-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathtrain.py
221 lines (187 loc) · 8.37 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
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
"""
@Project: PICR_Net
@File: train.py
@Author: chen zhang
@Institution: Beijing JiaoTong University
"""
import time
import os
import logging
from datetime import datetime
from mindspore import dataset,context,nn
import mindspore
import mindspore.nn as nn
from mindspore.ops import operations as P
from mindspore.ops import functional as F
from mindspore.ops import composite as C
from mindspore.common.tensor import Tensor
from mindspore.common.parameter import Parameter
# from torch import einsum
# from einops import rearrange, repeat
import mindspore.ops as ops
import numpy as np
from math import exp
from tensorboardX import SummaryWriter
from setting.dataLoader import get_loader
from setting.utils import clip_gradient, adjust_lr
from setting.utils import create_folder, random_seed_setting
from setting.options_ms import opt
from setting.loss import IOU, SSIM
from model.build_model_ms import PICR_Net
random_seed_setting()
os.environ["CUDA_VISIBLE_DEVICES"] = opt.gpu_id
context.set_context(mode=context.PYNATIVE_MODE, device_target="GPU")
# Logs
save_path = create_folder(opt.save_path)
logging.basicConfig(filename=save_path + 'log.log',
format='[%(asctime)s-%(filename)s-%(levelname)s:%(message)s]',
level=logging.INFO,
filemode='a',
datefmt='%Y-%m-%d %I:%M:%S %p')
logging.info(f'Config--epoch:{opt.epoch}; lr:{opt.lr}; batch_size:{opt.batchsize}; image_size:{opt.trainsize}')
writer = SummaryWriter(save_path + 'summary')
# load data
train_loader, train_num = get_loader(opt.rgb_root, opt.depth_root, opt.gt_root, opt.batchsize, opt.trainsize)
val_loader, val_num = get_loader(opt.val_rgb_root, opt.val_depth_root, opt.val_gt_root, 1, opt.trainsize)
print(f'Loading data, including {train_num} training images and {val_num} validation images.')
logging.info(f'Loading data, including {train_num} training images and {val_num} validation images.')
# model
model = PICR_Net()
n_parameters = sum(p.numel() for p in model.trainable_params())
logging.info(f"number of params: {n_parameters}")
# check model size
# if not os.path.exists('module_size'):
# os.makedirs('module_size')
# for name, module in model.named_children():
# mindspore.save_checkpoint(module, 'module_size/' + '%s' % name + '.pth')
# optimizer
optimizer = nn.optim.Adam(model.trainable_params(), opt.lr)
# Loss function
bce_loss = nn.BCELoss(reduction='mean')
ssim_loss = SSIM(window_size=11, size_average=True)
iou_loss = IOU(size_average=True)
def loss_bce_ssim_iou(pred, target):
bce_out = bce_loss(pred, target)
ssim_out = 1 - ssim_loss(pred, target)
iou_out = iou_loss(pred, target)
loss = bce_out + iou_out + ssim_out
return loss
def loss_bce_iou(pred, target):
bce_out = bce_loss(pred, target)
# ssim_out = 1 - ssim_loss(pred, target)
iou_out = iou_loss(pred, target)
loss = bce_out + iou_out
return loss
# # Restore training from checkpoints
# if opt.load is not None:
# checkpoint = torch.load(opt.load)
# opt.epoch = checkpoint['epoch']
# model.load_state_dict(checkpoint['state_dict'])
# optimizer.load_state_dict(checkpoint['optimizer'])
# print(f'Load model from [{opt.load}]')
# train function
class ComputeLoss(nn.Cell):
def __init__(self, network):
super(ComputeLoss, self).__init__(auto_prefix=False)
self.network = network
def construct(self,images, depths, gts):
s, sides = self.network(images, depths)
# for side in sides:
# print(side.shape)
# s = torch.sigmoid(s)
gts_s1 = F.interpolate(gts, (7, 7), mode='bilinear', align_corners=True)
gts_s2 = F.interpolate(gts, (14, 14), mode='bilinear', align_corners=True)
gts_s3 = F.interpolate(gts, (28, 28), mode='bilinear', align_corners=True)
gts_s4 = F.interpolate(gts, (56, 56), mode='bilinear', align_corners=True)
