-
Notifications
You must be signed in to change notification settings - Fork 0
/
trainer.py
299 lines (253 loc) · 13.4 KB
/
trainer.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
import os
import torch
import torch.nn as nn
import torch.nn.utils as nn_utils
import wandb
from tqdm import tqdm
from typing import Optional, Callable, List, Dict, Any, Union
from torch.cuda.amp import GradScaler, autocast
from torch.optim.optimizer import Optimizer
from torch.optim.lr_scheduler import _LRScheduler
class WarmupScheduler(_LRScheduler):
"""Cussom learning rate scheduler with linear warmup."""
def __init__(self, optimizer: Optimizer, warmup_steps: int, after_scheduler: _LRScheduler, last_epoch: int = -1):
self.warmup_steps = warmup_steps
self.after_scheduler = after_scheduler
self.finished_warmup = False
super(WarmupScheduler, self).__init__(optimizer, last_epoch)
def get_lr(self):
if self.last_epoch < self.warmup_steps:
warmup_factor = float(self.last_epoch) / float(max(1, self.warmup_steps))
return [base_lr * warmup_factor for base_lr in self.base_lrs]
else:
if not self.finished_warmup:
self.after_scheduler.base_lrs = [group['lr'] for group in self.optimizer.param_groups]
self.finished_warmup = True
return self.after_scheduler.get_last_lr()
def step(self, epoch=None):
if self.finished_warmup:
if epoch is None:
self.after_scheduler.step(None)
else:
self.after_scheduler.step(epoch - self.warmup_steps)
else:
return super(WarmupScheduler, self).step(epoch)
class Trainer:
def __init__(self,
device: str,
default_root_dir: str,
optimizer: torch.optim.Optimizer,
scheduler: Optional[torch.optim.lr_scheduler._LRScheduler] = None,
compute_metrics: Optional[Callable[[Dict[str, torch.Tensor], Dict[str, torch.Tensor]], Dict[str, float]]] = None,
logger: Optional[str] = None,
log: bool = False,
max_epochs: int = 500,
gradient_accumulation_steps: int = 1,
warmup_steps: Optional[int] = None,
save_every_n_steps: Optional[int] = None,
val_check_interval: Optional[int] = None,
evaluate_first: Optional[bool] = False,
max_grad_norm: Optional[float] = None,
use_mixed_precision: bool = False,
project_name: str = "my-awesome-project"):
"""
Initialize the Trainer class.
Args:
device (str): The device to use for training ('cpu' or 'cuda').
default_root_dir (str): The default root directory to save model weights.
optimizer (torch.optim.Optimizer): The optimizer for the model.
scheduler (torch.optim.lr_scheduler._LRScheduler, optional): The learning rate scheduler.
compute_metrics (Callable[[Dict[str, torch.Tensor], Dict[str, torch.Tensor]], Dict[str, float]], optional): A function to compute metrics.
continue_training (bool, optional): Whether to continue training from the last checkpoint.
logger (str, optional): The logger to use for logging ('wandb' or 'tensorboard').
log (bool, optional): Whether to log the training.
max_epochs (int, optional): The maximum number of epochs to train the model.
evaluate_first (bool, optional): Whether to evaluate the model before training.
gradient_accumulation_steps (int, optional): The number of gradient accumulation steps.
max_grad_norm (float, optional): The maximum gradient norm for gradient clipping.
use_mixed_precision (bool, optional): Whether to use mixed precision training.
project_name (str, optional): The name of the project for logging.
"""
self.device = device
self.default_root_dir = default_root_dir
self.optimizer = optimizer
self.scheduler = WarmupScheduler(optimizer, warmup_steps=warmup_steps, after_scheduler=scheduler) if warmup_steps!=None else scheduler
self.compute_metrics = compute_metrics
self.log = log
self.max_epochs = max_epochs
self.save_every_n_steps = save_every_n_steps
self.val_check_interval = val_check_interval
self.gradient_accumulation_steps = gradient_accumulation_steps
self.max_grad_norm = max_grad_norm
self.use_mixed_precision = use_mixed_precision
self.evaluate_first = evaluate_first
self.warmup_steps = warmup_steps
self.scaler = GradScaler() if self.use_mixed_precision else None
self.project_name = project_name
self.epoch = 0
self.metrics = {}
self.best_val_metrics = None
self._start_epoch = 0
self._has_test = False
self._model_name = None
assert logger in ["wandb", "tensorboard", None], "Invalid logger. Choose between 'wandb' and 'tensorboard'."
assert self.device in ["cpu", "cuda"], "Invalid device. Choose between 'cpu' and 'cuda'."
assert not (self.log and logger is None), "You need to specify a logger if you want to log the training."
if self.log:
self.logger = self._initialize_logger(logger)
else:
self.logger = None
def _initialize_logger(self, logger: str):
if logger == "wandb":
try:
wandb.init(
project=self.project_name,
config={
"learning_rate": self.optimizer.defaults['lr'],
"architecture": "CustomModel",
"dataset": "YourDataset",
"epochs": self.max_epochs,
}
)
return wandb
except ImportError:
self.log = False
raise ImportError("You need to install wandb to use it as a logger.")
elif logger == "tensorboard":
raise NotImplementedError("Tensorboard logging is not implemented yet.")
return None
def _save_checkpoint(self, model, optimizer, epoch, scheduler=None):
"""Save the model checkpoint."""
