forked from baumgach/ralis
-
Notifications
You must be signed in to change notification settings - Fork 0
/
train_supervised.py
193 lines (165 loc) · 7.86 KB
/
train_supervised.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
import os
import sys
import shutil
import numpy as np
import random
import math
import torch
import torch.nn as nn
from torch.backends import cudnn
from torch.optim.lr_scheduler import ExponentialLR
from models.model_utils import create_models, load_models
from data.data_utils import get_data
from utils.final_utils import check_mkdir, create_and_load_optimizers, train, validate, final_test
import utils.parser as parser
cudnn.benchmark = False
cudnn.deterministic = True
def main(args):
# Set seeds
torch.manual_seed(args.seed)
torch.cuda.manual_seed_all(args.seed)
random.seed(args.seed)
np.random.seed(args.seed)
####------ Create experiment folder ------####
check_mkdir(args.ckpt_path)
check_mkdir(os.path.join(args.ckpt_path, args.exp_name))
####------ Print and save arguments in experiment folder ------####
parser.save_arguments(args)
####------ Copy current config file to ckpt folder ------####
fn = sys.argv[0].rsplit('/', 1)[-1]
shutil.copy(sys.argv[0], os.path.join(args.ckpt_path, args.exp_name, fn))
print("torch.cuda.is_available()", torch.cuda.is_available())
####------ Create segmentation, query and target networks ------####
kwargs_models = {"dataset": args.dataset,
"al_algorithm": args.al_algorithm,
"region_size": args.region_size
}
net, _, _ = create_models(**kwargs_models)
#print(net)
####------ Load weights if necessary and create log file ------####
kwargs_load = {"net": net,
"load_weights": args.load_weights,
"exp_name_toload": args.exp_name_toload,
"snapshot": args.snapshot,
"exp_name": args.exp_name,
"ckpt_path": args.ckpt_path,
"checkpointer": args.checkpointer,
"exp_name_toload_rl": args.exp_name_toload_rl,
"policy_net": None,
"target_net": None,
"test": args.test,
"dataset": args.dataset,
"al_algorithm": 'None'}
logger, curr_epoch, best_record = load_models(**kwargs_load)
####------ Load training and validation data ------####
kwargs_data = {"data_path": args.data_path,
"code_path": args.code_path,
"tr_bs": args.train_batch_size,
"vl_bs": args.val_batch_size,
"n_workers": 4,
"scale_size": args.scale_size,
"input_size": args.input_size,
"num_each_iter": args.num_each_iter,
"only_last_labeled": args.only_last_labeled,
"dataset": args.dataset,
"test": args.test,
"al_algorithm": args.al_algorithm,
"full_res": args.full_res,
"region_size": args.region_size,
"supervised": True}
train_loader, _, val_loader, _ = get_data(**kwargs_data)
####------ Create losses ------####
criterion = nn.CrossEntropyLoss(ignore_index=train_loader.dataset.ignore_label).cuda()
####------ Create optimizers (and load them if necessary) ------####
kwargs_load_opt = {"net": net,
"opt_choice": args.optimizer,
"lr": args.lr,
"wd": args.weight_decay,
"momentum": args.momentum,
"ckpt_path": args.ckpt_path,
"exp_name_toload": args.exp_name_toload,
"exp_name": args.exp_name,
"snapshot": args.snapshot,
"checkpointer": args.checkpointer,
"load_opt": args.load_opt,
"policy_net": None,
"lr_dqn": args.lr_dqn,
"al_algorithm": 'None'}
optimizer, _ = create_and_load_optimizers(**kwargs_load_opt)
# Early stopping params initialization
es_val = 0
es_counter = 0
es_dice = 0
if args.train:
print('Starting training...')
scheduler = ExponentialLR(optimizer, gamma=0.998)
net.train()
if args.modality == '2D':
for epoch in range(curr_epoch, args.epoch_num + 1):
print('Epoch %i /%i' % (epoch, args.epoch_num + 1))
tr_loss, _, tr_acc, tr_iu = train(train_loader, net, criterion,
optimizer, supervised=True)
if epoch % 1 == 0:
vl_loss, val_acc, val_iu, iu_xclass, _ = validate(val_loader, net, criterion,
optimizer, epoch, best_record,
args)
## Append info to logger
info = [epoch, optimizer.param_groups[0]['lr'],
tr_loss,
0, vl_loss, tr_acc, val_acc, tr_iu, val_iu]
for cl in range(train_loader.dataset.num_classes):
info.append(iu_xclass[cl])
logger.append(info)
scheduler.step()
# Early stopping with val jaccard/dice
es_counter += 1
if val_iu > es_val and not math.isnan(val_iu):
torch.save(net.cpu().state_dict(),
os.path.join(args.ckpt_path, args.exp_name,
'best_jaccard_val.pth'))
net.cuda()
es_val = val_iu
es_counter = 0
elif es_counter > args.patience:
print('Patience for Early Stopping reached!')
break
logger.close()
# 3D brats
else:
for epoch in range(curr_epoch, args.epoch_num + 1):
print('Epoch %i /%i' % (epoch, args.epoch_num + 1))
# adapt train loss,, ...
tr_loss, _, tr_acc, tr_iu, tr_meanDice = train(train_loader, net, criterion,
optimizer, supervised=True)
if epoch % 1 == 0:
vl_loss, val_acc, val_iu, iu_xclass, meanDice, meanDiceWT, meanDiceTC, meanDiceET, _ = validate(val_loader, net, criterion,
optimizer, epoch, best_record, args)
## Append info to logger
info = [epoch, optimizer.param_groups[0]['lr'],
tr_loss, tr_acc, tr_meanDice, vl_loss, val_acc, meanDice, meanDiceWT, meanDiceTC, meanDiceET]
logger.append(info)
scheduler.step()
# Early stopping with val jaccard/dice
es_counter += 1
if meanDice > es_dice and not math.isnan(meanDice):
torch.save(net.cpu().state_dict(),
os.path.join(args.ckpt_path, args.exp_name,
'best_jaccard_val.pth'))
net.cuda()
es_dice = meanDice
es_counter = 0
elif es_counter > args.patience:
print('Patience for Early Stopping reached!')
break
del (tr_loss, tr_acc, tr_iu, tr_meanDice, vl_loss, val_acc, meanDice, meanDiceWT, meanDiceTC, meanDiceET)
logger.close()
if args.final_test:
final_test(args, net, criterion)
if __name__ == '__main__':
####------ Parse arguments from console ------####
print("torch.cuda.is_available()", torch.cuda.is_available())
torch.cuda.empty_cache()
from subprocess import call
args = parser.get_arguments()
main(args)
torch.cuda.empty_cache()