-
Notifications
You must be signed in to change notification settings - Fork 1
/
cedh.py
282 lines (234 loc) · 10.6 KB
/
cedh.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
import os
import time
import torch
import torch.optim as optim
import torch.nn.functional as F
from loguru import logger
from torch.optim.lr_scheduler import CosineAnnealingLR
import models.alexnet as alexnet
import utils.evaluate as evaluate
from data.data_loader import sample_dataloader
from utils.sim_matrix import cosine_S, smooth_S
import torch.nn as nn
class CEDH_Loss(nn.Module):
"""
Loss function of ADSH
Args:
code_length(int): Hashing code length.
gamma(float): Hyper-parameter.
"""
def __init__(self, code_length, gamma):
super(CEDH_Loss, self).__init__()
self.code_length = code_length
self.gamma = gamma
def forward(self, F, B, S, omega):
hash_loss = ((self.code_length * S - F @ B.t()) ** 2).sum()
quantization_loss = ((F - B[omega, :]) ** 2).sum()
loss = (hash_loss + self.gamma * quantization_loss) / (F.shape[0] * B.shape[0])
return loss
def train(
dataset,
dataset_root,
query_dataloader,
train_dataloader,
retrieval_dataloader,
original_method,
original_code_dir,
original_length,
target_length,
max_iter,
max_epoch,
batch_size,
num_samples,
lr,
W_lambda,
gamma,
topk,
device,
sim_S,
alpha,
eval_epoch,
save_name,
):
"""
Training model.
Args
dataset(str): name of dataset.
dataset_root(str): root path of dataset.
query_dataloader, train_dataloader, retrieval_dataloader(torch.utils.data.dataloader.DataLoader): Data loader.
original_method(str): original method used to generate Hashing code.
code_dir(str): dir path of original Hashing code.
original_length(int): original Hashing code length.
target_length(int): target Hashing code length.
num_samples(int): number of data sampled to train cnn.
batch_size(int): number of images in a batch.
lr(float): learning rate.
max_iter(int): number of iterations to train the whole framework.
max_epoch(int): number of epoch to train cnn.
W_lambda(float): hyper-parameter to regularize matrix W.
alpha(float): hyper-parameter to soft the constraint in multi-label scence.
gamma(float): hyper-parameter to trade-off quantization loss.
topk(int): Topk k map.
device(torch.device): GPU or CPU.
save_name(str): name of checkpoint save dir.
Returns
mAP(float): Mean Average Precision.
"""
# Initialization
step_size = target_length - original_length
if original_method == 'cedh':
B1 = torch.load(os.path.join(original_code_dir, 'training_code{}.t'.format(original_length))).to(device)
original_retrieval_code = torch.load(os.path.join(original_code_dir, 'retrieval_code{}.t'.format(original_length))).to(device)
retrieval_targets = torch.load(os.path.join(original_code_dir, 'retrieval_targets{}.t'.format(original_length))).to(device)
else:
#* original hashing code of training data
B1 = torch.load(os.path.join(original_code_dir, 'training_code{}.t'.format(original_length))).to(device)
#* original hashing code of retrieval data
original_retrieval_code = torch.load(os.path.join(original_code_dir, 'retrieval_code{}.t'.format(original_length))).to(device)
#* targets of retrieval data
retrieval_targets = torch.load(os.path.join(original_code_dir, 'retrieval_targets{}.t'.format(original_length))).to(device)
# retrieval_targets = retrieval_dataloader.dataset.get_onehot_targets().to(device)
# import ipdb;ipdb.set_trace()
model = alexnet.load_model(target_length).to(device)
criterion = CEDH_Loss(target_length, gamma)
optimizer = optim.Adam(
model.parameters(),
lr=lr,
weight_decay=1e-5,
)
scheduler = CosineAnnealingLR(optimizer, max_iter, 1e-7)
U = torch.zeros(num_samples, target_length).to(device)
W = torch.rand(original_length, step_size).to(device)
B2 = (B1@W)
B = torch.cat((B1, B2), 1)
Z = torch.zeros(B.shape[0], step_size).to(device)
best_mAP = 0.0
timestr = time.strftime('%Y-%m-%d-%H:%M', time.gmtime())
savedir = os.path.join('checkpoints', 'cedh', dataset, save_name+'-'+timestr)
if not os.path.exists(savedir):
os.makedirs(savedir)
for it in range(max_iter):
#! evaluate and save best
# '''
if (it)%eval_epoch == 0 or it == max_iter-1:
query_code = generate_code(model, query_dataloader, target_length, device)
new_retrieval_code = (original_retrieval_code @ W).sign()
retrieval_code = torch.cat((original_retrieval_code, new_retrieval_code),1)
mAP = evaluate.mean_average_precision(
query_code.to(device),
retrieval_code,
query_dataloader.dataset.get_onehot_targets().to(device),
retrieval_targets,
device,
topk,
)
#!save best