loss_s1 = loss_bce_ssim_iou(sides[0].sigmoid(), gts_s1)
loss_s2 = loss_bce_ssim_iou(sides[1].sigmoid(), gts_s2)
loss_s3 = loss_bce_ssim_iou(sides[2].sigmoid(), gts_s3)
loss_s4 = loss_bce_ssim_iou(sides[3].sigmoid(), gts_s4)
loss_side = loss_s4 / 2 + loss_s3 / 4 + loss_s2 / 8 + loss_s1 / 16
# loss_side = loss_s4 + loss_s3 + loss_s2 + loss_s1
loss_main = loss_bce_ssim_iou(s.sigmoid(), gts)
loss = loss_main + loss_side
return loss
def train(train_loader, model, optimizer, epoch, save_path, iteration):
model.set_train(True)
loss_all = 0
net = ComputeLoss(model)
T_net = nn.TrainOneStepCell(net, optimizer)
# print(iteration)
# iteration = len(train_loader)
try:
for i, data in enumerate(train_loader, start=1):
data["rgb"] = F.squeeze(data["rgb"], axis=(1))
data["gt"] = F.squeeze(data["gt"], axis=(1))
data["d"] = F.squeeze(data["d"], axis=(1))
images, depths, gts = data['rgb'], data['d'], data['gt']
loss_step = T_net(images, depths, gts)
loss_all += loss_step.asnumpy()
if i % (iteration // 4) == 0 or i == iteration:
print(f'{datetime.now()} Epoch [{epoch:03d}/{opt.epoch:03d}], '
f'Step [{i:04d}/{iteration:04d}], loss_step: {loss_step.asnumpy():.4f}')
logging.info(f'{datetime.now()} Epoch [{epoch:03d}/{opt.epoch:03d}], '
f'Step [{i:04d}/{iteration:04d}], loss_step: {loss_step.asnumpy():.4f}')
# loss_avg = loss_all / iteration
# print(f'Epoch [{epoch:03d}/{opt.epoch:03d}]:Loss_train_avg={loss_avg:.4f}')
# logging.info(f'Epoch [{epoch:03d}/{opt.epoch:03d}], Loss_train_avg: {loss_avg:.4f}')
# writer.add_scalar('Loss-train-avg', loss_avg, global_step=epoch)
if (epoch % 5 == 0 or epoch == opt.epoch):
mindspore.save_checkpoint(model, save_path + 'PICR_Net_epoch_{}'.format(epoch))
except KeyboardInterrupt:
print('Keyboard Interrupt: save model and exit.')
state = {
'epoch': epoch,
'state_dict': model.state_dict(),
'optimizer': optimizer.state_dict()
}
mindspore.save_checkpoint(state, save_path + 'PICR_Net_epoch_{}_checkpoint.pth.tar'.format(epoch + 1))
print('Save checkpoint successfully!')
raise
# test function
best_loss = 100
best_epoch = 1
def test(val_loader, model, epoch, save_path):
global best_loss, best_epoch
model.set_train(False)
loss_sum = 0
for i, (image, depth, gt) in enumerate(val_loader, start=1):
pre = model(image, depth)
loss = loss_bce_ssim_iou(pre, gt)
loss_sum += loss.detach()
loss_epoch = loss_sum / val_num
if loss_epoch < best_loss:
best_loss, best_epoch = loss_epoch, epoch
mindspore.save_checkpoint(model, save_path + 'PICR_Net_epoch_best.pth')
print(f'Epoch [{epoch:03d}/{opt.epoch:03d}]:Loss_val={loss_epoch:.4f},'
f' Best_loss={best_loss:.4f}, Best_epoch:{best_epoch:03d}')
if __name__ == '__main__':
print("-------------------Config-------------------\n"
f'epoch:\t\t{opt.epoch}\n'
f'lr:\t\t{opt.lr}\n'
f'batchsize:\t{opt.batchsize}\n'
f'image_size:\t{opt.trainsize}\n'
f'decay_epoch:\t{opt.decay_epoch}\n'
f'decay_rate:\t{opt.decay_rate}\n'
f'checkpoint:\t{opt.load}\n'
"--------------------------------------------\n")
print("Start train...")
import warnings
warnings.filterwarnings("ignore")
time_begin = time.time()
# warm_up_with_multistep_lr = lambda epoch: epoch / 5 if epoch <= 5 else \
# 0.2 ** len([m for m in [40, 80, 100] if m <= epoch])
# scheduler = torch.optim.lr_scheduler.LambdaLR(optimizer, lr_lambda=warm_up_with_multistep_lr)
for epoch in range(1, opt.epoch + 1):
decay = opt.decay_rate ** (epoch // opt.decay_epoch)
lr = decay * opt.lr
# cur_lr = adjust_lr(optimizer, opt.lr, epoch, opt.decay_rate, opt.decay_epoch)
ops.assign(optimizer.learning_rate, Tensor(lr, mindspore.float32))
# cur_lr = scheduler.get_lr()
writer.add_scalar('learning-rate', lr, global_step=epoch)
train(train_loader, model, optimizer, epoch, save_path,train_num)
# test(val_loader, model, epoch, save_path)
# scheduler.step()
time_epoch = time.time()
print(f"Time out:{time_epoch - time_begin:.2f}s\n")
logging.info(f"Time out:{time_epoch - time_begin:.2f}s\n")