-
Notifications
You must be signed in to change notification settings - Fork 1
/
train.py
246 lines (197 loc) · 9.34 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
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
import os
import time
import builtins
from tqdm import tqdm
import torch
import torch.nn as nn
import torchvision
from utils.data_loader import load_data_loaders
from utils.criterion import get_criterion
from utils.optimizers import adjust_learning_rate
from utils.optimizers import load_optimizer
from utils.utils import AverageMeterCollection
from utils.utils import compute_performance
from utils.utils import reduce_tensor
from torch.utils.tensorboard import SummaryWriter
from torch.cuda.amp import GradScaler, autocast
# Training class
class Trainer(object):
def __init__(self, args):
self.args = args
# Mute the other workers
if not self.is_primary_worker():
def print_pass(*args):
pass
builtins.print = print_pass
# Set training and validation data loaders
self.train_loader, self.val_loader, \
self.args.num_classes, self.args.train_len, self.args.val_len = \
load_data_loaders(dataset=self.args.dataset,
args=self.args)
# Set backbone network
if hasattr(torchvision.models, args.arch):
if args.finetune:
self.model = getattr(torchvision.models, args.arch)(pretrained=True)
# Replace FC layer
for fc_name, fc_layer in self.model.named_modules():
if type(fc_layer) is nn.Linear:
break
setattr(self.model, fc_name, nn.Linear(fc_layer.in_features, args.num_classes))
else:
self.model = getattr(torchvision.models, args.arch)(num_classes=args.num_classes)
else:
raise ValueError(
f"Not supported model architecture {args.arch}")
if self.is_primary_worker():
print(self.model)
self.writer = SummaryWriter(log_dir=os.path.join(args.output, 'logs'))
# Set criterion and opimizer
self.criterion = get_criterion(self.args)
self.optimizer = load_optimizer(self.args, self.model)
if self.is_primary_worker():
print(self.optimizer)
# Distributed data parallel
if self.args.distributed:
torch.cuda.set_device(device=self.args.gpu_no)
self.model.cuda(device=self.args.gpu_no)
self.model = torch.nn.parallel.DistributedDataParallel(
module=self.model,
device_ids=[self.args.gpu_no])
else:
self.model = torch.nn.DataParallel(self.model).cuda()
self.criterion.cuda(device=self.args.gpu_no)
torch.backends.cudnn.benchmark = True
# Training on one mini-batch
def train_minibatch(self, iteration, batch, epoch, batch_start_time,
scaler, meters):
image, target = batch
batch_size = image.size(0)
current_lr = adjust_learning_rate(
optimizer=self.optimizer,
epoch=epoch,
iteration=iteration,
lr_decay_type=self.args.lr_decay_type,
epochs=self.args.epochs,
train_len=self.args.train_len)
target = target.cuda()
image = image.cuda()
# forward
with autocast(): # mixed precision
output = self.model(image).float()
# loss
loss = self.criterion(output, target)
# backward
self.optimizer.zero_grad()
scaler.scale(loss).backward()
scaler.step(self.optimizer)
scaler.update()
self.optimizer.zero_grad()
# print intermediate results
if (iteration + 1) % self.args.print_freq == 0:
meters = compute_performance(output=output,
target=target,
meters=meters,
world_size=self.args.world_size)
reduced_loss = reduce_tensor(loss.data,
self.args.world_size)
if self.args.distributed:
torch.cuda.synchronize()
meters.get('losses').update(reduced_loss.item(), batch_size)
meters.get('batch_time').update( (time.time() - batch_start_time) )
return meters, current_lr
# Training on one epoch
def train(self, epoch):
meters = AverageMeterCollection('batch_time', 'losses', 'mapcls', 'mapsmp')
