VCWE/model.py at master · HSLCY/VCWE · 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
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.nn import init
import numpy as np

"""
    u_embedding: Embedding for center word.
    v_embedding: Embedding for neighbor words.
"""


class VCWEModel(nn.Module):

    def __init__(self, emb_size, emb_dimension, wordid2charid, char_size):
        super(VCWEModel, self).__init__()
        self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
        self.wordid2charid = wordid2charid
        self.emb_size = emb_size
        self.emb_dimension = emb_dimension
        self.u_embeddings = nn.Embedding(emb_size, emb_dimension, sparse=False)
        self.v_embeddings = nn.Embedding(emb_size, emb_dimension, sparse=False)
        self.char_embeddings = nn.Embedding(char_size, emb_dimension, sparse=False)
        self.cnn_model = CNNModel(32,32,self.emb_dimension)
        self.lstm_model = LSTMModel(self.emb_dimension)

        initrange = 1.0 / self.emb_dimension
        init.uniform_(self.u_embeddings.weight.data, -initrange, initrange)
        init.uniform_(self.v_embeddings.weight.data, -initrange, initrange)
        init.uniform_(self.char_embeddings.weight.data, -initrange, initrange)


    def forward(self, pos_u, pos_v, neg_v, img_data):
        img_emb = self.cnn_model.forward(img_data)
        emb_u = self.u_embeddings(pos_u)
        emb_v = self.v_embeddings(pos_v).mean(dim=1)
        emb_neg_v = self.v_embeddings(neg_v)
        pos_v=pos_v.view(-1).cpu()
        temp = self.wordid2charid[pos_v].reshape(1,-1)
        temp = torch.from_numpy(temp).to(self.device).long()
        lstm_input = img_emb[temp.reshape(1, -1)].view(len(pos_v), -1, self.emb_dimension)
        del temp
        lstm_input = torch.transpose(lstm_input, 0, 1)          #  self.data.maxwordlength, batch_size, embedding_dim
        emb_char_v = self.lstm_model.forward(lstm_input, len(pos_v))
        emb_char_v = emb_char_v.view(pos_u.size(0),-1,self.emb_dimension)
        emb_char_v = torch.mean(emb_char_v,dim=1)

        pos_neg_v=neg_v.view(-1).cpu()
        temp = self.wordid2charid[pos_neg_v].reshape(1,-1)
        temp = torch.from_numpy(temp).to(self.device).long()
        lstm_input2 = img_emb[temp.reshape(1, -1)].view(len(pos_neg_v), -1, self.emb_dimension)
        del temp
        lstm_input2 = torch.transpose(lstm_input2, 0, 1)
        emb_neg_char_v = self.lstm_model.forward(lstm_input2, len(pos_neg_v))
        emb_neg_char_v = emb_neg_char_v.view(pos_u.size(0),-1,self.emb_dimension)

        c_score = torch.sum(torch.mul(emb_u, emb_char_v), dim=1)
        c_score = torch.clamp(c_score, max=10, min=-10)
        c_score = -F.logsigmoid(c_score)

        neg_c_score = torch.bmm(emb_neg_char_v, emb_u.unsqueeze(2)).squeeze()
        neg_c_score = torch.clamp(neg_c_score, max=10, min=-10)
        neg_c_score = -torch.sum(F.logsigmoid(-neg_c_score), dim=1)

        score = torch.sum(torch.mul(emb_u, emb_v), dim=1)
        score = torch.clamp(score, max=10, min=-10)
        score = -F.logsigmoid(score)

        neg_score = torch.bmm(emb_neg_v, emb_u.unsqueeze(2)).squeeze()
        neg_score = torch.clamp(neg_score, max=10, min=-10)
        neg_score = -torch.sum(F.logsigmoid(-neg_score), dim=1)

        return torch.mean(c_score + neg_c_score + score + neg_score)

    def save_embedding(self, id2word, file_name):
        embedding = self.u_embeddings.weight.cpu().data.numpy()
        with open(file_name, 'w', encoding="utf-8") as f:
            f.write('%d %d\n' % (len(id2word), self.emb_dimension))
            for wid, w in id2word.items():
                e = ' '.join(map(lambda x: str(x), embedding[wid]))
                f.write('%s %s\n' % (w, e))


class LSTMModel(nn.Module):

    def __init__(self, emb_dimension, d_a=128):
        super(LSTMModel, self).__init__()
        self.emb_dimension = emb_dimension
        self.hidden_dim = emb_dimension
        self.lstm = nn.LSTM(input_size=self.emb_dimension, hidden_size=self.hidden_dim, num_layers=1, bidirectional=True)
        initrange = 1.0 / self.emb_dimension
        init.uniform_(self.lstm.all_weights[0][0], -initrange, initrange)
        init.uniform_(self.lstm.all_weights[0][1], -initrange, initrange)
        self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

        self.linear_first = torch.nn.Linear(2*self.hidden_dim, d_a)
        self.linear_first.bias.data.fill_(0)
        self.linear_second = torch.nn.Linear(d_a, 1)
        self.linear_second.bias.data.fill_(0)
        self.linear_third = torch.nn.Linear(2*self.hidden_dim, self.emb_dimension)
        self.linear_third.bias.data.fill_(0)

    def init_hidden(self, batch_size):
        # first is the hidden h
        # second is the cell c
        return (torch.zeros(2, batch_size, self.hidden_dim).to(self.device),
                torch.zeros(2, batch_size, self.hidden_dim).to(self.device))

    def forward(self, input, batch_size):
        self.hidden = self.init_hidden(batch_size)
        lstm_out, self.hidden = self.lstm(input, self.hidden)
        a=self.linear_first(lstm_out)
        a=torch.tanh(a)
        a=self.linear_second(a)
        a=F.softmax(a,dim=0)
        a=a.expand(5,batch_size,2*self.hidden_dim)
        y=(a*lstm_out).sum(dim=0)
        y=self.linear_third(y)
        return y

class CNNModel(nn.Module):
    def __init__(self, output1, output2, emb_dimension):    #output1=32 output2=32
        super(CNNModel, self).__init__()
        self.emb_dimension = emb_dimension
        self.conv1 = nn.Conv2d(1, output1, (3, 3))
        self.conv2 = nn.Conv2d(output1, output2, (3, 3))
        self.hidden2result = nn.Linear(output2*64, emb_dimension)
        self.bn1 = nn.BatchNorm2d(output1)
        self.bn2 = nn.BatchNorm2d(output2)
        self.bn3 = nn.BatchNorm1d(emb_dimension)
        initrange = 1.0 / self.emb_dimension
        initrange1 = 1e-4
        initrange2 = 1e-2
        init.uniform_(self.conv1.weight.data, -initrange1, initrange1)
        init.uniform_(self.conv2.weight.data, -initrange2, initrange2)
        init.uniform_(self.hidden2result.weight.data, -initrange, initrange)

    def forward(self, x):
        x = self.conv1(x)
        x = F.max_pool2d(self.bn1(x), 2)
        x = self.conv2(x)
        x = F.max_pool2d(self.bn2(x), 2)
        x = x.view(x.size()[0], -1)
        x = self.hidden2result(x)
        x = F.relu(self.bn3(x))
        return x