-
Notifications
You must be signed in to change notification settings - Fork 0
/
train.py
194 lines (172 loc) · 10.4 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
190
191
192
193
194
import torch
import torch.nn as nn
import torch.optim as optim
import yaml
import datetime
import numpy as np
import logging
import os
from tqdm import tqdm
from torch.utils.tensorboard import SummaryWriter
import torchvision
from Unet.model import Unet
from CRF.crfrnn import CrfRnn
from Loader.Loader2d import create_loader_2d
from Loader.Dataset3d import Dataset3d
from Utils.util import create_logger
from Utils.model_util import load_unet_checkpoint, save_unet_checkpoint, val, ckpt_parallel2single
def main():
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
data_config_path = '/home/kunhan/workspace/UCRFNet/Config/config_data.yml'
train_config_path = '/home/kunhan/workspace/UCRFNet/Config/config_train.yml'
data_config = yaml.load(open(data_config_path, 'r'), Loader=yaml.FullLoader)
train_config = yaml.load(open(train_config_path, 'r'), Loader=yaml.FullLoader)
train_config = train_config['train']
now = datetime.datetime.now()
date_time = now.strftime("%Y-%m-%d-%H-%M")
logging_path = os.path.join(train_config['logging_dir'], 'logging_train_' + date_time)
logger = create_logger(logging_path)
###################################################################################
# channel, class configuration
###################################################################################
n_rois = len(data_config['dataset']['roi_names'])
if n_rois > 1:
# add background as the first class
n_class = n_rois + 1
else:
n_class = 1
n_channel_unet = data_config['dataset']['n_slice_unet']
n_channel_crfrnn = data_config['dataset']['n_slice_crfrnn']
for name in os.listdir(train_config['ckpt_dir_unet']):
if os.path.isfile(os.path.join(train_config['ckpt_dir_unet'], name)):
ckpt_parallel2single(n_channel_unet, n_class, train_config['ckpt_dir_unet'], train_config['ckpt_dir_unet_crfrnn'], name, device)
return
###################################################################################
# construct net
###################################################################################
logger.info("Creating net...")
unet = Unet(n_channel_unet, n_class, bilinear=True)
crfrnn = CrfRnn(n_class, num_iterations=5).to('cpu')
###################################################################################
# parallel model and data
###################################################################################
if torch.cuda.device_count() > 1:
print("Let's use", torch.cuda.device_count(), "GPUs!")
# dim = 0 [30, xxx] -> [10, ...], [10, ...], [10, ...] on 3 GPUs
unet = nn.DataParallel(unet)
unet.to(device)
###################################################################################
# criterion, optimizer, scheduler
###################################################################################
if n_class > 1:
# input: (N, C) or (N, C, H, W)
# target: (N) or (N, H, W)
criterion = nn.CrossEntropyLoss()
else:
# input and target have the same shape. (N, *)
criterion = nn.BCEWithLogitsLoss()
logger.info("Creating optimizer...")
optimizer_unet = optim.RMSprop(unet.parameters(), lr=train_config['lr'], weight_decay=1e-8, momentum=0.9)
# optimizer_crfrnn = optim.RMSprop(unet.parameters(), lr=train_config['lr'], weight_decay=1e-8, momentum=0.9)
logger.info("Creating scheduler...")
scheduler_unet = optim.lr_scheduler.ReduceLROnPlateau(optimizer_unet, 'min', patience=2)
# scheduler_crfrnn = optim.lr_scheduler.ReduceLROnPlateau(optimizer_crfrnn, 'min', patience=2)
###################################################################################
# SummaryWriter
###################################################################################
logger.info("Creating writer")
writer = SummaryWriter(comment=f"LR_{train_config['lr']}_BS_{train_config['n_epoch']}")
###################################################################################
# train setup
###################################################################################
global_step = 0
best_loss = np.inf
train_phase = 'train'
epoch_start = 0
###################################################################################
# load previous model
###################################################################################
if train_config['load_checkpoint'] is not None:
print("Loading net...")
logging.info("Loading net...")
