-
Notifications
You must be signed in to change notification settings - Fork 11
/
train_div.py
318 lines (274 loc) · 14.4 KB
/
train_div.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
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.autograd import Variable
from metric import compress, calculate_top_map, calculate_map, p_topK, calculate_map_1
from models import ImgNet, TxtNet
import os.path as osp
from load_data import get_loader, get_loader_wiki
import numpy as np
import pdb
import time
import logging
def calc_dis(query_L, retrieval_L, query_dis, top_k=32):
num_query = query_L.shape[0]
map = 0
for iter in range(num_query):
gnd = (np.dot(query_L[iter, :], retrieval_L.transpose()) > 0).astype(np.float32)
tsum = np.sum(gnd)
if tsum == 0:
continue
hamm = query_dis[iter]
ind = np.argsort(hamm)[:top_k]
gnd = gnd[ind]
tsum = np.int(np.sum(gnd))
if tsum == 0:
continue
count = np.linspace(1, tsum, tsum)
tindex = np.asarray(np.where(gnd == 1)) + 1.0
map = map + np.mean(count / (tindex))
map = map / num_query
return map
class Session:
def __init__(self, opt):
self.opt = opt
if opt.data_name == 'wiki':
dataloader, data_train = get_loader_wiki('./', opt.batch_size)
else:
dataloader, data_train = get_loader(opt.data_name, opt.batch_size)
# Data Loader (Input Pipeline)
self.global_imgs, self.global_txts, self.global_labs = data_train
self.global_imgs = F.normalize(torch.Tensor(self.global_imgs)).cuda()
self.global_txts = F.normalize(torch.Tensor(self.global_txts)).cuda()
self.global_labs = torch.Tensor(self.global_labs).cuda()
self.gs, self.sa, self.ni = self.cal_similarity(self.global_imgs, self.global_txts)
self.train_loader = dataloader['train']
self.val_loader = dataloader['validation']
self.test_loader = dataloader['query']
self.database_loader = dataloader['database']
self.databasev_loader = dataloader['databasev']
txt_feat_len = self.global_txts.size(1)
self.CodeNet_I = ImgNet(code_len=opt.bit, txt_feat_len=txt_feat_len)
self.FeatNet_I = ImgNet(code_len=opt.bit, txt_feat_len=txt_feat_len)
self.CodeNet_T = TxtNet(code_len=opt.bit, txt_feat_len=txt_feat_len)
self.opt_I = torch.optim.SGD(self.CodeNet_I.parameters(), lr=opt.learning_rate, momentum=opt.momentum,
weight_decay=opt.weight_decay)
self.opt_T = torch.optim.SGD(self.CodeNet_T.parameters(), lr=opt.learning_rate, momentum=opt.momentum,
weight_decay=opt.weight_decay)
self.best = 0
# pdb.set_trace()
logger = logging.getLogger('train')
logger.setLevel(logging.INFO)
stream_log = logging.StreamHandler()
stream_log.setLevel(logging.INFO)
formatter = logging.Formatter('%(asctime)s - %(levelname)s - %(message)s')
stream_log.setFormatter(formatter)
logger.addHandler(stream_log)
self.logger = logger
def cal_similarity(self, F_I, F_T):
batch_size = F_I.size(0)
F_I = F.normalize(F_I)
S_I = F_I.mm(F_I.t())
F_T = F.normalize(F_T)
S_T = F_T.mm(F_T.t())
S_pair = self.opt.a1 * S_T + (1 - self.opt.a1) * S_I
pro = F_T.mm(F_T.t()) * self.opt.a1 + F_I.mm(F_I.t()) * (1. - self.opt.a1)
size = batch_size
top_size = self.opt.knn_number
m, n1 = pro.sort()
pro[torch.arange(size).view(-1, 1).repeat(1, top_size).view(-1), n1[:, :top_size].contiguous().view(
-1)] = 0.
pro[torch.arange(size).view(-1), n1[:, -1:].contiguous().view(
-1)] = 0.
