|
1
|
|
|
#!/usr/bin/env python |
|
2
|
|
|
# -*- coding: utf-8 -*- |
|
3
|
|
|
|
|
4
|
|
|
import numpy as np |
|
5
|
|
|
import theano.tensor as T |
|
6
|
|
|
from deepy.layers import NeuralLayer |
|
7
|
|
|
from deepy.layers.var import NeuralVar |
|
8
|
|
|
from deepy.utils import onehot_tensor, onehot |
|
9
|
|
|
from deepy.utils import FLOATX |
|
10
|
|
|
|
|
11
|
|
|
class OneHotEmbedding(NeuralLayer): |
|
12
|
|
|
""" |
|
13
|
|
|
One-hot embedding layer. |
|
14
|
|
|
Computation: [0,1,2] ---> [[1,0,0],[0,1,0],[0,0,1]] |
|
15
|
|
|
""" |
|
16
|
|
|
def __init__(self, vocab_size, cached=True, zero_index=None, mask=None): |
|
17
|
|
|
super(OneHotEmbedding, self).__init__("onehot") |
|
18
|
|
|
self.vocab_size = vocab_size |
|
19
|
|
|
self.output_dim = vocab_size |
|
20
|
|
|
self.cached = cached |
|
21
|
|
|
self.zero_index = zero_index |
|
22
|
|
|
self.mask = mask.tensor if type(mask) == NeuralVar else mask |
|
23
|
|
|
|
|
24
|
|
|
def prepare(self): |
|
25
|
|
|
if not self.cached: |
|
26
|
|
|
return |
|
27
|
|
|
onehot_matrix = [] |
|
28
|
|
|
for i in xrange(self.vocab_size): |
|
29
|
|
|
onehot_matrix.append(onehot(self.vocab_size, i)) |
|
30
|
|
|
onehot_matrix = np.array(onehot_matrix, dtype=FLOATX) |
|
31
|
|
|
self.onehot_list = self.create_matrix(self.vocab_size, self.vocab_size, "onehot_list") |
|
32
|
|
|
self.onehot_list.set_value(onehot_matrix) |
|
33
|
|
|
|
|
34
|
|
|
def output(self, x): |
|
35
|
|
|
if self.cached: |
|
36
|
|
|
if x.ndim == 1: |
|
37
|
|
|
ret_tensor = self.onehot_list[x] |
|
38
|
|
|
else: |
|
39
|
|
|
ret_tensor = self.onehot_list[x.flatten()].reshape((x.shape[0], x.shape[1], self.vocab_size)) |
|
40
|
|
|
else: |
|
41
|
|
|
ret_tensor = onehot_tensor(x, self.vocab_size) |
|
42
|
|
|
if self.zero_index != None: |
|
43
|
|
|
mask = T.neq(x, self.zero_index) |
|
44
|
|
|
if x.ndim == 1: |
|
45
|
|
|
ret_tensor *= mask[:, None] |
|
46
|
|
|
else: |
|
47
|
|
|
ret_tensor *= mask[:, :, None] |
|
48
|
|
|
if self.mask: |
|
49
|
|
|
if x.ndim == 1: |
|
50
|
|
|
ret_tensor *= self.mask[:, None] |
|
51
|
|
|
else: |
|
52
|
|
|
ret_tensor *= self.mask[:, :, None] |
|
53
|
|
|
return ret_tensor |
|
54
|
|
|
|
uaca /
deepy
