): Type of logging to use.
logging_dir (str): Directory for logging.
accelerate_kwargs (dict): Additional arguments for Accelerator.
"""
super(NARModelTrainer, self).__init__()
hp = get_config_from_file(config_path).hparams
self.save_path = save_path
self.save_checkpoint_path = os.path.join(save_path, 'checkpoints')
os.makedirs(self.save_checkpoint_path, exist_ok=True)
# save config
shutil.copyfile(config_path, os.path.join(self.save_path, 'config.yaml'))
logging_dir = os.path.join(save_path, logging_dir)
kwargs = DistributedDataParallelKwargs(find_unused_parameters=True)
self.accelerator = Accelerator(kwargs_handlers=[kwargs],
log_with=log_type,
project_dir=logging_dir,
gradient_accumulation_steps=hp.grad_accum_every,
**accelerate_kwargs)
self.accelerator.print(save_path)
self.hp = hp
model = VoxInstructNAR(hp=hp)
calculate_model_params(model)
self.model = model.to(self.accelerator.device)
self.register_buffer('steps', torch.Tensor([0]))
self.max_step = hp.max_step
self.log_step = hp.log_step
self.ckpt_step = hp.ckpt_step
self.val_step = hp.val_step
self.batch_size = hp.batch_size
self.grad_accum_every = hp.grad_accum_every
self.max_grad_norm = hp.max_grad_norm
self.num_workers = hp.num_workers
# optimizers
self.optim = get_optimizer(
self.model.parameters(),
lr=hp.learning_rate,
wd=hp.weight_decay)
self.scheduler_func = ScheduledOptim(
warmup_steps=hp.warmup_step,
num_gpu=self.accelerator.num_processes)
self.scheduler = LambdaLR(self.optim, self.scheduler_func.get_lr_scale)
if restore_path is not None:
self.load(restore_path)
self.trainset = VoxDataset(hp.train_path, hp=hp)
self.valset = VoxDataset(hp.val_path, hp=hp, eval_mode=True)
self.train_loader = DataLoader(self.trainset,
num_workers=self.num_workers,
shuffle=True,
batch_size=self.batch_size,
collate_fn=self.trainset.collate_fn)
self.val_loader = DataLoader(self.valset,
num_workers=self.num_workers,
shuffle=False,
batch_size=self.batch_size,
collate_fn=self.valset.collate_fn)
# prepare with accelerator
(self.model, self.train_loader, self.optim, self.scheduler) = self.accelerator.prepare(self.model,
self.train_loader,
self.optim,
self.scheduler)
self.train_loader_iter = cycle(self.train_loader)
self.val_loader_iter = cycle(self.val_loader)
hps = {
"num_train_steps": self.max_step,
"batch_size": self.batch_size
}
self.accelerator.init_trackers(hp.name, config=hps)
def save(self, path, steps, not_save_optim=False):
if not_save_optim:
ckpt = dict(model=self.accelerator.get_state_dict(self.model),
scheduler=self.scheduler.state_dict(),
steps=int(steps))
else:
ckpt = dict(model=self.accelerator.get_state_dict(self.model),
optim=self.optim.state_dict(),
scheduler=self.scheduler.state_dict(),
steps=int(steps))
torch.save(ckpt, path)
def load(self, path):
path = Path(path)
assert path is not None
ckpt = torch.load(str(path), map_location='cpu')
model = self.accelerator.unwrap_model(self.model)
model.load_state_dict(ckpt['model'])
if 'steps' in ckpt:
self.steps = torch.Tensor([ckpt['steps']])
if 'scheduler' in ckpt:
self.scheduler.load_state_dict(ckpt['scheduler'])
if 'optim' in ckpt:
self.optim.load_state_dict(ckpt['optim'])
def print(self, msg):
self.accelerator.print(msg)
@property
def unwrapped_model(self):
return self.accelerator.unwrap_model(self.model)
@property
def device(self):
return self.accelerator.device
@property
def is_main(self):
return self.accelerator.is_main_process
@property
def is_local_main(self):
return self.accelerator.is_local_main_process
def train_step(self):
device = self.device
self.model.train()
logs = {}
self.optim.zero_grad()
start = datetime.datetime.now()
steps = int(self.steps.item())
for _ in range(self.grad_accum_every):
with self.accelerator.accumulate(self.model):
loaded_data = next(self.train_loader_iter)
text_ids, text_id_lens, seqs, seq_lens, full_seqs, segment_ids, basenames = to_device(loaded_data, device=device)
text_attn_mask = sequence_mask(text_id_lens, max_len=hp.max_text_len, device=device)
seq_attn_mask = sequence_mask(seq_lens, max_len=None, device=device)
logits, target_x, target_mask, act_mask_ratio = self.model(
full_seqs,
segment_ids=segment_ids,
attention_mask=seq_attn_mask,
text_ids=text_ids,
text_attn_mask=text_attn_mask,
)
# Compute loss and backpropagate
loss, acc = compute_loss(logits, target_x, mask=target_mask, compute_acc=True, topk=(1,10))
