|
1
|
|
|
#!/usr/bin/env python |
|
2
|
|
|
# -*- coding: utf-8 -*- |
|
3
|
|
|
|
|
4
|
|
|
from layer import NeuralLayer |
|
5
|
|
|
import deepy.tensor as T |
|
6
|
|
|
|
|
7
|
|
|
class Attention(NeuralLayer): |
|
8
|
|
|
|
|
9
|
|
|
|
|
10
|
|
|
def __init__(self, hidden_size, input_dim): |
|
11
|
|
|
super(Attention, self).__init__("attention") |
|
12
|
|
|
self.input_dim = input_dim if input_dim else hidden_size |
|
13
|
|
|
self.hidden_size = hidden_size |
|
14
|
|
|
self.init(input_dim) |
|
15
|
|
|
|
|
16
|
|
|
def prepare(self): |
|
17
|
|
|
self.Ua = self.create_weight(self.input_dim, self.hidden_size, "ua") |
|
18
|
|
|
self.Wa = self.create_weight(self.hidden_size, self.hidden_size, "wa") |
|
19
|
|
|
self.Va = self.create_weight(label="va", shape=(self.hidden_size,)) |
|
20
|
|
|
self.register_parameters(self.Va, self.Wa, self.Ua) |
|
21
|
|
|
|
|
22
|
|
|
|
|
23
|
|
|
def precompute(self, inputs): |
|
24
|
|
|
""" |
|
25
|
|
|
Precompute partial values in the score function. |
|
26
|
|
|
""" |
|
27
|
|
|
return T.dot(inputs, self.Ua) |
|
28
|
|
|
|
|
29
|
|
|
def compute_alignments(self, prev_state, precomputed_values, mask=None): |
|
30
|
|
|
""" |
|
31
|
|
|
Compute the alignment weights based on the previous state. |
|
32
|
|
|
""" |
|
33
|
|
|
|
|
34
|
|
|
WaSp = T.dot(prev_state, self.Wa) |
|
35
|
|
|
UaH = precomputed_values |
|
36
|
|
|
# For test time the UaH will be (time, output_dim) |
|
37
|
|
|
if UaH.ndim == 2: |
|
38
|
|
|
preact = WaSp[:, None, :] + UaH[None, :, :] |
|
39
|
|
|
else: |
|
40
|
|
|
preact = WaSp[:, None, :] + UaH |
|
41
|
|
|
act = T.activate(preact, 'tanh') |
|
42
|
|
|
align_scores = T.dot(act, self.Va) # ~ (batch, time) |
|
43
|
|
|
if mask: |
|
44
|
|
|
mask = (1 - mask) * -99.00 |
|
45
|
|
|
if align_scores.ndim == 3: |
|
46
|
|
|
align_scores += mask[None, :] |
|
47
|
|
|
else: |
|
48
|
|
|
align_scores += mask |
|
49
|
|
|
align_weights = T.nnet.softmax(align_scores) |
|
50
|
|
|
return align_weights |
|
51
|
|
|
|
|
52
|
|
|
def compute_context_vector(self, prev_state, inputs, precomputed_values=None, mask=None): |
|
53
|
|
|
""" |
|
54
|
|
|
Compute the context vector with soft attention. |
|
55
|
|
|
""" |
|
56
|
|
|
precomputed_values = precomputed_values if precomputed_values else self.precompute(inputs) |
|
57
|
|
|
align_weights = self.compute_alignments(prev_state, precomputed_values, mask) |
|
58
|
|
|
context_vector = T.sum(align_weights[:, :, None] * inputs, axis=1) |
|
59
|
|
|
return context_vector |
zomux /
deepy