# Init trainer
self.model.train()
self.optimizer.zero_grad()
scaler = GradScaler()
# Init timer
tic = time.time()
batch_start_time = time.time()
if self.is_primary_worker():
tqdm_batch = tqdm(total=len(self.train_loader), desc="[Epoch {}]".format(epoch))
# Train over whole mini-batches
for iteration, batch in enumerate(self.train_loader):
meters, lr = self.train_minibatch(batch=batch, iteration=iteration, epoch=epoch,
batch_start_time=batch_start_time,
scaler=scaler, meters=meters)
batch_start_time = time.time()
if self.is_primary_worker():
tqdm_batch.set_postfix({'Time': '{batch_time.val:.2f} ({batch_time.avg:.2f})'.format(
batch_time=meters.batch_time),
'Loss': '{loss.val:.3f} ({loss.avg:.3f})'.format(loss=meters.losses),
'mAP@cls': '{map.val:.2f} ({map.avg:.2f})'.format(map=meters.mapcls),
'mAP@sample': '{map.val:.2f} ({map.avg:.2f})'.format(map=meters.mapsmp) })
tqdm_batch.update()
if self.is_primary_worker():
tqdm_batch.close()
time_spent = time.time() - tic
print('[Epoch {}] {:.3f} sec/epoch'.format(epoch, time_spent), end='\t')
print('remaining time: {:.3f} hours'.format( (self.args.epochs - epoch - 1) * time_spent / 3600))
self.writer.add_scalar('LearningRate', lr, epoch)
self.writer.add_scalar('Loss/train', meters.losses.avg, epoch)
self.writer.add_scalar('mAP_class/train', meters.mapcls.avg, epoch)
self.writer.add_scalar('mAP_sample/train', meters.mapsmp.avg, epoch)
self.writer.close()
# Validation
def validate(self, epoch=0):
meters = AverageMeterCollection('mapcls', 'mapsmp')
# Init
self.model.eval()
# Validate over whole mini-batches
output, target = [], []
for iteration, (image, target_) in enumerate(self.val_loader):
image = image.cuda()
target_ = target_.cuda()
# forward pass and compute loss
with torch.no_grad():
output_ = self.model(image)
output.append(output_)
target.append(target_)
output = torch.cat(output)
target = torch.cat(target)
meters = compute_performance(output=output,
target=target,
meters=meters,
world_size = self.args.world_size)
if self.is_primary_worker():
print('Test: mAP@cls {mapcls.avg:.3f}, mAP@sample {mapsmp.avg:.3f}'
.format(mapcls=meters.mapcls, mapsmp=meters.mapsmp))
self.writer.add_scalar('mAP_class/test', meters.mapcls.avg, epoch)
self.writer.add_scalar('mAP_sample/test', meters.mapsmp.avg, epoch)
self.writer.close()
self.save_checkpoint(epoch=epoch, meters=meters)
# Check primary worker or not
def is_primary_worker(self):
return not self.args.distributed or \
(self.args.distributed and
(self.args.rank % self.args.ngpus_per_node) == 0)
# Load checkpoint
def load_checkpoint(self, weight_file):
if os.path.isfile(weight_file):
print(f"=> loading checkpoint '{weight_file}'")
checkpoint = torch.load(weight_file)
if 'state_dict' in checkpoint:
checkpoint = checkpoint['state_dict']
self.model.load_state_dict(checkpoint)
print(f"=> checkpoint loaded '{weight_file}'")
else:
raise Exception(f"=> no checkpoint found at '{weight_file}'")
# Save checkpoint
def save_checkpoint(self, epoch, meters):
if self.is_primary_worker():
save_dict = {
'epoch': epoch + 1,
'arch': self.args.arch,
'state_dict': self.model.module.state_dict() if hasattr(self.model, 'module')
else self.model.state_dict(),
'mAPclass': meters.mapcls.avg,
'mAPsample': meters.mapsmp.avg,
'optimizer': self.optimizer.state_dict(),
}
checkpoint_dir = os.path.join(self.args.output, 'checkpoints')
if not os.path.exists(checkpoint_dir):
os.makedirs(checkpoint_dir)
filepath = f"{checkpoint_dir}/checkpoint-{self.args.arch}-last.pth"
torch.save(save_dict, filepath)