File size: 8,483 Bytes
19b8775
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""
Pytorch implementation of basic sequence to Sequence modules.
"""

import logging
import torch
import torch.nn as nn
import math
import numpy as np

import stanza.models.common.seq2seq_constant as constant

logger = logging.getLogger('stanza')

class BasicAttention(nn.Module):
    """
    A basic MLP attention layer.
    """
    def __init__(self, dim):
        super(BasicAttention, self).__init__()
        self.linear_in = nn.Linear(dim, dim, bias=False)
        self.linear_c = nn.Linear(dim, dim)
        self.linear_v = nn.Linear(dim, 1, bias=False)
        self.linear_out = nn.Linear(dim * 2, dim, bias=False)
        self.tanh = nn.Tanh()
        self.sm = nn.Softmax(dim=1)

    def forward(self, input, context, mask=None, attn_only=False):
        """
        input: batch x dim
        context: batch x sourceL x dim
        """
        batch_size = context.size(0)
        source_len = context.size(1)
        dim = context.size(2)
        target = self.linear_in(input) # batch x dim
        source = self.linear_c(context.contiguous().view(-1, dim)).view(batch_size, source_len, dim)
        attn = target.unsqueeze(1).expand_as(context) + source
        attn = self.tanh(attn) # batch x sourceL x dim
        attn = self.linear_v(attn.view(-1, dim)).view(batch_size, source_len)

        if mask is not None:
            attn.masked_fill_(mask, -constant.INFINITY_NUMBER)

        attn = self.sm(attn)
        if attn_only:
            return attn

        weighted_context = torch.bmm(attn.unsqueeze(1), context).squeeze(1)
        h_tilde = torch.cat((weighted_context, input), 1)
        h_tilde = self.tanh(self.linear_out(h_tilde))

        return h_tilde, attn

class SoftDotAttention(nn.Module):
    """Soft Dot Attention.

    Ref: http://www.aclweb.org/anthology/D15-1166
    Adapted from PyTorch OPEN NMT.
    """

    def __init__(self, dim):
        """Initialize layer."""
        super(SoftDotAttention, self).__init__()
        self.linear_in = nn.Linear(dim, dim, bias=False)
        self.sm = nn.Softmax(dim=1)
        self.linear_out = nn.Linear(dim * 2, dim, bias=False)
        self.tanh = nn.Tanh()
        self.mask = None

    def forward(self, input, context, mask=None, attn_only=False, return_logattn=False):
        """Propagate input through the network.

        input: batch x dim
        context: batch x sourceL x dim
        """
        target = self.linear_in(input).unsqueeze(2)  # batch x dim x 1

        # Get attention
        attn = torch.bmm(context, target).squeeze(2)  # batch x sourceL

        if mask is not None:
            # sett the padding attention logits to -inf
            assert mask.size() == attn.size(), "Mask size must match the attention size!"
            attn.masked_fill_(mask, -constant.INFINITY_NUMBER)

        if return_logattn:
            attn = torch.log_softmax(attn, 1)
            attn_w = torch.exp(attn)
        else:
            attn = self.sm(attn)
            attn_w = attn
        if attn_only:
            return attn

        attn3 = attn_w.view(attn_w.size(0), 1, attn_w.size(1))  # batch x 1 x sourceL

        weighted_context = torch.bmm(attn3, context).squeeze(1)  # batch x dim
        h_tilde = torch.cat((weighted_context, input), 1)

        h_tilde = self.tanh(self.linear_out(h_tilde))

        return h_tilde, attn


class LinearAttention(nn.Module):
    """ A linear attention form, inspired by BiDAF:
        a = W (u; v; u o v)
    """

    def __init__(self, dim):
        super(LinearAttention, self).__init__()
        self.linear = nn.Linear(dim*3, 1, bias=False)
        self.linear_out = nn.Linear(dim * 2, dim, bias=False)
        self.sm = nn.Softmax(dim=1)
        self.tanh = nn.Tanh()
        self.mask = None

    def forward(self, input, context, mask=None, attn_only=False):
        """
        input: batch x dim
        context: batch x sourceL x dim
        """
        batch_size = context.size(0)
        source_len = context.size(1)
        dim = context.size(2)
        u = input.unsqueeze(1).expand_as(context).contiguous().view(-1, dim)  # batch*sourceL x dim
        v = context.contiguous().view(-1, dim)
        attn_in = torch.cat((u, v, u.mul(v)), 1)
        attn = self.linear(attn_in).view(batch_size, source_len)