model_name = model.__class__.__name__
checkpoint = {
'epoch': epoch,
'model_state_dict': model.state_dict(),
'optimizer_state_dict': optimizer.state_dict(),
}
if scheduler:
checkpoint['scheduler_state_dict'] = scheduler.state_dict()
os.makedirs(os.path.join(self.default_root_dir, 'weights'), exist_ok=True)
torch.save(checkpoint, os.path.join(self.default_root_dir, 'weights', f'{model_name}_{epoch}.model'))
def _train_one_epoch(self, model, trainloader, testloader):
"""Train the model for one epoch."""
model.train()
training_loss = 0.0
accumulated_loss = 0.0
total_steps = len(trainloader) // self.gradient_accumulation_steps
with tqdm(total=total_steps) as pbar:
for i, data in enumerate(trainloader):
# Move data to the device
if isinstance(data, list):
data = [item.to(self.device) for item in data]
elif isinstance(data, dict):
data = {key: value.to(self.device) for key, value in data.items()}
else:
data = data.to(self.device)
# Forward pass
with autocast(enabled=self.use_mixed_precision, dtype=torch.float16):
output = model.train_step(data)
loss = output[0] if isinstance(output, tuple) else output
loss /= self.gradient_accumulation_steps
# Backward pass
if self.use_mixed_precision:
self.scaler.scale(loss).backward()
else:
loss.backward()
real_loss = loss.item() * self.gradient_accumulation_steps
accumulated_loss += real_loss
# Update the weights
if (i + 1) % self.gradient_accumulation_steps == 0:
if self.use_mixed_precision:
self.scaler.unscale_(self.optimizer)
if self.max_grad_norm is not None:
nn_utils.clip_grad_norm_(model.parameters(), self.max_grad_norm)
self.scaler.step(self.optimizer)
self.scaler.update()
else:
if self.max_grad_norm is not None:
nn_utils.clip_grad_norm_(model.parameters(), self.max_grad_norm)
self.optimizer.step()
self.optimizer.zero_grad()
pbar.update(1)
if self.scheduler:
self.scheduler.step()
if self.log:
self._log_metrics({"Training Loss": real_loss,
"Learning Rate": self.optimizer.param_groups[0]['lr']})
pbar.set_postfix({'Training Loss': real_loss})
#if (self.epoch + i+1) % self.val_check_interval == 0:
# if self._has_test:
# test_loss, test_metrics = self.test(model, testloader)
# self._log_metrics({"Test Loss": test_loss})
# self._log_metrics(test_metrics)
training_loss = accumulated_loss / len(trainloader)
self.optimizer.zero_grad()
return training_loss
def test(self, model, dataloader):
"""Test the model for one epoch."""
model.eval()
total_loss = 0.0
total_metrics = {}
with torch.no_grad():
with tqdm(total=len(dataloader)) as pbar:
for i, data in enumerate(dataloader):
# Move data to the device
if isinstance(data, list):
data = [item.to(self.device) for item in data]
elif isinstance(data, dict):
data = {key: value.to(self.device) for key, value in data.items()}
else:
data = data.to(self.device)
# Forward pass
with autocast(enabled=self.use_mixed_precision, dtype=torch.float16):
output = model.test_step(data)
loss = output[0] if isinstance(output, tuple) else output # Loss is the first element in the output tuple
total_loss += loss.item() # Accumulate the loss
# Compute metrics
metrics = self.compute_metrics(data, output) if self.compute_metrics else {}
for key, value in metrics.items():
if key not in total_metrics:
total_metrics[key] = 0.0
total_metrics[key] += value
pbar.update(1)
# Update progress bar with current average loss
avg_loss = total_loss / (i + 1)
avg_metrics = {key: total_metrics[key] / (i + 1) for key in total_metrics}
pbar.set_postfix({'Test Loss': avg_loss})
# Calculate average loss and metrics for the epoch
avg_loss = total_loss / len(dataloader)
avg_metrics = {key: total_metrics[key] / len(dataloader) for key in total_metrics}
# add prefix to metrics
avg_metrics = {f"Test {key}": value for key, value in avg_metrics.items()}
return avg_loss, avg_metrics
def _log_metrics(self, metrics: Dict[str, float]):
if self.logger:
self.logger.log(metrics)
def train(self, model, train_dataloader, test_dataloader=None):
batch_size = train_dataloader.batch_size
steps_per_epoch = len(train_dataloader) // self.gradient_accumulation_steps
total_steps = steps_per_epoch * self.max_epochs
if test_dataloader:
self._has_test = True
model.to(self.device)
self._model_name = model.__class__.__name__
if self.evaluate_first:
first_train_loss, _ = self.test(model, train_dataloader)
self._log_metrics({"Training Loss per epoch": first_train_loss})
first_test_loss, first_test_metrics = self.test(model, test_dataloader)
self._log_metrics({"Test Loss": first_test_loss})
self._log_metrics(first_test_metrics)
for epoch in range(self._start_epoch, self.max_epochs):
print("Epoch:", epoch)
train_loss = self._train_one_epoch(model, train_dataloader, test_dataloader)
self._log_metrics({"Training Loss per epoch": train_loss})
self.epoch += 1
if self._has_test:
test_loss, test_metrics = self.test(model, test_dataloader)
self._log_metrics({"Test Loss": test_loss})
self._log_metrics(test_metrics)
self._save_checkpoint(model, self.optimizer, epoch, self.scheduler)
model_name = model.__class__.__name__
torch.save(model.state_dict(), os.path.join(self.default_root_dir, 'weights', f'{model_name}_final.model'))
if self.logger == "wandb" and self.log:
self.logger.finish()