self.accelerator.backward(loss)
# Clip gradients if max_grad_norm is set
if self.max_grad_norm is not None:
grad_norm = self.accelerator.clip_grad_norm_(self.model.parameters(), self.max_grad_norm)
else:
grad_norm = None
accum_log(logs, {'loss': loss.item()})
accum_log(logs, {'acc_top1': acc[0].item()})
accum_log(logs, {'acc_top10': acc[1].item()})
accum_log(logs, {'grad_norm': grad_norm.item()})
accum_log(logs, {'samples': seqs.shape[0]})
accum_log(logs, {'min_len': seq_lens.min().item()})
accum_log(logs, {'max_len': seq_lens.max().item()})
accum_log(logs, {'avg_mask_ratio': act_mask_ratio.mean().item()})
# Step optimizer and scheduler after accumulation
self.optim.step()
self.scheduler.step()
logs['loss'] = logs['loss'] / hp.grad_accum_every
logs['acc_top1'] = logs['acc_top1'] / hp.grad_accum_every
logs['acc_top10'] = logs['acc_top10'] / hp.grad_accum_every
logs['min_len'] = logs['min_len'] / hp.grad_accum_every
logs['max_len'] = logs['max_len'] / hp.grad_accum_every
logs['avg_mask_ratio'] = logs['avg_mask_ratio'] / hp.grad_accum_every
# log
times = datetime.datetime.now() - start
if steps % self.log_step == 0:
self.print("NAR training! step: {}, loss: {:.4f}, acc_top1: {:4f}, acc_top10: {:4f}, lr: {}, samples: {}, min_len:{}, max_len: {}, avg_mask_ratio: {}, grad_norm:{:4f}, times:{}\n".format(\
steps, logs['loss'], logs['acc_top1'], logs['acc_top10'], self.scheduler.get_last_lr(), logs['samples'], logs['min_len'], logs['max_len'], logs['avg_mask_ratio'], logs['grad_norm'], times))
self.accelerator.log(
{
"Train/loss": logs['loss'],
"Train/acc_top1": logs['acc_top1'],
"Train/acc_top10": logs['acc_top10'],
"Train/lr": self.scheduler.get_last_lr()[0].item(),
"Train/avg_mask_ratio": logs["avg_mask_ratio"],
"Train/grad_norm": logs['grad_norm'],
},
step=steps)
if self.is_main and self.val_step > 0 and steps % self.val_step == 0:
eval_model = self.unwrapped_model
eval_model.eval()
val_loss = 0.
token_cnt = 0
all_acc_0 = 0.
all_acc_1 = 0.
for loaded_data in tqdm(self.val_loader):
text_ids, text_id_lens, seqs, seq_lens, full_seqs, segment_ids, basenames = to_device(loaded_data, device=device)
text_attn_mask = sequence_mask(text_id_lens, max_len=hp.max_text_len, device=device)
seq_attn_mask = sequence_mask(seq_lens, max_len=None, device=device)
with torch.no_grad():
logits, target_x, target_mask, avg_mask_ratio = self.model(
full_seqs,
segment_ids=segment_ids,
attention_mask=attention_mask,
text_ids = text_ids,
text_attn_mask=text_attn_mask,
)
# Compute loss
loss, acc = compute_loss(logits, target_x, mask=target_mask, compute_acc=True, topk=(1,10))
token_cnt += target_mask.sum().item()
val_loss += loss.item() * target_mask.sum().item()
all_acc_0 += acc[0].item() * target_mask.sum().item()
all_acc_1 += acc[1].item() * target_mask.sum().item()
val_loss = val_loss / token_cnt
all_acc_0 = all_acc_0 / token_cnt
all_acc_1 = all_acc_1 / token_cnt
self.accelerator.log({
"Val/loss": val_loss,
"Val/acc_top1": all_acc_0,
"Val/acc_top10": all_acc_1,
}, step=steps)
self.print("NAR validation! step: {}k, loss: {:.4f}, acc_top1: {:4f}, acc_top10: {:4f}\n".format(steps // 1000, val_loss, all_acc_0, all_acc_1))
self.accelerator.wait_for_everyone()
if self.is_main and steps % self.ckpt_step == 0:
model_path = os.path.join(self.save_checkpoint_path, "{}k_ckpt.pyt".format(steps // 1000))
self.save(model_path, steps, not_save_optim=True)
self.print(f'{steps}: saving model at {model_path}')
self.steps += 1
return logs
def train(self):
while self.steps < self.max_step:
logs = self.train_step()
self.print('training complete')
self.accelerator.end_training()
if __name__ == "__main__":
torch.distributed.init_process_group(backend="nccl", timeout=datetime.timedelta(seconds=3600))
parser = argparse.ArgumentParser()
parser.add_argument('-c', '--config', type=str, required=True, help='Config yaml')
parser.add_argument('--save_path', type=str, required=True, help='Path to save checkpoints')
parser.add_argument('--restore_path', default=None, type=str, help='restore checkpoints')
args = parser.parse_args()
hp = get_config_from_file(args.config).hparams
trainer = NARModelTrainer(
config_path=args.config,
save_path=args.save_path,
restore_path=args.restore_path,
)
trainer.train()