SCH/utils.py at main · hutt94/SCH · GitHub

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
import numpy as np
import torch
from torch.autograd import Variable


def compress(train_loader, test_loader, model_I, model_T, train_dataset, test_dataset):
    re_BI = list([])
    re_BT = list([])
    re_L = list([])
    for _, (data_I, data_T, _, _) in enumerate(train_loader):
        with torch.no_grad():
            var_data_I = Variable(data_I.cuda())
            _, _, code_I = model_I(var_data_I)
        code_I = torch.sign(code_I)
        re_BI.extend(code_I.cpu().data.numpy())

        var_data_T = Variable(torch.FloatTensor(data_T.numpy()).cuda())
        _, _, code_T = model_T(var_data_T)
        code_T = torch.sign(code_T)
        re_BT.extend(code_T.cpu().data.numpy())

    qu_BI = list([])
    qu_BT = list([])
    qu_L = list([])
    for _, (data_I, data_T, _, _) in enumerate(test_loader):
        with torch.no_grad():
            var_data_I = Variable(data_I.cuda())
            _, _, code_I = model_I(var_data_I)
        code_I = torch.sign(code_I)
        qu_BI.extend(code_I.cpu().data.numpy())

        var_data_T = Variable(torch.FloatTensor(data_T.numpy()).cuda())
        _, _, code_T = model_T(var_data_T)
        code_T = torch.sign(code_T)
        qu_BT.extend(code_T.cpu().data.numpy())

    re_BI = np.array(re_BI)
    re_BT = np.array(re_BT)
    re_L = train_dataset.train_labels

    qu_BI = np.array(qu_BI)
    qu_BT = np.array(qu_BT)
    qu_L = test_dataset.train_labels
    return re_BI, re_BT, re_L, qu_BI, qu_BT, qu_L


def calculate_hamming(B1, B2):
    """
    :param B1:  vector [n]
    :param B2:  vector [r*n]
    :return: hamming distance [r]
    """
    leng = B2.shape[1]  # max inner product value
    distH = 0.5 * (leng - np.dot(B1, B2.transpose()))
    return distH


def calculate_map(qu_B, re_B, qu_L, re_L):
    """
       :param qu_B: {-1,+1}^{mxq} query bits
       :param re_B: {-1,+1}^{nxq} retrieval bits
       :param qu_L: {0,1}^{mxl} query label
       :param re_L: {0,1}^{nxl} retrieval label
       :return:
    """
    num_query = qu_L.shape[0]
    map = 0
    top_result = 100
    results = []

    for iter in range(num_query):
        gnd = (np.dot(qu_L[iter, :], re_L.transpose()) > 0).astype(np.float32)
        tsum = int(np.sum(gnd))
        if tsum == 0:
            continue
        hamm = calculate_hamming(qu_B[iter, :], re_B)
        ind = np.argsort(hamm)
        results.append(ind[:top_result])
        gnd = gnd[ind]

        count = np.linspace(1, tsum, tsum)  # [1,2, tsum]
        tindex = np.asarray(np.where(gnd == 1)) + 1.0
        map_ = np.mean(count / (tindex))
        map = map + map_
    map = map / num_query
    return map, np.array(results)


def calculate_top_map(qu_B, re_B, qu_L, re_L, topk):
    """
    :param qu_B: {-1,+1}^{mxq} query bits
    :param re_B: {-1,+1}^{nxq} retrieval bits
    :param qu_L: {0,1}^{mxl} query label
    :param re_L: {0,1}^{nxl} retrieval label
    :param topk:
    :return:
    """
    num_query = qu_L.shape[0]
    topkmap = 0
    for iter in range(num_query):
        gnd = (np.dot(qu_L[iter, :], re_L.transpose()) > 0).astype(np.float32)
        hamm = calculate_hamming(qu_B[iter, :], re_B)
        ind = np.argsort(hamm)
        gnd = gnd[ind]

        tgnd = gnd[0:topk]
        tsum = int(np.sum(tgnd))
        if tsum == 0:
            continue
        count = np.linspace(1, tsum, tsum)
        tindex = np.asarray(np.where(tgnd == 1)) + 1.0
        topkmap_ = np.mean(count / (tindex))
        topkmap = topkmap + topkmap_
    topkmap = topkmap / num_query
    return topkmap


def calc_hamming_dist(B1, B2):
    q = B2.shape[1]
    if len(B1.shape) < 2:
        B1 = B1.unsqueeze(0)
    distH = 0.5 * (q - B1.mm(B2.t()))
    return distH


def p_topK(qB, rB, query_label, retrieval_label, K):
    num_query = query_label.shape[0]
    p = [0] * len(K)
    query_label = torch.Tensor(query_label)
    retrieval_label = torch.Tensor(retrieval_label)
    qB = torch.Tensor(qB)
    rB = torch.Tensor(rB)
    for iter in range(num_query):
        gnd = (query_label[iter].unsqueeze(0).mm(retrieval_label.t()) > 0).float().squeeze()
        tsum = torch.sum(gnd)
        if tsum == 0:
            continue
        hamm = calc_hamming_dist(qB[iter, :], rB).squeeze()
        hamm = torch.Tensor(hamm)
        for i in range(len(K)):
            total = min(K[i], retrieval_label.shape[0])
            ind = torch.sort(hamm).indices[:int(total)]
            gnd_ = gnd[ind]
            p[i] += gnd_.sum() / total
    p = torch.Tensor(p) / num_query
    return p


def pr_Curve(qB, rB, query_label, retrieval_label, K):
    num_query = query_label.shape[0]
    p = [0] * len(K)
    r = [0] * len(K)
    query_label = torch.Tensor(query_label)
    retrieval_label = torch.Tensor(retrieval_label)
    qB = torch.Tensor(qB)
    rB = torch.Tensor(rB)
    for iter in range(num_query):
        gnd = (query_label[iter].unsqueeze(0).mm(retrieval_label.t()) > 0).float().squeeze()
        tsum = torch.sum(gnd)
        if tsum == 0:
            continue
        hamm = calc_hamming_dist(qB[iter, :], rB).squeeze()
        hamm = torch.Tensor(hamm)
        for i in range(len(K)):
            total = min(K[i], retrieval_label.shape[0])
            ind = torch.sort(hamm).indices[:int(total)]
            gnd_ = gnd[ind]
            p[i] += gnd_.sum() / total
            r[i] += gnd_.sum() / tsum
    p = torch.Tensor(p) / num_query
    r = torch.Tensor(r) / num_query
    return p, r