-
Notifications
You must be signed in to change notification settings - Fork 1
/
trainer.py
221 lines (202 loc) · 6.94 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
"""Trainer
"""
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import os
import argparse
import importlib
import torch
import torch.nn as nn
from torch.utils.data import DataLoader
from tensorboardX import SummaryWriter
try:
from apex.parallel import DistributedDataParallel, convert_syncbn_model
from apex import amp
except ImportError:
raise ImportError(
"Please install apex from https://www.github.com/nvidia/apex to run this script."
)
import common.modes
import common.meters
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument(
'--dataset',
help='Dataset name',
default=None,
type=str,
)
parser.add_argument(
'--model',
help='Model name',
default=None,
type=str,
)
parser.add_argument(
'--job_dir',
help='GCS location to write checkpoints and export models',
required=True)
# Experiment arguments
parser.add_argument(
'--save_checkpoints_epochs',
help='Number of epochs to save checkpoint',
default=1,
type=int)
parser.add_argument(
'--train_epochs',
help='Number of epochs to run training totally',
default=10,
type=int)
parser.add_argument(
'--log_steps',
help='Number of steps for training logging',
default=100,
type=int)
parser.add_argument(
'--random_seed',
help='Random seed for TensorFlow',
default=None,
type=int)
# Performance tuning parameters
parser.add_argument(
'--opt_level',
help='Number of GPUs for experiments',
default='O2',
type=str)
parser.add_argument(
'--sync_bn',
default=False,
action='store_true',
help='Enabling apex sync BN.')
parser.add_argument("--local_rank", default=0, type=int)
parser.add_argument("--node_rank", default=0, type=int)
# Parse arguments
args, _ = parser.parse_known_args()
dataset_module = importlib.import_module(
'datasets.' + args.dataset if args.dataset else 'datasets')
dataset_module.update_argparser(parser)
model_module = importlib.import_module('models.' +
args.model if args.model else 'models')
model_module.update_argparser(parser)
params = parser.parse_args()
print(params)
torch.backends.cudnn.benchmark = True
params.distributed = False
params.master_proc = True
if 'WORLD_SIZE' in os.environ:
params.distributed = int(os.environ['WORLD_SIZE']) > 1
if params.distributed:
torch.cuda.set_device(params.local_rank)
torch.distributed.init_process_group(backend='nccl', init_method='env://')
if params.node_rank or params.local_rank:
params.master_proc = False
train_dataset = dataset_module.get_dataset(common.modes.TRAIN, params)
eval_dataset = dataset_module.get_dataset(common.modes.EVAL, params)
if params.distributed:
train_sampler = torch.utils.data.distributed.DistributedSampler(
train_dataset)
eval_sampler = torch.utils.data.distributed.DistributedSampler(eval_dataset)
eval_sampler = None
else:
train_sampler = None
eval_sampler = None
train_data_loader = DataLoader(
dataset=train_dataset,
num_workers=params.num_data_threads,
batch_size=params.train_batch_size,
shuffle=(train_sampler is None),
drop_last=True,
pin_memory=True,
sampler=train_sampler,
)
eval_data_loader = DataLoader(
dataset=eval_dataset,
num_workers=params.num_data_threads,
batch_size=params.eval_batch_size,
shuffle=False,
drop_last=False,
pin_memory=True,
sampler=eval_sampler,
)
model, criterion, optimizer, lr_scheduler, metrics = model_module.get_model_spec(
params)
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model = model.to(device)
criterion = criterion.to(device)
model, optimizer = amp.initialize(
model, optimizer, opt_level=params.opt_level, loss_scale='dynamic')
if os.path.exists(os.path.join(params.job_dir, 'latest.pth')):
checkpoint = torch.load(
os.path.join(params.job_dir, 'latest.pth'),
map_location=lambda storage, loc: storage.cuda())
model.load_state_dict(checkpoint['model_state_dict'])
optimizer.load_state_dict(checkpoint['optimizer_state_dict'])
latest_epoch = checkpoint['epoch']
print('Loaded checkpoint from epoch {}.'.format(latest_epoch))
else:
latest_epoch = 0
if params.distributed:
if params.sync_bn:
model = convert_syncbn_model(model)
model = DistributedDataParallel(model)
if params.master_proc:
writer = SummaryWriter(params.job_dir)
def train(epoch):
if params.distributed:
train_sampler.set_epoch(epoch)
lr_scheduler.step(epoch - 1)
loss_meter = common.meters.AverageMeter()
model.train()
for batch_idx, (data, target) in enumerate(train_data_loader):
data = data.to(device, non_blocking=True)
target = target.to(device, non_blocking=True)
optimizer.zero_grad()
output = model(data)
loss = criterion(output, target)
with amp.scale_loss(loss, optimizer) as scaled_loss:
scaled_loss.backward()
optimizer.step()
if batch_idx % params.log_steps == 0:
loss_meter.update(loss.item(), data.size(0))
print('Train Epoch: {} [{}/{} ({:.0f}%)]\tLoss: {:.6f}'.format(
epoch, batch_idx * len(data),
len(train_data_loader) * len(data),
100. * batch_idx / len(train_data_loader), loss.item()))
if params.master_proc:
writer.add_scalar('training_loss', loss_meter.avg, epoch)
def eval(epoch):
metric_meters = {}
for metric_name in metrics.keys():
metric_meters[metric_name] = common.meters.AverageMeter()
with torch.no_grad():
model.eval()
for data, target in eval_data_loader:
data = data.to(device, non_blocking=True)
target = target.to(device, non_blocking=True)
output = model(data)
for metric_name in metrics.keys():
metric_meters[metric_name].update(
metrics[metric_name](output, target).item(), data.size(0))
for metric_name in metrics.keys():
print('Eval set: Average {}: {:.4f}\n'.format(
metric_name, metric_meters[metric_name].avg))
writer.add_scalar(metric_name, metric_meters[metric_name].avg, epoch)
for epoch in range(latest_epoch + 1, params.train_epochs + 1):
train(epoch)
if epoch % params.save_checkpoints_epochs == 0:
if params.master_proc:
if not os.path.exists(params.job_dir):
os.makedirs(params.job_dir)
torch.save({
'epoch':
epoch,
'model_state_dict':
model.module.state_dict()
if params.distributed else model.state_dict(),
'optimizer_state_dict':
optimizer.state_dict(),
}, os.path.join(params.job_dir, 'latest.pth'))
eval(epoch)
if params.master_proc:
writer.close()