-
Notifications
You must be signed in to change notification settings - Fork 0
/
train.py
189 lines (159 loc) · 6.42 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
#!/usr/bin/python
# -*- encoding: utf-8 -*-
from datasets import get_datasets
from utils.AttrDict import AttrDict
from config import load_cfg
from torch.utils.data import DataLoader
from models import get_model
from optimizer import Optimizer
import torch.nn as nn
import torch
import os
import sys
import logging
import pickle
import numpy as np
import cv2
## TODO:
# 1. image_siza -- done
# 2. optimizer class encapsuling optimizer and scheduler -- done
# 3. encapsule load checkpoint and model to a method
# 4. scheduler checkpoint -- done
# 5. deeplab lfov metric all 0
# 6. use core/ to collect other files
def resize_label(label, size):
if label.shape[-2] == size[0] and label.shape[-1] == size[1]:
return label
size = (label.shape[0], size[0], size[1])
new_label = np.empty(size, dtype = np.int32)
for i, lb in enumerate(label.numpy()):
new_label[i, ...] = cv2.resize(lb, size[1:], interpolation = cv2.INTER_NEAREST)
return torch.from_numpy(new_label).long()
def train(cfg_file):
## config
cfg = load_cfg(cfg_file)
## logging
FORMAT = '%(levelname)s %(filename)s: %(message)s'
logging.basicConfig(level=logging.INFO, format=FORMAT, stream=sys.stdout)
logger = logging.getLogger(__name__)
## data loader
trainset = get_datasets(cfg.train.datasets, 'train')
valset = get_datasets(cfg.val.datasets, 'val')
trainloader = DataLoader(trainset,
batch_size = cfg.train.batch_size,
shuffle = True,
num_workers = 4,
drop_last = True)
valloader = DataLoader(valset,
batch_size = 4,
shuffle = True,
num_workers = 4,
drop_last = True)
## network and checkpoint
model = get_model(cfg).float().cuda()
print(model)
if cfg.model.class_weight is not None:
weight = torch.Tensor(cfg.model.class_weight) # ignore some labels or set weight
Loss = nn.CrossEntropyLoss(weight = weight).cuda()
## optimizer
optimizer = Optimizer(model.parameters(), cfg)
## checkpoint
save_path = cfg.out_path
if not os.path.exists(save_path): os.makedirs(save_path)
save_name = os.path.join(save_path, 'model.pytorch')
if os.path.exists(save_name): return
model_ckpts = os.listdir(save_path)
its = [int(os.path.splitext(el)[0].split('_')[2]) for el in model_ckpts if el[:5] == 'model']
ckpt_max_it = 0
if len(its) > 0:
ckpt_max_it = max(its)
logger.info('resume from checkpoint iter: {}'.format(ckpt_max_it))
model_ckpt = os.path.join(save_path, ''.join(['model_iter_', str(ckpt_max_it), '.pytorch']))
model.load_state_dict(torch.load(model_ckpt))
optimizer.load_checkpoint(save_path, ckpt_max_it)
start_it = ckpt_max_it + 1
## multi-gpu
model = nn.DataParallel(model, device_ids = None)
## train
result = AttrDict({
'train_loss': [],
'val_loss': [],
'cfg': cfg
})
trainiter = iter(trainloader)
for it in range(start_it, cfg.train.max_iter):
model.train()
optimizer.zero_grad()
try:
im, label = next(trainiter)
if not im.shape[0] == cfg.train.batch_size: continue
except StopIteration:
trainiter = iter(trainloader)
im, label = next(trainiter)
im = im.cuda().float()
num_class = cfg.model.num_class
logits = model(im)
label = resize_label(label, logits.shape[2:])
label = label.cuda().long().contiguous().view(-1, )
logits = logits.permute(0, 2, 3, 1).contiguous().view(-1, num_class)
loss = Loss(logits, label)
loss_value = loss.detach().cpu().numpy()
result.train_loss.append(loss_value)
loss.backward()
optimizer.step()
if it == 0: continue
if it % 20 == 0:
logger.info('iter: {}/{}, loss: {}'.format(it, cfg.train.max_iter, loss_value))
if it % cfg.val.valid_iter == 0:
valid_loss, acc_clss, acc_all = val_one_epoch(model, Loss, valloader, cfg)
result.val_loss.append(valid_loss)
if it % cfg.train.snapshot_iter == 0:
save_model_name = os.path.join(save_path, ''.join(['model_iter_', str(it), '.pytorch']))
logger.info('saving snapshot to: {}'.format(save_path))
torch.save(model.module.state_dict(), save_model_name)
optimizer.save_checkpoint(save_path, it)
logger.info('training done')
save_name = os.path.join(save_path, 'model.pytorch')
logger.info('saving model to: {}'.format(save_name))
model.cpu()
torch.save(model.module.state_dict(), save_name)
with open(save_path + '/result.pkl', 'wb') as fw:
pickle.dump(result, fw)
while True:
try:
im, label = next(trainiter)
except StopIteration:
break
print('everything done')
def val_one_epoch(model, Loss, valid_loader, cfg):
model.eval()
val_loss = []
acc_list_list = []
acc_list = [[] for i in range(cfg.model.num_class)]
for img, label in valid_loader:
logits = model(img.cuda().float())
label = resize_label(label, logits.shape[2:])
logits = logits.permute(0, 2, 3, 1).contiguous().view(-1, cfg.model.num_class)
label = label.cuda().long().contiguous().view(-1, )
loss = Loss(logits, label)
val_loss.append(loss.detach().cpu().numpy())
clsses = logits.detach().cpu().numpy().argmax(axis = 1)
lbs = label.cpu().numpy().astype(np.int64)
for idx in range(cfg.model.num_class - 1):
indices = np.where(lbs == idx)
clss = clsses[indices]
acc_list[idx].extend(list(clss == idx))
valid_loss = sum(val_loss) / len(val_loss)
acc_per_class = np.array([sum(el) / len(el) if not len(el) == 0 else None for el in acc_list])
acc_list_all = []
[acc_list_all.extend(el) for el in acc_list]
acc_all = sum(acc_list_all) / len(acc_list_all)
print('=======================================')
print('result on validation set:')
print('accuracy per class:\n {}'.format(acc_per_class.reshape(-1, 1)))
print('accuracy all: {}'.format(acc_all))
print('validation loss: {}'.format(valid_loss))
print('=======================================')
return valid_loss, acc_per_class, acc_all
if __name__ == '__main__':
train()