|
1
|
|
|
#!/usr/bin/env python |
|
2
|
|
|
# -*- coding: utf-8 -*- |
|
3
|
|
|
|
|
4
|
|
|
from layer import NeuralLayer |
|
5
|
|
|
|
|
6
|
|
|
|
|
7
|
|
|
class Chain(NeuralLayer): |
|
8
|
|
|
""" |
|
9
|
|
|
Stack many layers to form a chain. |
|
10
|
|
|
This is useful to reuse layers in a customized layer. |
|
11
|
|
|
Usage: |
|
12
|
|
|
As part of the main pipe line: |
|
13
|
|
|
chain = Chain(layer1, layer2) |
|
14
|
|
|
model.stack(chain) |
|
15
|
|
|
As part of the computational graph: |
|
16
|
|
|
chain = Chain(layer1, layer2) |
|
17
|
|
|
y = chain.compute(x) |
|
18
|
|
|
""" |
|
19
|
|
|
|
|
20
|
|
|
def __init__(self, *layers): |
|
21
|
|
|
super(Chain, self).__init__("chain") |
|
22
|
|
|
self.layers = [] |
|
23
|
|
|
self._layers_to_stack = [] |
|
24
|
|
|
if len(layers) == 1 and type(layers[0]) == int: |
|
25
|
|
|
# This is a deprecated using of Chain |
|
26
|
|
|
self.input_dim = layers[0] |
|
27
|
|
|
else: |
|
28
|
|
|
self.stack(*layers) |
|
29
|
|
|
|
|
30
|
|
|
def stack(self, *layers): |
|
31
|
|
|
if self.input_dim is None or self.input_dim == 0: |
|
32
|
|
|
# Don't know the input dimension until connect |
|
33
|
|
|
self._layers_to_stack.extend(layers) |
|
34
|
|
|
else: |
|
35
|
|
|
self._register_layers(*layers) |
|
36
|
|
|
return self |
|
37
|
|
|
|
|
38
|
|
|
def _register_layers(self, *layers): |
|
39
|
|
|
for layer in layers: |
|
40
|
|
|
if not self.layers: |
|
41
|
|
|
layer.connect(self.input_dim) |
|
42
|
|
|
else: |
|
43
|
|
|
layer.connect(self.layers[-1].output_dim) |
|
44
|
|
|
self.layers.append(layer) |
|
45
|
|
|
self.output_dim = layer.output_dim |
|
46
|
|
|
self.register_inner_layers(*self.layers) |
|
47
|
|
|
|
|
48
|
|
|
def prepare(self, *layers): |
|
49
|
|
|
if self._layers_to_stack: |
|
50
|
|
|
self._register_layers(*self._layers_to_stack) |
|
51
|
|
|
self._layers_to_stack = [] |
|
52
|
|
|
|
|
53
|
|
|
def output(self, x): |
|
54
|
|
|
return self._output(x, False) |
|
55
|
|
|
|
|
56
|
|
|
def test_output(self, x): |
|
57
|
|
|
return self._output(x, True) |
|
58
|
|
|
|
|
59
|
|
|
def _output(self, x, test): |
|
60
|
|
|
y = x |
|
61
|
|
|
for layer in self.layers: |
|
62
|
|
|
y = layer.call(y, test=test) |
|
63
|
|
|
return y |
|
64
|
|
|
|
uaca /
deepy