pro = pro / pro.sum(1).view(-1, 1)
pro_dis = pro.mm(pro.t())
pro_dis = pro_dis * self.opt.scale
# pdb.set_trace()
S = (S_pair * (1 - self.opt.a2) + pro_dis * self.opt.a2)
S = S * 2.0 - 1
return S, S_pair, pro_dis
def loss_cal(self, code_I, code_T, S, I):
B_I = F.normalize(code_I)
B_T = F.normalize(code_T)
BI_BI = B_I.mm(B_I.t())
BT_BT = B_T.mm(B_T.t())
BI_BT = B_I.mm(B_T.t())
# pdb.set_trace()
diagonal = BI_BT.diagonal()
all_1 = torch.rand((BT_BT.size(0))).fill_(1).cuda()
loss_pair = F.mse_loss(diagonal, self.opt.K * all_1)
loss_dis_1 = F.mse_loss(BT_BT * (1-I), S * (1-I))
loss_dis_2 = F.mse_loss(BI_BT * (1-I), S * (1-I))
loss_dis_3 = F.mse_loss(BI_BI * (1-I), S * (1-I))
loss_cons = F.mse_loss(BI_BT, BI_BI) + \
F.mse_loss(BI_BT, BT_BT) + \
F.mse_loss(BI_BI, BT_BT) + \
F.mse_loss(BI_BT, BI_BT.t())
loss = loss_pair + (loss_dis_1 + loss_dis_2 + loss_dis_3) * self.opt.dw \
+ loss_cons * self.opt.cw
# loss = loss_pair
loss = loss
return loss, (loss_pair, loss_dis_1, loss_dis_2, loss_dis_3, loss_cons, loss_cons)
def train(self, epoch):
self.CodeNet_I.cuda().train()
self.CodeNet_T.cuda().train()
top_mAP = 0.0
num = 0.0
# self.CodeNet_I.set_alpha(epoch)
# self.CodeNet_T.set_alpha(epoch * 2.0)
self.logger.info('Epoch [%d/%d], alpha for ImgNet: %.3f, alpha for TxtNet: %.3f' % (
epoch + 1, self.opt.num_epochs, self.CodeNet_I.alpha, self.CodeNet_T.alpha))
for idx, (img, txt, labels, index) in enumerate(self.train_loader):
img = Variable(img.cuda())
txt = Variable(torch.FloatTensor(txt.numpy()).cuda())
batch_size = img.size(0)
I = torch.eye(batch_size).cuda()
_, code_I = self.CodeNet_I(img)
_, code_T = self.CodeNet_T(txt)
S = self.gs[index, :][:, index].cuda()
loss, all_los = self.loss_cal(code_I, code_T, S, I)
self.opt_I.zero_grad()
self.opt_T.zero_grad()
loss.backward(retain_graph=True)
self.opt_I.step()
self.opt_T.step()
_, code_I = self.CodeNet_I(img)
_, code_T = self.CodeNet_T(txt)
loss_i, _ = self.loss_cal(code_I, code_T.sign().detach(), S, I)
self.opt_I.zero_grad()
loss_i.backward(retain_graph=True)
self.opt_I.step()
loss_t, _ = self.loss_cal(code_I.sign().detach(), code_T, S, I)
self.opt_T.zero_grad()
loss_t.backward()
self.opt_T.step()
loss1, loss2, loss3, loss4, loss5, loss6 = all_los
top_mAP += calc_dis(labels.cpu().numpy(), labels.cpu().numpy(), -S.cpu().numpy())
num += 1.
if (idx + 1) % (len(self.train_loader)) == 0:
self.logger.info(
'Epoch [%d/%d], Iter [%d/%d] '
'Loss1: %.4f Loss2: %.4f Loss3: %.4f '
'Loss4: %.4f '
'Loss5: %.4f Loss6: %.4f '
'Total Loss: %.4f '
'mAP: %.4f'
% (
epoch + 1, self.opt.num_epochs, idx + 1,
len(self.train_loader) // self.opt.batch_size,
loss1.mean().item(), loss2.mean().item(), loss3.mean().item(),
loss4.item(),
code_T.abs().mean().item(),
code_I.abs().mean().item(),
loss.item(),
top_mAP / num))
def eval(self, step=0, num_epoch=0, last=False):
