|
1
|
|
|
#!/usr/bin/env python |
|
2
|
|
|
# -*- coding: utf-8 -*- |
|
3
|
|
|
|
|
4
|
|
|
import theano.tensor as T |
|
5
|
|
|
from deepy import NeuralLayer, AutoEncoder, Dense |
|
6
|
|
|
from deepy import GaussianInitializer, global_theano_rand |
|
7
|
|
|
|
|
8
|
|
|
|
|
9
|
|
|
class ReparameterizationLayer(NeuralLayer): |
|
10
|
|
|
""" |
|
11
|
|
|
Reparameterization layer in a Variational encoder. |
|
12
|
|
|
Only binary output cost function is supported now. |
|
13
|
|
|
The prior value is recorded after the computation graph created. |
|
14
|
|
|
""" |
|
15
|
|
|
|
|
16
|
|
|
def __init__(self, size, sample=False): |
|
17
|
|
|
""" |
|
18
|
|
|
:param size: the size of latent variable |
|
19
|
|
|
:param sample: whether to get a clean latent variable |
|
20
|
|
|
""" |
|
21
|
|
|
super(ReparameterizationLayer, self).__init__("VariationalEncoder") |
|
22
|
|
|
self.size = size |
|
23
|
|
|
self.output_dim = size |
|
24
|
|
|
self.sample = sample |
|
25
|
|
|
self._prior = None |
|
26
|
|
|
|
|
27
|
|
|
def prepare(self): |
|
28
|
|
|
self._mu_encoder = Dense(self.size, 'linear', init=GaussianInitializer(), random_bias=True).connect(self.input_dim) |
|
29
|
|
|
self._log_sigma_encoder = Dense(self.size, 'linear', init=GaussianInitializer(), random_bias=True).connect(self.input_dim) |
|
30
|
|
|
self.register_inner_layers(self._mu_encoder, self._log_sigma_encoder) |
|
31
|
|
|
|
|
32
|
|
|
def output(self, x): |
|
33
|
|
|
# Compute p(z|x) |
|
34
|
|
|
mu = self._mu_encoder.output(x) |
|
35
|
|
|
log_sigma = 0.5 * self._log_sigma_encoder.output(x) |
|
36
|
|
|
self._prior = 0.5* T.sum(1 + 2*log_sigma - mu**2 - T.exp(2*log_sigma)) |
|
37
|
|
|
# Reparameterization |
|
38
|
|
|
eps = global_theano_rand.normal((x.shape[0], self.size)) |
|
39
|
|
|
|
|
40
|
|
|
if self.sample: |
|
41
|
|
|
z = mu |
|
42
|
|
|
else: |
|
43
|
|
|
z = mu + T.exp(log_sigma) * eps |
|
44
|
|
|
return z |
|
45
|
|
|
|
|
46
|
|
|
def prior(self): |
|
47
|
|
|
""" |
|
48
|
|
|
Get the prior value. |
|
49
|
|
|
""" |
|
50
|
|
|
return self._prior |
|
51
|
|
|
|
|
52
|
|
|
|
|
53
|
|
|
class VariationalAutoEncoder(AutoEncoder): |
|
54
|
|
|
""" |
|
55
|
|
|
Variational Auto Encoder. |
|
56
|
|
|
Only binary output cost function is supported now. |
|
57
|
|
|
""" |
|
58
|
|
|
|
|
59
|
|
|
def __init__(self, input_dim, latent_dim=None, sample=False): |
|
60
|
|
|
""" |
|
61
|
|
|
:param latent_dim: latent variable size, this is not required in training |
|
62
|
|
|
:param sample: add no randomness when this flag is on |
|
63
|
|
|
""" |
|
64
|
|
|
super(VariationalAutoEncoder, self).__init__(input_dim, rep_dim=latent_dim) |
|
65
|
|
|
self.sample = sample |
|
66
|
|
|
self._setup_monitors = True |
|
67
|
|
|
|
|
68
|
|
|
|
|
69
|
|
|
def stack_reparameterization_layer(self, layer_size): |
|
70
|
|
|
""" |
|
71
|
|
|
Perform reparameterization trick for latent variables. |
|
72
|
|
|
:param layer_size: the size of latent variable |
|
73
|
|
|
""" |
|
74
|
|
|
self.rep_layer = ReparameterizationLayer(layer_size, sample=self.sample) |
|
75
|
|
|
self.stack_encoders(self.rep_layer) |
|
76
|
|
|
|
|
77
|
|
|
def _cost_func(self, y): |
|
78
|
|
|
logpxz = - T.nnet.binary_crossentropy(y, self.input_variables[0]).sum() |
|
79
|
|
|
logp = logpxz + self.rep_layer.prior() |
|
80
|
|
|
# the lower bound is the mean value of logp |
|
81
|
|
|
cost = - logp |
|
82
|
|
|
if self._setup_monitors: |
|
83
|
|
|
self._setup_monitors = False |
|
84
|
|
|
self.training_monitors.append(("lower_bound", logp / y.shape[0])) |
|
85
|
|
|
self.testing_monitors.append(("lower_bound", logp / y.shape[0])) |
|
86
|
|
|
return cost |
|
87
|
|
|
|
uaca /
deepy