# load unet from checkpoint
unet, optimizer_unet, scheduler_unet, epoch_loss, epoch_start, global_step = load_unet_checkpoint(unet=unet,
optimizer_unet=optimizer_unet,
scheduler_unet=scheduler_unet,
ckpt_dir_unet=
train_config[
'ckpt_dir_unet'],
ckpt_fn_unet=
train_config[
'checkpoint_unet'],
device=device)
logger.info(f"Start from epoch: {epoch_start}, global_step: {global_step}")
print(f"Start from epoch: {epoch_start}, global step{global_step}")
###################################################################################
# train
###################################################################################
for epoch in range(epoch_start, train_config['n_epoch']):
epoch_loss = 0
dataset3d = Dataset3d(data_config['dataset'], train_phase)
n_train = len(dataset3d)
logger.info(f"Epoch: {epoch}/{train_config['n_epoch']}")
with tqdm(total=n_train, desc=f"Epoch {epoch + 1}/{train_config['n_epoch']}", unit='batch') as pbar:
for data in dataset3d:
print()
logger.info(f"Processing {data['pid']}...")
loader2d = create_loader_2d(data, data_config, train_phase)
for batch_id, batch in enumerate(loader2d):
img = batch['img_patch'].to(device=device, dtype=torch.float32) # [N, n_channel_unet, H, W]
mask_type = torch.float32 if n_class == 1 else torch.long
mask_gt = batch['mask'].to(device=device, dtype=mask_type) # [N, H, W]
mask_pred = unet(img)
loss_unet = criterion(mask_pred, mask_gt)
loss = loss_unet
optimizer_unet.zero_grad()
loss.backward()
nn.utils.clip_grad_value_(unet.parameters(), 0.1)
optimizer_unet.step()
global_step += 1
loss_scalar = loss.item()
epoch_loss += loss_scalar
logger.info(f"\tBatch: {batch_id}/{len(loader2d)}, Batch Loss: {loss_scalar}")
pbar.set_postfix(**{'loss (batch)': loss_scalar, 'pid': data['pid']})
if (global_step + 1) % train_config['write_summary_loss_batch_step'] == 0:
writer.add_scalar('Loss_unet_batch/train', loss_scalar, global_step)
if (global_step + 1) % train_config['write_summary_2d_batch_step'] == 0:
images_grid = torchvision.utils.make_grid(torch.unsqueeze(img[:, n_channel_unet // 2], 1))
gt_masks_grid = torchvision.utils.make_grid(torch.unsqueeze(mask_gt, 1))
pred_mask_grid = torchvision.utils.make_grid(torch.argmax(mask_pred, dim=1, keepdim=True))
writer.add_image('images', images_grid, global_step)
writer.add_image('gt_masks', gt_masks_grid, global_step)
writer.add_image('pred_masks', pred_mask_grid, global_step)
pbar.update()
# logging
if (epoch + 1) % train_config['logging_epoch_step'] == 0:
writer.add_scalar('Loss_unet_epoch/train', epoch_loss, global_step)
print(f"Epoch: {epoch}/{train_config['n_epoch']}, Epoch Loss: {epoch_loss}")
logger.info(f"Epoch: {epoch}/{train_config['n_epoch']}, Epoch Loss: {epoch_loss}")
# do validation and save model
if (epoch + 1) % train_config['save_model_epoch_step'] == 0:
# validation
val_loss = val(unet, crfrnn, criterion, data_config, n_class, device, logger)
writer.add_scalar('Loss_unet_epoch/val', val_loss, global_step)
print(f"Epoch: {epoch}/{train_config['n_epoch']}, Validation Loss: {val_loss}")
logger.info(f"Epoch: {epoch}/{train_config['n_epoch']}, Validation Loss: {val_loss}")
# save model
scheduler_unet.step(val_loss)
if best_loss > epoch_loss:
best_loss = epoch_loss
save_unet_checkpoint(unet=unet, optimizer_unet=optimizer_unet, scheduler_unet=scheduler_unet,
epoch_loss=epoch_loss, epoch=epoch, global_step=global_step,
ckpt_dir_unet=train_config['ckpt_dir_unet'],
ckpt_fn_unet='best_unet_ckpt' + date_time + '.ckpt')
save_unet_checkpoint(unet=unet, optimizer_unet=optimizer_unet, scheduler_unet=scheduler_unet,
epoch_loss=epoch_loss, epoch=epoch, global_step=global_step,
ckpt_dir_unet=train_config['ckpt_dir_unet'],
ckpt_fn_unet=f'unet_ckpt_{date_time}_Epoch_{epoch}.ckpt')
print(f"Epoch: {epoch}/{train_config['n_epoch']}, Save model.")
logger.info(f"Epoch: {epoch}/{train_config['n_epoch']}, Save model.")
writer.close()
if __name__ == '__main__':
main()