if mAP >= best_mAP:
best_mAP = mAP
training_code = generate_code(model, train_dataloader, target_length, device)
torch.save(training_code.cpu(), os.path.join(savedir, 'training_code{}.t'.format(target_length)))
torch.save(query_code.cpu(), os.path.join(savedir, 'query_code{}.t'.format(target_length)))
query_targets = query_dataloader.dataset.get_onehot_targets().to(device)
torch.save(query_targets.cpu(), os.path.join(savedir, 'query_targets{}.t'.format(target_length)))
# retrieval集扩展后新的长度的hashcode
torch.save(retrieval_code.cpu(), os.path.join(savedir, 'retrieval_code{}.t'.format(target_length)))
torch.save(retrieval_targets.cpu(), os.path.join(savedir, 'retrieval_targets{}.t'.format(target_length)))
torch.save(model, os.path.join(savedir, 'model-{}.t'.format(target_length)))
torch.save(W.cpu(), os.path.join(savedir, 'W-{}.t'.format(target_length)))
logger.info('[iter:{}/{}][mAP:{:4f}]'.format(it+1, max_iter, mAP))
#! evaluate end
# '''
iter_start = time.time()
#* Sample training data for cnn learning
training_dataloader, sample_index = sample_dataloader(train_dataloader, num_samples, batch_size, dataset_root, dataset)
#* Create Similarity matrix
#* the targets of the sampled training data
training_targets = training_dataloader.dataset.get_onehot_targets().to(device)
#* the targets of the whole training data
train_targets = train_dataloader.dataset.get_onehot_targets().to(device)
if sim_S == 'cosine':
S = cosine_S(training_targets, train_targets)
elif sim_S == 'smooth':
S = smooth_S(training_targets, train_targets, alpha)
# Training CNN model
CNN_time_start = time.time()
for epoch in range(max_epoch):
for batch, (data, targets, index) in enumerate(training_dataloader):
data, targets, index = data.to(device), targets.to(device), index.to(device)
optimizer.zero_grad()
f = model(data)
U[index, :] = f.data
# cnn_loss = criterion(f, B, S[index, :], sample_index[index])
hash_loss = ((target_length * S[index, :] - f @ B.t()) **2).sum()
# import ipdb; ipdb.set_trace()
quantization_loss = ((f[:, :original_length] - B[sample_index[index]][:,:original_length]) ** 2).sum()+\
((f[:, original_length:] - Z[sample_index[index]])**2).sum()
cnn_loss = (hash_loss + gamma * quantization_loss) / (f.shape[0] * B.shape[0])
cnn_loss.backward()
optimizer.step()
scheduler.step()
CNN_time_end = time.time()
# update W
U1 = U[:, :original_length]
U2 = U[:, original_length:]
W_time_start = time.time()
temp1 = B1.t() @ B1 + W_lambda * torch.eye(original_length).to(device)
temp2 = target_length * S - U1 @ B1.t()
temp3 = U2.t() @ U2 + W_lambda * torch.eye(step_size).to(device)
W = torch.inverse(temp1) @ (B1.t() @ temp2.t() @ U2 + B1.t() @ Z) @ torch.inverse(temp3)
W_time_end = time.time()
# update Z
expand_U2 = torch.zeros(B2.shape).to(device)
expand_U2[sample_index] = U2
Z = (expand_U2 + B1 @ W).sign()
B2 = (B1 @ W)
B[:, original_length:] = B2
# Total loss
iter_loss = calc_loss(U, B, S, target_length, sample_index, gamma)
logger.debug(
'[iter:{}/{}][loss:{:.2f}][iter_time:{:.2f}]'.
format(it + 1, max_iter, iter_loss, time.time() - iter_start))
logger.debug(
'[iter:{}/{}][CNN_time:{:.2f}][W_time:{:.2f}]'.format(it + 1, max_iter, CNN_time_end-CNN_time_start, W_time_end-W_time_start))
logger.info('Best checkpoint saved at: {}'.format(savedir))
def solve_dcc(B1, B2, U1, U2, expand_U2, S, code_length, step_size, gamma, alpha, W):
"""
Solve DCC problem.
"""
Z = code_length * S - U1 @ (B1.t())
P = -2*(Z.t() @ U2 + gamma * expand_U2 + alpha * (B1 @ W))
for bit in range(step_size):
p = P[:, bit]
u2 = U2[:, bit]
B_prime = torch.cat((B2[:, :bit], B2[:, bit + 1:]), dim=1)
U2_prime = torch.cat((U2[:, :bit], U2[:, bit + 1:]), dim=1)
B2[:, bit] = -(2 * B_prime @ U2_prime.t() @ u2 + p).sign()
return B2
def calc_loss(U, B, S, code_length, omega, gamma):
"""
Calculate loss.
"""
hash_loss = ((code_length * S - U @ B.t()) ** 2).sum()
quantization_loss = ((U - B[omega, :]) ** 2).sum()
loss = (hash_loss + gamma * quantization_loss) / (U.shape[0] * B.shape[0])
return loss.item()
def generate_code(model, dataloader, code_length, device):
"""
Generate hash code
Args
dataloader(torch.utils.data.DataLoader): Data loader.
code_length(int): Hash code length.
device(torch.device): Using gpu or cpu.
Returns
code(torch.Tensor): Hash code.
"""
model.eval()
with torch.no_grad():
N = len(dataloader.dataset)
code = torch.zeros([N, code_length])
for data, _, index in dataloader:
data = data.to(device)
hash_code = model(data)
code[index, :] = hash_code.sign().cpu()
model.train()
return code