-
Notifications
You must be signed in to change notification settings - Fork 0
/
train.py
110 lines (92 loc) · 4.1 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
#!/usr/bin/env python3
#
# Copyright, TU Dortmund 2020
#
# MIT License
# Permission is hereby granted, free of charge, to any person obtaining a copy
# of this software and associated documentation files (the "Software"), to deal
# in the Software without restriction, including without limitation the rights
# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
# copies of the Software, and to permit persons to whom the Software is
# furnished to do so, subject to the following conditions:
#
# The above copyright notice and this permission notice shall be included in
# all copies or substantial portions of the Software.
#
# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
# SOFTWARE.
import argparse
import pandas as pd
import torch
import torch.optim as optim
from torch.backends import cudnn
from torch.nn import CrossEntropyLoss
from torchvision.models.segmentation import deeplabv3_resnet101, fcn_resnet101
from tqdm import tqdm
from dataset import get_train_loader
from predict import predict
from utils import path_join
aux_loss = True
parser = argparse.ArgumentParser(
description='Training deeplab or fcn on cityscapes dataset')
parser.add_argument('--config', type=str, default='cityscapes_deeplab.yaml',
help='config file')
args0 = parser.parse_args()
cudnn.benchmark = True
args, data_loader = get_train_loader(args0.config)
if args.architecture == 'deeplab':
model = deeplabv3_resnet101(pretrained=False, progress=True,
num_classes=args.classes, aux_loss=aux_loss)
elif args.architecture == 'fcn':
model = fcn_resnet101(pretrained=False, progress=True,
num_classes=args.classes, aux_loss=aux_loss)
else:
raise ValueError('Unknown architecture specified. '
'Please choose either \'fcn\' or \'deeplab\'.')
model.cuda()
print(model)
criterion = CrossEntropyLoss(ignore_index=args.ignore_label)
optimizer = optim.SGD(model.parameters(), lr=args.base_lr,
momentum=args.momentum)
loss_log = []
for epoch in range(args.epochs):
running_loss_ep = 0.0
# Use tqdm progress bar for iterations
data_generator = tqdm(enumerate(data_loader, 0), total=len(data_loader),
ncols=80)
for i, (inputs, labels) in data_generator:
# For some reason .cuda() did not work,
# so this is how we move the data to the GPU
inputs = torch.cuda.FloatTensor(inputs.numpy())
labels = torch.cuda.FloatTensor(labels.numpy()).long()
optimizer.zero_grad()
# forward + backward + optimize
outputs = model(inputs)
loss_regular = criterion(outputs['out'], labels)
loss_aux = criterion(outputs['aux'], labels)
loss = loss_regular + args.aux_weight * loss_aux
loss.backward()
optimizer.step()
loss_val = loss.item()
running_loss_ep += loss_val
# print statistics
data_generator.set_description(
'Epoch {}/{}'.format(epoch + 1, args.epochs))
data_generator.set_postfix(loss=loss_val)
loss_log.append([epoch, (running_loss_ep / len(data_loader))])
if epoch % 5 == 0:
torch.save(model.state_dict(),
path_join(args.output_dir, f'model_ep{epoch}.dict'))
final_model_path = path_join(args.output_dir, f'model_ep{args.epochs - 1}.dict')
torch.save(model.state_dict(), final_model_path)
loss_df = pd.DataFrame(data=loss_log, columns=['epoch', 'cross_entropy_loss'])
loss_df.to_csv(path_join(args.output_dir, f'train_loss.csv'), index=None)
print('Finished Training')
print('Starting Evaluation')
predict(args0.config, final_model_path, args.output_dir, args.classes,
args.architecture, aux_loss=aux_loss)