-
Notifications
You must be signed in to change notification settings - Fork 1
/
train.py
172 lines (159 loc) · 6.52 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
# standard modules
import os
import sys
import time
import random
import numpy as np
# torch modules
import torch
import torch.optim as optim
import torch.nn.functional as F
from torch.utils.data import DataLoader
from tensorboardX import SummaryWriter
from torchsummary import summary
# custom modules
from net.utils import *
from net.model import get_model
from net.dataset import get_dataset
from config import get_config
PATH = os.getcwd()
EXP_ID = sys.argv[1] # 1st param, experiment ID
MODEL = EXP_ID.split('-')[0] # model of experiment
CONFIG = get_config(EXP_ID) # get model configurations
class Trainer():
def __init__(self, config):
self.config = config
# init trainer
self.best_pred = 0.0
self.start_time = time.time()
self.path = os.path.join(PATH, 'results', MODEL, EXP_ID)
if not os.path.isdir(self.path):
os.makedirs(self.path)
self.init_seed(self.config['seed'])
print('[Seed]: %s' % self.config['seed'])
print('[Path]: %s' % self.path)
print('[Config]: %s\n' % self.config)
# init dataloader
trainset, valset = get_dataset(
dataset=self.config['dataset'],
aug=self.config['aug']
)
self.train_loader = DataLoader(
dataset=trainset,
batch_size=self.config['batch_size'],
shuffle=True,
num_workers=2
)
self.val_loader = DataLoader(
dataset=valset,
batch_size=self.config['batch_size'],
shuffle=False,
num_workers=2
)
self.log_interval = len(self.train_loader) // 4
# init model
self.use_cuda = (self.config['use_cuda'] and torch.cuda.is_available())
self.device = torch.device('cuda:0' if self.use_cuda else 'cpu')
self.model = get_model(
model_name=self.config['model'],
model_config=self.config['model_config']
).to(self.device)
summary(self.model, (3, 32, 32))
print(self.model)
# init optim
if self.config['optim']['method'] == 'sgd':
self.optimizer = optim.SGD(
params=self.model.parameters(),
lr=self.config['optim']['lr'],
momentum=self.config['optim']['momentum'],
weight_decay=self.config['optim']['weight_decay']
)
elif self.config['optim']['method'] == 'adam':
self.optimizer = optim.Adam(
params=self.model.parameters(),
lr=self.config['optim']['lr']
)
else:
raise ValueError('Invalid optim: %s.' % self.config['optim']['method'])
self.scheduler = Poly_LR_Scheduler(
base_lr=self.config['optim']['lr'],
base_poly=self.config['poly'],
num_epochs=self.config['epochs'],
iters_per_epoch=len(self.train_loader),
)
print(self.optimizer)
# init summary writer
if self.config['dump_summary']:
self.writer = SummaryWriter(self.path)
# init random seed
def init_seed(self, seed):
random.seed(seed)
np.random.seed(seed)
torch.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
# train for a single epoch
def train_one_epoch(self, epoch):
self.model.train()
self.curr_epoch = epoch
for batch_idx, (data, target) in enumerate(self.train_loader):
if self.use_cuda:
data, target = data.to(self.device), target.to(self.device)
self.scheduler(self.optimizer, self.writer, batch_idx, epoch, self.best_pred)
self.optimizer.zero_grad()
output = self.model(data)
loss = F.cross_entropy(output, target)
loss.backward()
self.optimizer.step()
if batch_idx % self.log_interval == 0:
print('Train Step: [%5d/%5d (%3.0f%s)], Loss: %.6f' %(
batch_idx * self.config['batch_size'], len(self.train_loader.dataset),
100.0 * batch_idx / len(self.train_loader), '%', loss.item()
))
self.check_accuracy()
# evaluate current model
def eval(self, data_loader):
self.model.eval()
loss, correct = 0, 0
for data, target in data_loader:
if self.use_cuda:
data, target = data.to(self.device), target.to(self.device)
output = self.model(data)
loss += F.cross_entropy(output, target, reduction='sum').item()
pred = output.data.max(1, keepdim=True)[1]
correct += pred.eq(target.data.view_as(pred)).cpu().sum()
accuracy = 100. * correct / len(data_loader.dataset)
loss /= len(data_loader.dataset)
return accuracy, loss
# print accuracy per epoch
def check_accuracy(self):
train_acc, train_loss = self.eval(self.train_loader)
val_acc, val_loss = self.eval(self.val_loader)
print('Train Accuracy: %.2f%s\tTrain Loss: %.6f' % (train_acc, '%', train_loss))
print('Val Accuracy: %.2f%s\tVal Loss: %.6f' % (val_acc, '%', val_loss))
if val_acc > self.best_pred:
self.best_pred = val_acc
if self.config['dump_summary']:
self.writer.add_scalar('accuracy/train', train_acc, self.curr_epoch)
self.writer.add_scalar('loss/train', train_loss, self.curr_epoch)
self.writer.add_scalar('accuracy/val', val_acc, self.curr_epoch)
self.writer.add_scalar('loss/val', val_loss, self.curr_epoch)
# main function for training
def train(self):
for epoch in range(1, self.config['epochs']+1):
print('\n============ train epoch [%2d/%2d] =================' % (epoch, self.config['epochs']))
self.train_one_epoch(epoch)
print('==================================================')
if self.config['export_best']:
weight_path = os.path.join(self.path, 'weights_%.2f.pth' % self.best_pred)
torch.save(self.model.state_dict(), weight_path)
print('[Weights]: best state dict exported.')
runtime = int(time.time() - self.start_time) / 60
print('\n[Time]: %d mins\n[Best Pred]: %.2f%s' % (runtime, self.best_pred, '%'))
if __name__ == '__main__':
if len(sys.argv) > 2:
print('testing mode...')
CONFIG['use_cuda'] = False
CONFIG['batch_size'] = 2
print(sys.argv)
trainer = Trainer(config=CONFIG) # init trainer
trainer.train() # training