-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathtrainer.py
201 lines (153 loc) · 6.48 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
"""
loss.py
Mar 4 2023
Gabriel Moreira
"""
import os
import torch
import numpy as np
from tqdm import tqdm
from tracker import Tracker
class Trainer:
def __init__(
self,
model,
epochs,
optimizer,
scheduler,
train_loss,
val_loss,
train_loader,
val_loader,
val_freq: int,
best_on: str,
device: str,
name: str,
resume: bool):
self.model = model
self.epochs = epochs
self.optimizer = optimizer
self.scheduler = scheduler
self.train_loss = train_loss
self.val_loss = val_loss
self.train_loader = train_loader
self.val_loader = val_loader
self.val_freq = val_freq
self.device = device
self.name = name
self.start_epoch = 1
self.best_on = best_on
self.tracker = Tracker(os.path.join('./experiments', name), load=resume)
self.scaler = torch.cuda.amp.GradScaler()
if resume:
self.resume_checkpoint()
def fit(self):
"""
Fit model to training set over #epochs
"""
is_best = False
for epoch in range(self.start_epoch, self.epochs+1):
train_out = self.train_epoch(epoch)
if epoch == 1 or epoch % self.val_freq == 0:
val_out = self.validate_epoch()
self.epoch_verbose(epoch, **train_out, **val_out)
is_best = self.tracker.is_better(self.best_on, val_out[self.best_on])
if is_best:
self.save_best()
self.save_checkpoint(epoch)
self.tracker.update(epoch=epoch,
lr=self.optimizer.param_groups[0]['lr'],
**train_out,
**val_out)
self.scheduler.step()
def train_epoch(self, epoch):
self.model.train()
batch_bar = tqdm(total=len(self.train_loader), dynamic_ncols=True, desc='Train')
total_loss = 0.0
total_correct = 0
total = 0
out = {}
for i_batch, batch_dict in enumerate(self.train_loader):
batch_data = batch_dict['data'].to(self.device)
self.optimizer.zero_grad()
with torch.cuda.amp.autocast():
batch_features = self.model(batch_data)
if 'target' in batch_dict:
batch_target = batch_dict['target'].to(self.device)
loss_batch = self.train_loss(batch_features, batch_target)
total_loss += loss_batch.detach()
out['train_loss'] = float(total_loss / len(self.train_loader))
tc, t = self.train_loss.scores()
total_correct += tc
total += t
out['train_acc'] = float(total_correct / total)
batch_bar.set_postfix(
loss="{:1.5e}".format(out['train_loss']),
acc="{:.4f}".format(out['train_acc']),
lr="{:1.2e}".format(float(self.optimizer.param_groups[0]['lr'])))
batch_bar.update()
else:
loss_batch = self.train_loss(batch_features)
total_loss += loss_batch.detach()
out['train_loss'] = float(total_loss / len(self.train_loader))
batch_bar.set_postfix(
loss="{:1.5e}".format(total_loss / (i_batch + 1)),
lr="{:1.2e}".format(float(self.optimizer.param_groups[0]['lr'])))
batch_bar.update()
self.scaler.scale(loss_batch).backward()
self.scaler.step(self.optimizer)
self.scaler.update()
batch_bar.close()
return out
@torch.no_grad()
def validate_epoch(self):
self.model.eval()
total_loss = 0.0
total_correct = 0
total = 0
out = {}
for i_batch, batch_dict in enumerate(tqdm(self.val_loader)):
batch_data = batch_dict['data'].to(self.device)
batch_features = self.model(batch_data)
if 'target' in batch_dict.keys():
batch_target = batch_dict['target'].to(self.device)
loss_batch = self.val_loss(batch_features, batch_target)
tc, t = self.val_loss.scores()
total_correct += tc
total += t
out['val_acc'] = float(total_correct / total)
else:
loss_batch = self.val_loss(batch_features)
total_loss += loss_batch.detach()
out['val_loss'] = float(total_loss / len(self.val_loader))
return out
def save_checkpoint(self, epoch):
checkpoint = {"epoch" : epoch,
"model" : self.model.state_dict(),
"optimizer" : self.optimizer,
"scheduler" : self.scheduler,
"scaler" : self.scaler}
# Save checkpoint to resume training later
checkpoint_path = os.path.join('./experiments', self.name, 'checkpoint.pt')
torch.save(checkpoint, checkpoint_path)
print('Checkpoint saved: {}'.format(checkpoint_path))
def save_best(self):
# Save best model weights
best_path = os.path.join('./experiments/', self.name, "best_weights.pt")
torch.save(self.model.state_dict(), best_path)
print("Saving best model: {}".format(best_path))
def resume_checkpoint(self):
resume_path = os.path.join('./experiments/', self.name, 'checkpoint.pt')
print('Loading checkpoint: {} ...'.format(resume_path))
checkpoint = torch.load(resume_path)
self.start_epoch = checkpoint['epoch'] + 1
self.optimizer = checkpoint['optimizer']
self.scheduler = checkpoint['scheduler']
self.scaler = checkpoint['scaler']
self.model.load_state_dict(checkpoint["model"])
print('Checkpoint loaded. Resume training from epoch {}'.format(self.start_epoch))
def epoch_verbose(self, epoch, **kwargs):
log = "\nEpoch: {}/{} summary:".format(epoch, self.epochs)
for k, v in kwargs.items():
log += "\n {} | {:1.6e}".format(k,v)
print(log)