|
1
|
|
|
#!/usr/bin/env python |
|
2
|
|
|
# -*- coding: utf-8 -*- |
|
3
|
|
|
|
|
4
|
|
|
from deepy.layers import NeuralLayer |
|
5
|
|
|
from deepy.utils import build_activation |
|
6
|
|
|
import theano.tensor as T |
|
7
|
|
|
|
|
8
|
|
|
class HighwayLayer(NeuralLayer): |
|
9
|
|
|
""" |
|
10
|
|
|
Highway network layer. |
|
11
|
|
|
See http://arxiv.org/abs/1505.00387. |
|
12
|
|
|
""" |
|
13
|
|
|
|
|
14
|
|
|
def __init__(self, activation='relu', init=None, gate_bias=-5): |
|
15
|
|
|
super(HighwayLayer, self).__init__("highway") |
|
16
|
|
|
self.activation = activation |
|
17
|
|
|
self.init = init |
|
18
|
|
|
self.gate_bias = gate_bias |
|
19
|
|
|
|
|
20
|
|
|
def prepare(self): |
|
21
|
|
|
self.output_dim = self.input_dim |
|
22
|
|
|
self._act = build_activation(self.activation) |
|
23
|
|
|
self.W_h = self.create_weight(self.input_dim, self.input_dim, "h", initializer=self.init) |
|
24
|
|
|
self.W_t = self.create_weight(self.input_dim, self.input_dim, "t", initializer=self.init) |
|
25
|
|
|
self.B_h = self.create_bias(self.input_dim, "h") |
|
26
|
|
|
self.B_t = self.create_bias(self.input_dim, "t", value=self.gate_bias) |
|
27
|
|
|
|
|
28
|
|
|
self.register_parameters(self.W_h, self.B_h, self.W_t, self.B_t) |
|
29
|
|
|
|
|
30
|
|
|
def output(self, x): |
|
31
|
|
|
t = self._act(T.dot(x, self.W_t) + self.B_t) |
|
32
|
|
|
h = self._act(T.dot(x, self.W_h) + self.B_h) |
|
33
|
|
|
return h * t + x * (1 - t) |
|
34
|
|
|
|
|
35
|
|
|
|
uaca /
deepy
