|
1
|
|
|
#!/usr/bin/env python |
|
2
|
|
|
# -*- coding: utf-8 -*- |
|
3
|
|
|
|
|
4
|
|
|
|
|
5
|
|
|
import logging as loggers |
|
6
|
|
|
|
|
7
|
|
|
import theano |
|
8
|
|
|
import theano.tensor as T |
|
9
|
|
|
|
|
10
|
|
|
from deepy.utils import build_activation |
|
11
|
|
|
from deepy.trainers import THEANO_LINKER |
|
12
|
|
|
from deepy.layers.layer import NeuralLayer |
|
13
|
|
|
|
|
14
|
|
|
logging = loggers.getLogger(__name__) |
|
15
|
|
|
|
|
16
|
|
|
from network import NeuralNetwork |
|
17
|
|
|
|
|
18
|
|
|
# TODO: repair additional_h mode |
|
19
|
|
|
class RecursiveAutoEncoder(NeuralNetwork): |
|
20
|
|
|
""" |
|
21
|
|
|
Recursive auto encoder (Recursively encode a sequence by combining two children). |
|
22
|
|
|
Parameters: |
|
23
|
|
|
rep_dim - dimension of representation |
|
24
|
|
|
""" |
|
25
|
|
|
def __init__(self, input_dim, rep_dim=None, activation='tanh', unfolding=True, additional_h=False, |
|
26
|
|
|
config=None): |
|
27
|
|
|
super(RecursiveAutoEncoder, self).__init__(input_dim, config=config, input_tensor=3) |
|
28
|
|
|
|
|
29
|
|
|
self.rep_dim = rep_dim |
|
30
|
|
|
self.stack(RecursiveAutoEncoderCore(rep_dim, unfolding=unfolding, additional_h=additional_h)) |
|
31
|
|
|
self._encode_func = None |
|
32
|
|
|
self._decode_func = None |
|
33
|
|
|
|
|
34
|
|
|
def _cost_func(self, y): |
|
35
|
|
|
# As the core returns cost |
|
36
|
|
|
return y |
|
37
|
|
|
|
|
38
|
|
|
@property |
|
39
|
|
|
def cost(self): |
|
40
|
|
|
return self._cost_func(self.output) |
|
41
|
|
|
|
|
42
|
|
|
@property |
|
43
|
|
|
def test_cost(self): |
|
44
|
|
|
return self._cost_func(self.test_output) |
|
45
|
|
|
|
|
46
|
|
|
def encode(self, x): |
|
47
|
|
|
""" |
|
48
|
|
|
Encode given input. |
|
49
|
|
|
""" |
|
50
|
|
|
if not self._encode_func: |
|
51
|
|
|
x_var = T.vector() |
|
52
|
|
|
self._encode_func = theano.function([x_var], self.layers[0].encode_func(x), |
|
53
|
|
|
allow_input_downcast=True, mode=theano.Mode(linker=THEANO_LINKER)) |
|
54
|
|
|
return self._encode_func(x) |
|
55
|
|
|
|
|
56
|
|
|
def decode(self, rep, n_steps): |
|
57
|
|
|
""" |
|
58
|
|
|
Decode given representation. |
|
59
|
|
|
""" |
|
60
|
|
|
if not self._decode_func: |
|
61
|
|
|
rep_var = T.vector() |
|
62
|
|
|
n_var = T.iscalar() |
|
63
|
|
|
self._decode_func = theano.function([rep_var, n_var], self.layers[0].decode_func(rep_var, n_var), |
|
64
|
|
|
allow_input_downcast=True, mode=theano.Mode(linker=THEANO_LINKER)) |
|
65
|
|
|
return self._decode_func(rep, n_steps) |
|
66
|
|
|
|
|
67
|
|
|
class RecursiveAutoEncoderCore(NeuralLayer): |
|
68
|
|
|
|
|
69
|
|
|
def __init__(self, rep_dim=None, activation='tanh', unfolding=True, additional_h=True): |
|
70
|
|
|
""" |
|
71
|
|
|
Binarized Recursive Encoder Core layer |
|
72
|
|
|
Input: |
|
73
|
|
|
A sequence of terminal nodes in vectore representations. |
|
74
|
|
|
Output: |
|
75
|
|
|
Cost |
|
76
|
|
|
""" |
|
77
|
|
|
super(RecursiveAutoEncoderCore, self).__init__("RAE") |
|
78
|
|
|
self.rep_dim = rep_dim |
|
79
|
|
|
self.unfolding = unfolding |
|
80
|
|
|
self.additional_h = additional_h |
|
81
|
|
|
self.activation = activation |
|
82
|
|
|
|
|
83
|
|
|
def prepare(self): |
|
84
|
|
|
self._setup_params() |
|
85
|
|
|
self._setup_functions() |
|
86
|
|
|
|
|
87
|
|
|
def output(self, x): |
|
88
|
|
|
rep, cost = self._recursive_func(x) |
|
89
|
|
|
self.register_monitors(("mean(rep)", abs(rep).mean())) |
|
90
|
|
|
return cost |
|
91
|
|
|
|
|
92
|
|
|
def _recursive_step(self, i, p, x): |
|
93