        if mask is not None:
            # sett the padding attention logits to -inf
            assert mask.size() == attn.size(), "Mask size must match the attention size!"
            attn.masked_fill_(mask, -constant.INFINITY_NUMBER)

        attn = self.sm(attn)
        if attn_only:
            return attn

        attn3 = attn.view(batch_size, 1, source_len)  # batch x 1 x sourceL

        weighted_context = torch.bmm(attn3, context).squeeze(1)  # batch x dim
        h_tilde = torch.cat((weighted_context, input), 1)
        h_tilde = self.tanh(self.linear_out(h_tilde))
        return h_tilde, attn

class DeepAttention(nn.Module):
    """ A deep attention form, invented by Robert:
        u = ReLU(Wx)
        v = ReLU(Wy)
        a = V.(u o v)
    """

    def __init__(self, dim):
        super(DeepAttention, self).__init__()
        self.linear_in = nn.Linear(dim, dim, bias=False)
        self.linear_v = nn.Linear(dim, 1, bias=False)
        self.linear_out = nn.Linear(dim * 2, dim, bias=False)
        self.relu = nn.ReLU()
        self.sm = nn.Softmax(dim=1)
        self.tanh = nn.Tanh()
        self.mask = None

    def forward(self, input, context, mask=None, attn_only=False):
        """
        input: batch x dim
        context: batch x sourceL x dim
        """
        batch_size = context.size(0)
        source_len = context.size(1)
        dim = context.size(2)
        u = input.unsqueeze(1).expand_as(context).contiguous().view(-1, dim)  # batch*sourceL x dim
        u = self.relu(self.linear_in(u))
        v = self.relu(self.linear_in(context.contiguous().view(-1, dim)))
        attn = self.linear_v(u.mul(v)).view(batch_size, source_len)

        if mask is not None:
            # sett the padding attention logits to -inf
            assert mask.size() == attn.size(), "Mask size must match the attention size!"
            attn.masked_fill_(mask, -constant.INFINITY_NUMBER)

        attn = self.sm(attn)
        if attn_only:
            return attn

        attn3 = attn.view(batch_size, 1, source_len)  # batch x 1 x sourceL

        weighted_context = torch.bmm(attn3, context).squeeze(1)  # batch x dim
        h_tilde = torch.cat((weighted_context, input), 1)
        h_tilde = self.tanh(self.linear_out(h_tilde))
        return h_tilde, attn

class LSTMAttention(nn.Module):
    r"""A long short-term memory (LSTM) cell with attention."""

    def __init__(self, input_size, hidden_size, batch_first=True, attn_type='soft'):
        """Initialize params."""
        super(LSTMAttention, self).__init__()
        self.input_size = input_size
        self.hidden_size = hidden_size
        self.batch_first = batch_first

        self.lstm_cell = nn.LSTMCell(input_size, hidden_size)

        if attn_type == 'soft':
            self.attention_layer = SoftDotAttention(hidden_size)
        elif attn_type == 'mlp':
            self.attention_layer = BasicAttention(hidden_size)
        elif attn_type == 'linear':
            self.attention_layer = LinearAttention(hidden_size)
        elif attn_type == 'deep':
            self.attention_layer = DeepAttention(hidden_size)
        else:
            raise Exception("Unsupported LSTM attention type: {}".format(attn_type))
        logger.debug("Using {} attention for LSTM.".format(attn_type))

    def forward(self, input, hidden, ctx, ctx_mask=None, return_logattn=False):
        """Propagate input through the network."""
        if self.batch_first:
            input = input.transpose(0,1)

        output = []
        attn = []
        steps = range(input.size(0))
        for i in steps:
            hidden = self.lstm_cell(input[i], hidden)
            hy, cy = hidden
            h_tilde, alpha = self.attention_layer(hy, ctx, mask=ctx_mask, return_logattn=return_logattn)
            output.append(h_tilde)
            attn.append(alpha)
        output = torch.cat(output, 0).view(input.size(0), *output[0].size())

        if self.batch_first:
            output = output.transpose(0,1)

        if return_logattn:
            attn = torch.stack(attn, 0)
            if self.batch_first:
                attn = attn.transpose(0, 1)
            return output, hidden, attn

        return output, hidden