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
|
"""Training script for PyTorch models."""
from pathlib import Path
import time
from typing import Dict, Optional, Type
from loguru import logger
import numpy as np
import torch
from tqdm import tqdm, trange
from training.util import RunningAverage
import wandb
from text_recognizer.models import Model
torch.backends.cudnn.benchmark = True
np.random.seed(4711)
torch.manual_seed(4711)
torch.cuda.manual_seed(4711)
class Trainer:
"""Trainer for training PyTorch models."""
# TODO implement wandb.
# TODO implement Bayesian parameter search.
def __init__(
self,
model: Type[Model],
model_dir: Path,
epochs: int,
val_metric: str = "accuracy",
checkpoint_path: Optional[Path] = None,
use_wandb: Optional[bool] = False,
) -> None:
"""Initialization of the Trainer.
Args:
model (Type[Model]): A model object.
model_dir (Path): Path to the model directory.
epochs (int): Number of epochs to train.
val_metric (str): The validation metric to evaluate the model on. Defaults to "accuracy".
checkpoint_path (Optional[Path]): The path to a previously trained model. Defaults to None.
use_wandb (Optional[bool]): Sync training to wandb.
"""
self.model = model
self.model_dir = model_dir
self.epochs = epochs
self.checkpoint_path = checkpoint_path
self.start_epoch = 0
if self.checkpoint_path is not None:
self.start_epoch = self.model.load_checkpoint(self.checkpoint_path)
if use_wandb:
# TODO implement wandb logging.
pass
self.val_metric = val_metric
self.best_val_metric = 0.0
# Parse the name of the experiment.
experiment_dir = str(self.model_dir.parents[1]).split("/")
self.experiment_name = experiment_dir[-2] + "/" + experiment_dir[-1]
def train(self) -> None:
"""Training loop."""
# Set model to traning mode.
self.model.train()
# Running average for the loss.
loss_avg = RunningAverage()
data_loader = self.model.data_loaders("train")
with tqdm(
total=len(data_loader),
leave=False,
unit="step",
bar_format="{n_fmt}/{total_fmt} |{bar:20}| {remaining} {rate_inv_fmt}{postfix}",
) as t:
for data, targets in data_loader:
data, targets = (
data.to(self.model.device),
targets.to(self.model.device),
)
# Forward pass.
# Get the network prediction.
output = self.model.network(data)
# Compute the loss.
loss = self.model.criterion(output, targets)
# Backward pass.
# Clear the previous gradients.
self.model.optimizer.zero_grad()
# Compute the gradients.
loss.backward()
# Perform updates using calculated gradients.
self.model.optimizer.step()
# Compute metrics.
loss_avg.update(loss.item())
output = output.data.cpu()
targets = targets.data.cpu()
metrics = {
metric: self.model.metrics[metric](output, targets)
for metric in self.model.metrics
}
metrics["loss"] = loss_avg()
# Update Tqdm progress bar.
t.set_postfix(**metrics)
t.update()
# If the model has a learning rate scheduler, compute a step.
if self.model.lr_scheduler is not None:
self.model.lr_scheduler.step()
def validate(self) -> Dict:
"""Evaluation loop.
Returns:
Dict: A dictionary of evaluation metrics.
"""
# Set model to eval mode.
self.model.eval()
# Running average for the loss.
data_loader = self.model.data_loaders("val")
# Running average for the loss.
loss_avg = RunningAverage()
# Summary for the current eval loop.
summary = []
with tqdm(
total=len(data_loader),
leave=False,
unit="step",
bar_format="{n_fmt}/{total_fmt} |{bar:20}| {remaining} {rate_inv_fmt}{postfix}",
) as t:
for data, targets in data_loader:
data, targets = (
data.to(self.model.device),
targets.to(self.model.device),
)
with torch.no_grad():
# Forward pass.
# Get the network prediction.
output = self.model.network(data)
# Compute the loss.
loss = self.model.criterion(output, targets)
# Compute metrics.
loss_avg.update(loss.item())
output = output.data.cpu()
targets = targets.data.cpu()
metrics = {
metric: self.model.metrics[metric](output, targets)
for metric in self.model.metrics
}
metrics["loss"] = loss.item()
summary.append(metrics)
# Update Tqdm progress bar.
t.set_postfix(**metrics)
t.update()
# Compute mean of all metrics.
metrics_mean = {
metric: np.mean([x[metric] for x in summary]) for metric in summary[0]
}
metrics_str = " - ".join(f"{k}: {v}" for k, v in metrics_mean.items())
logger.debug(metrics_str)
return metrics_mean
def fit(self) -> None:
"""Runs the training and evaluation loop."""
logger.debug(f"Running an experiment called {self.experiment_name}.")
t_start = time.time()
# Run the training loop.
for epoch in trange(
self.epochs,
initial=self.start_epoch,
leave=False,
bar_format="{desc}: {n_fmt}/{total_fmt} |{bar:10}| {remaining}{postfix}",
desc="Epoch",
):
# Perform one training pass over the training set.
self.train()
# Evaluate the model on the validation set.
val_metrics = self.validate()
# The validation metric to evaluate the model on, e.g. accuracy.
val_metric = val_metrics[self.val_metric]
is_best = val_metric >= self.best_val_metric
self.best_val_metric = val_metric if is_best else self.best_val_metric
# Save checkpoint.
self.model.save_checkpoint(self.model_dir, is_best, epoch, self.val_metric)
if self.start_epoch > 0 and epoch + self.start_epoch == self.epochs:
logger.debug(f"Trained the model for {self.epochs} number of epochs.")
break
t_end = time.time()
t_training = t_end - t_start
logger.info(f"Training took {t_training:.2f} s.")
|