|
|
|
x_t = x[i] |
|
94
|
|
|
# Encoding |
|
95
|
|
|
rep = self._activation_func(T.dot(p, self.W_e1) + T.dot(x_t, self.W_e2) + self.B_e) |
|
96
|
|
|
if self.unfolding: |
|
97
|
|
|
x_decs = self._unfold(rep, i) |
|
98
|
|
|
distance = T.sum((x_decs - x[: i + 1]) ** 2) |
|
99
|
|
|
else: |
|
100
|
|
|
# Decoding |
|
101
|
|
|
p_dec, x_dec = self._decode_step(rep) |
|
102
|
|
|
# Euclidean distance |
|
103
|
|
|
distance = T.sum((p_dec - p)**2 + (x_dec - x_t)**2) |
|
104
|
|
|
return rep, distance |
|
105
|
|
|
|
|
106
|
|
|
def _unfold(self, p, n): |
|
107
|
|
|
if self.additional_h: |
|
108
|
|
|
n += 1 |
|
109
|
|
|
[ps, xs], _ = theano.scan(self._decode_step, outputs_info=[p, None], n_steps=n) |
|
110
|
|
|
if self.additional_h: |
|
111
|
|
|
return xs[::-1] |
|
112
|
|
|
else: |
|
113
|
|
|
return T.concatenate([xs, [ps[-1]]])[::-1] |
|
114
|
|
|
|
|
115
|
|
|
def _recursive_func(self, x): |
|
116
|
|
|
# Return total error |
|
117
|
|
|
if self.additional_h: |
|
118
|
|
|
h0 = self.h0 |
|
119
|
|
|
start_index = 0 |
|
120
|
|
|
else: |
|
121
|
|
|
h0 = x[0] |
|
122
|
|
|
start_index = 1 |
|
123
|
|
|
[reps, distances], _ = theano.scan(self._recursive_step, sequences=[T.arange(start_index, x.shape[0])], |
|
124
|
|
|
outputs_info=[h0, None], non_sequences=[x]) |
|
125
|
|
|
return reps[-1], T.sum(distances) |
|
126
|
|
|
|
|
127
|
|
|
def encode_func(self, x): |
|
128
|
|
|
if self.additional_h: |
|
129
|
|
|
h0 = self.h0 |
|
130
|
|
|
start_index = 0 |
|
131
|
|
|
else: |
|
132
|
|
|
h0 = x[0] |
|
133
|
|
|
start_index = 1 |
|
134
|
|
|
[reps, _], _ = theano.scan(self._recursive_step, sequences=[T.arange(start_index, x.shape[0])], |
|
135
|
|
|
outputs_info=[h0, None], non_sequences=[x]) |
|
136
|
|
|
return reps[-1] |
|
137
|
|
|
|
|
138
|
|
|
def _decode_step(self, p): |
|
139
|
|
|
p_dec = self._activation_func(T.dot(p, self.W_d1) + self.B_d1) |
|
140
|
|
|
x_dec = self._activation_func(T.dot(p, self.W_d2) + self.B_d2) |
|
141
|
|
|
return p_dec, x_dec |
|
142
|
|
|
|
|
143
|
|
|
def decode_func(self, rep, n): |
|
144
|
|
|
return self._unfold(rep, n) |
|
145
|
|
|
|
|
146
|
|
|
def _setup_functions(self): |
|
147
|
|
|
self._assistive_params = [] |
|
148
|
|
|
self._activation_func = build_activation(self.activation) |
|
149
|
|
|
self._softmax_func = build_activation('softmax') |
|
150
|
|
|
|
|
151
|
|
|
def _setup_params(self): |
|
152
|
|
|
if not self.rep_dim or self.rep_dim < 0: |
|
153
|
|
|
self.rep_dim = self.input_dim |
|
154
|
|
|
if not self.additional_h and self.rep_dim != self.input_dim: |
|
155
|
|
|
raise Exception("rep_dim must be input_dim when additional_h is not used") |
|
156
|
|
|
|
|
157
|
|
|
self.W_e1 = self.create_weight(self.rep_dim, self.rep_dim, "enc1") |
|
158
|
|
|
self.W_e2 = self.create_weight(self.input_dim, self.rep_dim, "enc2") |
|
159
|
|
|
self.B_e = self.create_bias(self.rep_dim, "enc") |
|
160
|
|
|
|
|
161
|
|
|
self.W_d1 = self.create_weight(self.rep_dim, self.rep_dim, "dec1") |
|
162
|
|
|
self.W_d2 = self.create_weight(self.rep_dim, self.input_dim, "dec2") |
|
163
|
|
|
self.B_d1 = self.create_bias(self.rep_dim, "dec1") |
|
164
|
|
|
self.B_d2 = self.create_bias(self.input_dim, "dec2") |
|
165
|
|
|
|
|
166
|
|
|
self.h0 = None |
|
167
|
|
|
if self.additional_h: |
|
168
|
|
|
self.h0 = self.create_vector(self.output_dim, "h0") |
|
169
|
|
|
|
|
170
|
|
|
self.register_parameters(self.W_e1, self.W_e2, self.W_d1, self.W_d2, |
|
171
|
|
|
self.B_e, self.B_d1, self.B_d2) |
|
172
|
|
|
|
|
173
|
|
|
|
uaca /
deepy