# Change model to 'eval' mode (BN uses moving mean/var).
self.CodeNet_I.eval().cuda()
self.CodeNet_T.eval().cuda()
if self.opt.EVAL == False:
re_BI, re_BT, re_L, qu_BI, qu_BT, qu_L = compress(self.databasev_loader, self.val_loader, self.CodeNet_I,
self.CodeNet_T)
MAP_I2T = calculate_top_map(qu_B=qu_BI, re_B=re_BT, qu_L=qu_L, re_L=re_L, topk=50)
MAP_T2I = calculate_top_map(qu_B=qu_BT, re_B=re_BI, qu_L=qu_L, re_L=re_L, topk=50)
MAP_I2Ta = calculate_map(qu_B=qu_BI, re_B=re_BT, qu_L=qu_L, re_L=re_L)
MAP_T2Ia = calculate_map(qu_B=qu_BT, re_B=re_BI, qu_L=qu_L, re_L=re_L)
MAP_I2Ta_1 = calculate_map_1(qu_B=qu_BI, re_B=re_BT, qu_L=qu_L, re_L=re_L)
MAP_T2Ia_1 = calculate_map_1(qu_B=qu_BT, re_B=re_BI, qu_L=qu_L, re_L=re_L)
K = [1, 200, 500, 1000, 1500, 2000, 2500, 3000, 3500, 4000, 4500, 5000]
self.logger.info('--------------------Evaluation: Calculate top MAP-------------------')
self.logger.info('MAP@50 of Image to Text: %.3f, MAP of Text to Image: %.3f' % (MAP_I2T, MAP_T2I))
self.logger.info('MAP@All of Image to Text: %.3f, MAP of Text to Image: %.3f' % (MAP_I2Ta, MAP_T2Ia))
self.logger.info('MAP@All_1 of Image to Text: %.3f, MAP of Text to Image: %.3f' % (MAP_I2Ta_1, MAP_T2Ia_1))
self.logger.info('--------------------------------------------------------------------')
if MAP_I2Ta + MAP_T2Ia > self.best:
num_epoch = 0
self.save_checkpoints(step=step, best=True)
self.best = MAP_T2Ia + MAP_I2Ta
self.logger.info("#########is best:%.3f #########" % self.best)
else:
num_epoch += 1
if self.opt.EVAL:
re_BI, re_BT, re_L, qu_BI, qu_BT, qu_L = compress(self.database_loader, self.test_loader, self.CodeNet_I,
self.CodeNet_T)
MAP_I2T = calculate_top_map(qu_B=qu_BI, re_B=re_BT, qu_L=qu_L, re_L=re_L, topk=5)
MAP_T2I = calculate_top_map(qu_B=qu_BT, re_B=re_BI, qu_L=qu_L, re_L=re_L, topk=5)
MAP_I2Ta = calculate_map(qu_B=qu_BI, re_B=re_BT, qu_L=qu_L, re_L=re_L)
MAP_T2Ia = calculate_map(qu_B=qu_BT, re_B=re_BI, qu_L=qu_L, re_L=re_L)
MAP_I2Ta_1 = calculate_map_1(qu_B=qu_BI, re_B=re_BT, qu_L=qu_L, re_L=re_L)
MAP_T2Ia_1 = calculate_map_1(qu_B=qu_BT, re_B=re_BI, qu_L=qu_L, re_L=re_L)
if self.opt.data_name == 'wiki':
K = [1, 200, 400, 500, 1000, 1500, 2000]
else:
K = [1, 200, 500, 1000, 1500, 2000, 2500, 3000, 3500, 4000, 4500, 5000]
self.logger.info('--------------------Evaluation: Calculate top MAP-------------------')
self.logger.info('MAP@50 of Image to Text: %.3f, MAP of Text to Image: %.3f' % (MAP_I2T, MAP_T2I))
self.logger.info('MAP@All of Image to Text: %.3f, MAP of Text to Image: %.3f' % (MAP_I2Ta, MAP_T2Ia))
self.logger.info('MAP@All_1 of Image to Text: %.3f, MAP of Text to Image: %.3f' % (MAP_I2Ta_1, MAP_T2Ia_1))
self.logger.info('--------------------------------------------------------------------')
retI2T = p_topK(qu_BI, re_BT, qu_L, re_L, K)
