|
1
|
|
|
#!/usr/bin/env python |
|
2
|
|
|
# -*- coding: utf-8 -*- |
|
3
|
|
|
|
|
4
|
|
|
from deepy import * |
|
5
|
|
|
import theano.tensor as T |
|
6
|
|
|
|
|
7
|
|
|
class AggregationLayer(NeuralLayer): |
|
8
|
|
|
""" |
|
9
|
|
|
Aggregation layer. |
|
10
|
|
|
""" |
|
11
|
|
|
|
|
12
|
|
|
def __init__(self, size, activation='relu', init=None, layers=3): |
|
13
|
|
|
super(AggregationLayer, self).__init__("aggregation") |
|
14
|
|
|
self.size = size |
|
15
|
|
|
self.activation = activation |
|
16
|
|
|
self.init = init |
|
17
|
|
|
self.layers = layers |
|
18
|
|
|
|
|
19
|
|
|
def prepare(self): |
|
20
|
|
|
self.output_dim = self.size |
|
21
|
|
|
self._act = build_activation(self.activation) |
|
22
|
|
|
self._inner_layers = [Dense(self.size, self.activation, init=self.init).connect(self.input_dim)] |
|
23
|
|
|
for _ in range(self.layers - 1): |
|
24
|
|
|
self._inner_layers.append(Dense(self.size, self.activation, init=self.init).connect(self.size)) |
|
25
|
|
|
self.register_inner_layers(*self._inner_layers) |
|
26
|
|
|
|
|
27
|
|
|
self._chain2 = Chain(self.input_dim).stack( |
|
28
|
|
|
Dense(self.size, self.activation, init=self.init), |
|
29
|
|
|
Dense(self.layers, 'linear', init=self.init), |
|
30
|
|
|
Softmax() |
|
31
|
|
|
) |
|
32
|
|
|
|
|
33
|
|
|
self.register_inner_layers(self._chain2) |
|
34
|
|
|
self._dropout = Dropout(0.1) |
|
35
|
|
|
|
|
36
|
|
|
def _output(self, x, test=False): |
|
37
|
|
|
seq = [] |
|
38
|
|
|
v = x |
|
39
|
|
|
for layer in self._inner_layers: |
|
40
|
|
|
v = layer.call(v, test) |
|
41
|
|
|
v = self._dropout.call(v, test) |
|
42
|
|
|
seq.append(v.dimshuffle(0, "x", 1)) |
|
43
|
|
|
|
|
44
|
|
|
seq_v = T.concatenate(seq, axis=1) |
|
45
|
|
|
|
|
46
|
|
|
eva = self._chain2.call(x, test) |
|
47
|
|
|
|
|
48
|
|
|
result = seq_v * eva.dimshuffle((0, 1, "x")) |
|
49
|
|
|
result = result.sum(axis=1) |
|
50
|
|
|
return result |
|
51
|
|
|
|
|
52
|
|
|
def output(self, x): |
|
53
|
|
|
return self._output(x, False) |
|
54
|
|
|
|
|
55
|
|
|
def test_output(self, x): |
|
56
|
|
|
return self._output(x, True) |
|
57
|
|
|
|
|
58
|
|
|
|
|
59
|
|
|
|
|
60
|
|
|
|
uaca /
deepy
