-
Notifications
You must be signed in to change notification settings - Fork 56
/
train.py
218 lines (197 loc) · 10 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
# -*- coding: utf-8 -*-
"""
Created on Wed Jun 27 11:49:42 2018
@author: ZK
"""
import torch
from torch import nn
from torch.autograd import Variable as V
import os
from axis import SmoothL1Loss
from axis import Myloss
#%%
def train_model(dataloader, model, optimizer, lmbda, scheduler, num_epochs, pth_dir, use_gpu):
if not os.path.exists(pth_dir):
os.makedirs(pth_dir)
dirlist = os.listdir(pth_dir)
if (dirlist):
# del dirlist[dirlist.index('record.txt')]
l = [int(i.split('.')[0].split('_')[-1]) for i in dirlist]
former_epoch = max(l)
model.load_state_dict(torch.load(pth_dir+'/epoch_'+str(former_epoch)+'.pth'))
print('former_epoch %d loaded.' % former_epoch)
else:
former_epoch = 0
print('first train begin.')
for epoch in range(former_epoch+1, num_epochs+1):
print('-' * 20)
print('Epoch {}/{}'.format(epoch, num_epochs))
# for phase in ['train', 'valid']:
phase = 'train'
epoch_loss = 0
epoch_closs = 0
epoch_rloss = 0
if phase == 'train':
print('-----train-----')
if scheduler:
scheduler.step()
model.train(True) # Set model to training mode
else:
print('-----valid-----')
model.train(False) # Set model to evaluate mode
phase = 'train'
for i, tvdata in enumerate(dataloader[phase]):
template, detection, clabel, rlabel, pcc, ratio = tvdata
""""""
# template, detection, clabel, rlabel, pcc, ratio = template.squeeze(), detection.squeeze(), clabel.squeeze(), rlabel.squeeze(), pcc.squeeze(), ratio.squeeze()
# template, detection, clabel, rlabel, pcc, ratio = template.numpy(), detection.numpy(), clabel.numpy(), rlabel.numpy(), pcc.numpy(), ratio.numpy()
# import cv2
# import numpy as np
# import math
# from axis import xywh_to_x1y1x2y2
# template = np.transpose(template,(1,2,0))
# template = cv2.cvtColor(template, cv2.COLOR_RGB2BGR)
# cv2.imshow('img', template)
# cv2.waitKey(0)
#
# detection = np.transpose(detection,(1,2,0))
# detection = cv2.cvtColor(detection, cv2.COLOR_RGB2BGR)
## cv2.imshow('img', detection)
## cv2.waitKey(0)
##
# a = 64
# s = a**2
# r = [[3*math.sqrt(s/3.),math.sqrt(s/3.)], [2*math.sqrt(s/2.),math.sqrt(s/2.)], [a,a], [math.sqrt(s/2.),2*math.sqrt(s/2.)], [math.sqrt(s/3.),3*math.sqrt(s/3.)]]
# r = [list(map(round, i)) for i in r]
#
# loc1 = np.where(clabel > 0.5)
## img = cv2.imread('./lq/JPEGImages/'+os.listdir('./lq/JPEGImages/')[i])
# for where in range(len(loc1[0])):
# loc = [loc1[0][where], loc1[1][where], loc1[2][where]]
#
# anchor = [7+15*loc[1], 7+15*loc[2]] + r[loc[0]] #根据loc确定anchor
# "根据loc确定对anchor的修正："
# reg = [rlabel[loc[0]*4, loc[1], loc[2]], rlabel[loc[0]*4+1, loc[1], loc[2]], rlabel[loc[0]*4+2, loc[1], loc[2]], rlabel[loc[0]*4+3, loc[1], loc[2]]]
# "根据anchor及reg确定proposals"
# pro = [anchor[0]+reg[0]*anchor[2], anchor[1]+reg[1]*anchor[3], anchor[2]*math.exp(reg[2]), anchor[3]*math.exp(reg[3])]
## pro = anchor
## "把在255X255中的proposals转换成原图的对应位置"
## pro2 = [pro[0]*ratio+pcc[2]-pcc[0], pro[1]*ratio+pcc[3]-pcc[1], pro[2]*ratio, pro[3]*ratio]
# list1 = xywh_to_x1y1x2y2(pro)
# list1 = list(map(lambda x:int(round(x)), list1))
# cv2.rectangle(detection, (list1[0],list1[1]), (list1[2],list1[3]), (0,255,0), 1)
# cv2.imshow('img', detection)
# cv2.waitKey(0)
# detection = Image.fromarray(cv2.cvtColor(detection,cv2.COLOR_BGR2RGB))
# detection.save('./tmp/'+str(i)+'.jpg')
# cv2.imwrite('./tmp/'+str(i)+'.jpg', detection, [int(cv2.IMWRITE_JPEG_QUALITY), 95])
""""""
if use_gpu:
target = torch.zeros(clabel.shape).cuda()+1
template = V(template.cuda())
detection = V(detection.cuda())
clabel = V(clabel.cuda())
rlabel = V(rlabel.cuda())
model = model.cuda()