retT2I = p_topK(qu_BT, re_BI, qu_L, re_L, K)
self.logger.info(retI2T)
self.logger.info(retT2I)
now = time.strftime("%Y%m%d%H%M%S", time.localtime(time.time()))
log_name = 'log/' + self.opt.data_name + '_' + \
str(self.opt.batch_size) + '_' + str(self.opt.bit) + '_' + \
str(self.opt.dw) + '_' + str(self.opt.cw) \
+ '_' + str(self.opt.a1) + '_' + str(self.opt.a2) + '_' + \
str(self.opt.knn_number) + '_' + str(self.opt.scale) + \
'.txt'
fi = open(log_name, 'a')
fi.write('--------------------Evaluation: Calculate top MAP-------------------')
fi.write('\n')
fi.write('MAP@50 of Image to Text: %.3f, MAP of Text to Image: %.3f' % (MAP_I2T, MAP_T2I))
fi.write('\n')
fi.write('MAP@All of Image to Text: %.3f, MAP of Text to Image: %.3f' % (MAP_I2Ta, MAP_T2Ia))
fi.write('\n')
fi.write('MAP@All_1 of Image to Text: %.3f, MAP of Text to Image: %.3f' % (MAP_I2Ta_1, MAP_T2Ia_1))
fi.write('\n')
fi.write(str(retI2T.cpu().numpy()))
fi.write('\n')
fi.write(str(retT2I.cpu().numpy()))
fi.close()
log_name = 'result/' + self.opt.data_name + 'batch_size' + \
str(self.opt.batch_size) + 'bit' + str(self.opt.bit) + '.txt'
parameter = 'dw:' + str(self.opt.dw) + ' ' + 'cw:' + str(self.opt.cw) \
+ ' ' + 'a1:' + str(self.opt.a1) + ' ' + 'a2:' + str(self.opt.a2) + ' ' \
+ 'knn:' + str(self.opt.knn_number) + ' ' + 'scale:' + \
str(self.opt.scale)
fi = open(log_name, 'a')
fi.write(parameter)
fi.write('\n')
fi.write('MAP@50 of Im2Te: %.3f, MAP of Te2Im: %.3f' % (MAP_I2T, MAP_T2I))
fi.write('\n')
fi.write('MAP@All of Im2Te: %.3f, MAP of Te2Im: %.3f' % (MAP_I2Ta, MAP_T2Ia))
fi.write('\n')
fi.write('MAP@All_1 of Im2Te: %.3f, MAP of Te2Im: %.3f' % (MAP_I2Ta_1, MAP_T2Ia_1))
fi.write('\n')
fi.close()
return num_epoch
def save_checkpoints(self, step,
best=False):
file_name = '%s_%d_bit_latest.pth' % (self.opt.data_name, self.opt.bit)
if best:
file_name = '%s_%d_bit_best_epoch.pth' % (self.opt.data_name, self.opt.bit)
ckp_path = osp.join(self.opt.save_model_path, file_name)
obj = {
'ImgNet': self.CodeNet_I.state_dict(),
'TxtNet': self.CodeNet_T.state_dict(),
'step': step,
}
torch.save(obj, ckp_path)
self.logger.info('**********Save the trained model successfully.**********')
def load_checkpoints(self):
file_name = '%s_%d_bit_best_epoch.pth' % (self.opt.data_name, self.opt.bit)
ckp_path = osp.join(self.opt.save_model_path, file_name)
try:
obj = torch.load(ckp_path, map_location=lambda storage, loc: storage.cuda())
self.logger.info('**************** Load checkpoint %s ****************' % ckp_path)
except IOError:
self.logger.error('********** No checkpoint %s!*********' % ckp_path)
return
self.CodeNet_I.load_state_dict(obj['ImgNet'])
self.CodeNet_T.load_state_dict(obj['TxtNet'])
self.logger.info('********** The loaded model has been trained for %d epochs.*********' % obj['step'])