-
Notifications
You must be signed in to change notification settings - Fork 116
/
Copy pathPCDH.py
157 lines (120 loc) · 5.09 KB
/
PCDH.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
from utils.tools import *
import os
import torch
import torch.optim as optim
import torch.nn as nn
from torchvision import models
import time
import numpy as np
torch.multiprocessing.set_sharing_strategy('file_system')
# PCDH(Neurocomputing 2020)
# paper [Deep discrete hashing with pairwise correlation learning](https://www.sciencedirect.com/science/article/pii/S092523121931793X)
# [PCDH] epoch:720, bit:48, dataset:nuswide_21, MAP:0.653, Best MAP: 0.659
# [PCDH] epoch:1785, bit:48, dataset:cifar10-1, MAP:0.166, Best MAP: 0.168
def get_config():
config = {
"alpha": 1,
"beta": 1,
# "optimizer":{"type": optim.SGD, "optim_params": {"lr": 0.05, "weight_decay": 10 ** -5}},
"optimizer": {"type": optim.RMSprop, "optim_params": {"lr": 1e-5, "weight_decay": 10 ** -5}},
"info": "[PCDH]",
"resize_size": 144,
"crop_size": 128,
"batch_size": 64,
"net": Net,
# "dataset": "cifar10-1",
"dataset": "nuswide_21",
"epoch": 2000,
"test_map": 15,
# "device":torch.device("cpu"),
"device": torch.device("cuda:1"),
"bit_list": [48],
}
config = config_dataset(config)
return config
class Net(nn.Module):
def __init__(self, hash_bit, num_classes, pretrained=True):
super(Net, self).__init__()
self.conv_layer = nn.Sequential(
nn.Conv2d(3, 128, kernel_size=3, padding=1),
nn.BatchNorm2d(128),
nn.ReLU(inplace=True),
nn.MaxPool2d((2, 2)),
nn.Conv2d(128, 128, kernel_size=3, padding=1),
nn.BatchNorm2d(128),
nn.ReLU(inplace=True),
nn.MaxPool2d((2, 2)),
nn.Conv2d(128, 256, kernel_size=3, padding=1),
nn.BatchNorm2d(256),
nn.ReLU(inplace=True),
nn.MaxPool2d((2, 2)),
nn.Conv2d(256, 256, kernel_size=3, padding=1),
nn.BatchNorm2d(256),
nn.ReLU(inplace=True),
nn.MaxPool2d((2, 2)),
)
self.feature_layer = nn.Linear(8 * 8 * 256, 1024)
self.hash_like_layer = nn.Sequential(nn.Linear(1024, hash_bit), nn.Tanh())
self.discrete_hash_layer = nn.Linear(hash_bit, hash_bit)
self.classification_layer = nn.Linear(hash_bit, num_classes, bias=False)
def forward(self, x, istraining=False):
x = self.conv_layer(x)
x = x.view(x.size(0), -1)
feature = self.feature_layer(x)
h = self.hash_like_layer(feature)
b = self.discrete_hash_layer(h).add(1).mul(0.5).clamp(min=0, max=1)
b = (b >= 0.5).float() * 2 - 1
y_pre = self.classification_layer(b)
if istraining:
return feature, h, y_pre
else:
return b
class PCDHLoss(torch.nn.Module):
def __init__(self, config, bit):
super(PCDHLoss, self).__init__()
self.m = 2 * bit
def forward(self, feature, h, y_pre, y, ind, config):
dist = (h.unsqueeze(1) - h.unsqueeze(0)).pow(2).sum(dim=2)
s = (y @ y.t() == 0).float()
loss1 = (1 - s) / 2 * dist + s / 2 * (self.m - dist).clamp(min=0).pow(2)
loss1 = loss1.mean()
dist2 = (feature.unsqueeze(1) - feature.unsqueeze(0)).pow(2).sum(dim=2)
loss2 = (1 - s) / 2 * dist2 + s / 2 * (self.m - dist2).clamp(min=0).pow(2)
loss2 = loss2.mean()
if "nuswide" in config["dataset"]:
Lc = (y_pre - y * y_pre + ((1 + (-y_pre).exp()).log())).sum(dim=1).mean()
else:
Lc = (-y_pre.softmax(dim=1).log() * y).sum(dim=1).mean()
return loss1 + config["alpha"] * loss2 + config["beta"] * Lc
def train_val(config, bit):
device = config["device"]
train_loader, test_loader, dataset_loader, num_train, num_test, num_dataset = get_data(config)
config["num_train"] = num_train
net = config["net"](bit, config["n_class"]).to(device)
optimizer = config["optimizer"]["type"](net.parameters(), **(config["optimizer"]["optim_params"]))
criterion = PCDHLoss(config, bit)
Best_mAP = 0
for epoch in range(config["epoch"]):
current_time = time.strftime('%H:%M:%S', time.localtime(time.time()))
print("%s[%2d/%2d][%s] bit:%d, dataset:%s, training...." % (
config["info"], epoch + 1, config["epoch"], current_time, bit, config["dataset"]), end="")
net.train()
train_loss = 0
for image, label, ind in train_loader:
image = image.to(device)
label = label.to(device)
optimizer.zero_grad()
feature, h, y_pre = net(image, istraining=True)
loss = criterion(feature, h, y_pre, label.float(), ind, config)
train_loss += loss.item()
loss.backward()
optimizer.step()
train_loss = train_loss / len(train_loader)
print("\b\b\b\b\b\b\b loss:%.3f" % (train_loss))
if (epoch + 1) % config["test_map"] == 0:
Best_mAP = validate(config, Best_mAP, test_loader, dataset_loader, net, bit, epoch, num_dataset)
if __name__ == "__main__":
config = get_config()
print(config)
for bit in config["bit_list"]:
train_val(config, bit)