else:
target = torch.zeros(clabel.shape)+1
template = V(template)
detection = V(detection)
clabel = V(clabel)
rlabel = V(rlabel)
optimizer.zero_grad()
# forward
coutput, routput = model(template, detection)
# coutput, routput, clabel, rlabel = coutput.squeeze(), routput.squeeze(), clabel.squeeze(), rlabel.squeeze()
coutput, clabel = coutput.squeeze(), clabel.squeeze()
coutput = coutput.view(5, 2, 17, 17) # Batch*k*2*17*17
# routput0 = routput[0].data.numpy()
# rlabel0 = rlabel[0].data.numpy()
closs = nn.CrossEntropyLoss()(coutput, clabel)
rloss = SmoothL1Loss(use_gpu = use_gpu)(clabel, target, routput, rlabel)
# rloss = nn.SmoothL1Loss()(routput, rlabel)
loss = Myloss()(coutput, clabel, target, routput, rlabel, lmbda)
# loss = closs + lmbda * rloss
loss2 = torch.add(closs, lmbda, rloss)
epoch_loss += loss2.data.item()
epoch_closs += closs.data.item()
epoch_rloss += rloss.data.item()
# epoch_rloss += 0
# backward + optimize only if in training
if phase == 'train':
loss.backward()
optimizer.step()
# statistics
# top1num, top1acc = accuracy(outputs, labels, 1)
# top3num, top3acc = accuracy(outputs, labels, 3)
# epoch_top1num += top1num
# epoch_top3num += top3num
if (phase == 'train'):
if(i+1 == 2 or (i+1) % 100 == 0):
print('batch %d, train loss:%.6f' % (i+1, loss.data.item()))
# duration = time.time() - since
# print('step %d in %.0f seconds. loss: %.6f' % (i+1, duration, loss.data[0]))
# print(' * top1acc:{top1acc:.6f}; top3acc:{top3acc:.6f}'
# .format(top1acc=top1acc, top3acc=top3acc))
if (i+1 == len(dataloader[phase])):
print('train loss:%.6f' % (epoch_loss/len(dataloader[phase])))
print('closs:%.6f' % (epoch_closs/len(dataloader[phase])))
print('rloss:%.6f' % (epoch_rloss/len(dataloader[phase])))
# with open(RECORD_FILE, 'a') as f:
# f.write('-'*20 + '\nEpoch %d/%d\n' % (epoch,num_epochs))
# f.write('Epoch %d: loss:%.6f; top1acc:%.6f; top3acc:%.6f\n'
# % (epoch, epoch_loss/len(dataloader), epoch_top1num/len(dataset_train), epoch_top3num/len(dataset_train)))
# elif (phase == 'valid'):
# if (i+1 == len(valid_dataloader)):
# print('\nvalid loss:%.6f;\ntop1acc:%.6f; top3acc:%.6f'
# % (epoch_loss/len(valid_dataloader), epoch_top1num/len(dataset_valid), epoch_top3num/len(dataset_valid)))
# with open(RECORD_FILE, 'a') as f:
# f.write('Epoch %d: loss:%.6f; top1acc:%.6f; top3acc:%.6f\n'
# % (epoch, epoch_loss/len(valid_dataloader), epoch_top1num/len(dataset_valid), epoch_top3num/len(dataset_valid)))
# deep copy the model
# if epoch_acc > best_acc:
# best_acc = epoch_acc
# best_model_wts = model_conv.state_dict()
torch.save(model.state_dict(), (pth_dir + 'epoch_%d.pth')% epoch)
# print('current model saved to epoch_%d.pth' % epoch)
#%%
from SRPN import SiameseRPN
#from data import dataloader
from data_otb import dataloader
import torch.optim as optim
from torch.optim import lr_scheduler
#%%
if __name__ == '__main__':
model = SiameseRPN()
params = []
# params += list(model.features[0].parameters())
# params += list(model.features[3].parameters())
# params += list(model.features[6].parameters())
params += list(model.features[8].parameters())
params += list(model.features[10].parameters())
params += list(model.conv1.parameters())
params += list(model.conv2.parameters())
params += list(model.conv3.parameters())
params += list(model.conv4.parameters())
optimizer = optim.Adam(params, lr=1e-3, eps=1e-8, weight_decay=0)
# optimizer = optim.SGD(params, lr=1e-3)
scheduler = lr_scheduler.StepLR(optimizer, step_size=40, gamma=0.1)
train_model(
dataloader = dataloader
,
model = model
,
optimizer = optimizer
,
# scheduler = scheduler
scheduler = None
,
lmbda = 1
,
num_epochs = 100
,
pth_dir = './pth_OTB2015/'
,
use_gpu = True
)
#%%