This repository has been archived by the owner on Apr 26, 2023. It is now read-only.
-
Notifications
You must be signed in to change notification settings - Fork 6
/
Copy pathclassifier.py
executable file
·152 lines (144 loc) · 7.17 KB
/
classifier.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
#!/bin/env python3
import argparse
import os
import numpy as np
import torch
import torch.nn as nn
from torch.optim import Adam
from torch.utils.data import DataLoader
from torch.utils.tensorboard import SummaryWriter
from torchvision import transforms
from torchvision.datasets import ImageFolder
from tqdm import tqdm
from data_parallel import get_data_parallel
from helpers import load_epoch
from models import CaptchaClassifierCNN40x40
from running_log import RunningLog
def eval_model(model, valid_data_loader, device):
criterion = nn.CrossEntropyLoss().to(device)
total_count, correct_count = 0, 0
losses = []
for data in tqdm(valid_data_loader, desc='Eval'):
data = [x.to(device) for x in data]
total_count += data[0].size(0)
output = model(data[0])
loss = criterion(output, data[1])
losses.append(loss.item())
# noinspection PyUnresolvedReferences
correct_count += (torch.argmax(output, dim=1) == data[1]).sum().item()
return np.mean(losses), correct_count / total_count
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--task', choices=['train', 'valid', 'train-all'],
default='train', help='task to run')
parser.add_argument('--dataset_path', help='path to the dataset folder',
default='dataset/segmented')
parser.add_argument('--save_path', help='path for saving models and codes',
default='save/classifier')
parser.add_argument('--gpu', type=lambda x: list(map(int, x.split(','))),
default=[], help="GPU ids separated by `,'")
parser.add_argument('--load', type=int, default=0,
help='load module training at give epoch')
parser.add_argument('--epoch', type=int, default=20, help='epoch to train')
parser.add_argument('--batch_size', type=int, default=64, help='batch size')
parser.add_argument('--learning_rate', type=float, default=0.0001,
help='learning rate')
parser.add_argument('--log_every_iter', type=int, default=100,
help='log loss every numbers of iteration')
parser.add_argument('--valid_every_epoch', type=int, default=1,
help='run validation every numbers of epoch; '
'0 for disabling')
parser.add_argument('--save_every_epoch', type=int, default=5,
help='save model every numbers of epoch; '
'0 for disabling')
parser.add_argument('--comment', default='', help='comment for tensorboard')
args = parser.parse_args()
running_log = RunningLog(args.save_path)
running_log.set('parameters', vars(args))
os.makedirs(args.save_path, exist_ok=True)
model = get_data_parallel(CaptchaClassifierCNN40x40(), args.gpu)
device = torch.device("cuda:%d" % args.gpu[0] if args.gpu else "cpu")
optimizer_state_dict = None
if args.load > 0:
model_state_dict, optimizer_state_dict = \
load_epoch(args.save_path, args.load)
model.load_state_dict(model_state_dict)
model.to(device)
running_log.set('state', 'interrupted')
if args.task == 'train' or args.task == 'train-all':
model.train()
# noinspection PyUnresolvedReferences
train_dataset = ImageFolder(os.path.join(
args.dataset_path, 'train' if args.task == 'train' else 'all'),
transform=transforms.Compose([
transforms.Grayscale(),
transforms.ToTensor(),
]))
train_data_loader = DataLoader(train_dataset,
batch_size=args.batch_size,
shuffle=True)
valid_data_loader = None
optimizer = Adam(model.parameters(), lr=args.learning_rate)
if optimizer_state_dict is not None:
optimizer.load_state_dict(optimizer_state_dict)
criterion = nn.CrossEntropyLoss().to(device)
writer = SummaryWriter(comment=args.comment or os.path.basename(args.save_path))
step = 0
for epoch in tqdm(range(args.load + 1, args.epoch + 1), desc='Epoch'):
losses = []
for iter, data in enumerate(tqdm(train_data_loader, desc='Iter'), 1):
data = [x.to(device) for x in data]
output = model(data[0])
loss = criterion(output, data[1])
losses.append(loss.item())
writer.add_scalar('train/loss', loss.item(), step)
optimizer.zero_grad()
loss.backward()
optimizer.step()
if iter % args.log_every_iter == 0:
# noinspection PyStringFormat
tqdm.write('epoch:[%d/%d] iter:[%d/%d] Loss=%.5f' %
(epoch, args.epoch, iter, len(train_data_loader),
np.mean(losses)))
losses = []
step += 1
if args.valid_every_epoch and epoch % args.valid_every_epoch == 0:
if valid_data_loader is None:
# noinspection PyUnresolvedReferences
valid_dataset = ImageFolder(os.path.join(args.dataset_path, 'test'),
transform=transforms.Compose([
transforms.Grayscale(),
transforms.ToTensor(),
]))
valid_data_loader = DataLoader(valid_dataset,
batch_size=args.batch_size,
shuffle=False)
model.eval()
loss, acc = eval_model(model, valid_data_loader, device)
# noinspection PyStringFormat
tqdm.write('Loss=%f Accuracy=%f' % (loss, acc))
writer.add_scalar('eval/loss', loss, epoch)
writer.add_scalar('eval/acc', acc, epoch)
model.train()
if args.save_every_epoch and epoch % args.save_every_epoch == 0:
tqdm.write('saving to epoch.%04d.pth' % epoch)
torch.save((model.state_dict(), optimizer.state_dict()),
os.path.join(args.save_path,
'epoch.%04d.pth' % epoch))
elif args.task == 'valid':
model.eval()
# noinspection PyUnresolvedReferences
valid_dataset = ImageFolder(os.path.join(args.dataset_path, 'test'),
transform=transforms.Compose([
transforms.Grayscale(),
transforms.ToTensor(),
]))
valid_data_loader = DataLoader(valid_dataset,
batch_size=args.batch_size,
shuffle=False)
loss, acc = eval_model(model, valid_data_loader, device)
# noinspection PyStringFormat
tqdm.write('Loss=%f Accuracy=%f' % (loss, acc))
running_log.set('state', 'succeeded')
if __name__ == '__main__':
main()