|
1
|
|
|
#!/usr/bin/env python |
|
2
|
|
|
# -*- coding: utf-8 -*- |
|
3
|
|
|
|
|
4
|
|
|
import theano.tensor as TT |
|
5
|
|
|
from theano.ifelse import ifelse as theano_ifelse |
|
6
|
|
|
|
|
7
|
|
|
from deepy.core.tensor_conversion import neural_computation |
|
8
|
|
|
from deepy.layers.concat import Concatenate |
|
9
|
|
|
from wrapper import deepy_tensor |
|
10
|
|
|
|
|
11
|
|
|
|
|
12
|
|
|
def concat(vars, axis=-1): |
|
13
|
|
|
""" |
|
14
|
|
|
A shortcut for concatenation. |
|
15
|
|
|
""" |
|
16
|
|
|
return concatenate(vars, axis) |
|
17
|
|
|
|
|
18
|
|
|
@neural_computation |
|
19
|
|
|
def reverse(tensor, axis=-1): |
|
20
|
|
|
ndim = tensor.ndim |
|
21
|
|
|
selectors = [slice(None)] * ndim |
|
22
|
|
|
selectors[axis] = slice(None, None, -1) |
|
23
|
|
|
ret = tensor[selectors] |
|
24
|
|
|
if hasattr(tensor.tag, "last_dim"): |
|
25
|
|
|
ret.tag.last_dim = tensor.tag.last_dim |
|
26
|
|
|
return ret |
|
27
|
|
|
|
|
28
|
|
|
@neural_computation |
|
29
|
|
|
def activate(var, method): |
|
30
|
|
|
""" |
|
31
|
|
|
An activation function. |
|
32
|
|
|
:param var: input var |
|
33
|
|
|
:param method: type of activation, such as `relu`,`tanh`,`sigmoid` |
|
34
|
|
|
""" |
|
35
|
|
|
from activations import get_activation |
|
36
|
|
|
return get_activation(method)(var) |
|
37
|
|
|
|
|
38
|
|
|
def concatenate(vars, axis=-1): |
|
39
|
|
|
""" |
|
40
|
|
|
A utility function of concatenate. |
|
41
|
|
|
""" |
|
42
|
|
|
from deepy.core.neural_var import NeuralVariable |
|
43
|
|
|
if isinstance(vars[0], NeuralVariable): |
|
44
|
|
|
concat_var = Concatenate(axis=axis).compute(*vars) |
|
45
|
|
|
if axis == -1 or axis == vars[0].tensor.ndim - 1: |
|
46
|
|
|
concat_var.output_dim = sum([x.output_dim for x in vars], 0) |
|
47
|
|
|
else: |
|
48
|
|
|
concat_var = TT.concatenate(vars, axis) |
|
49
|
|
|
return concat_var |
|
50
|
|
|
|
|
51
|
|
|
@neural_computation |
|
52
|
|
|
def ifelse(condition, then_branch, else_branch): |
|
53
|
|
|
return theano_ifelse(condition, then_branch, else_branch) |
|
54
|
|
|
|
|
55
|
|
|
|
|
56
|
|
|
def apply(func, *args, **kwargs): |
|
57
|
|
|
from deepy.core.neural_var import NeuralVariable |
|
58
|
|
|
dim = kwargs['dim'] if 'dim' in kwargs else args[0].dim() |
|
59
|
|
|
return NeuralVariable(func(*[x.tensor for x in args]), dim) |
|
60
|
|
|
|
|
61
|
|
|
def repeat(*args, **kwargs): |
|
62
|
|
|
return deepy_tensor.repeat(*args, **kwargs) |
|
63
|
|
|
|
|
64
|
|
|
def vars(*tensor_types): |
|
65
|
|
|
""" |
|
66
|
|
|
Create multiple variables without specifying last dimension and shape. |
|
67
|
|
|
:rtype: list of deepy.core.neural_var.NeuralVariable |
|
68
|
|
|
""" |
|
69
|
|
|
return map(var, tensor_types) |
|
70
|
|
|
|
|
71
|
|
|
|
|
72
|
|
|
def var(tensor_type, last_dim=0, test_shape=None): |
|
73
|
|
|
""" |
|
74
|
|
|
Wrap a Theano tensor into the variable for defining neural network. |
|
75
|
|
|
:param last_dim: last dimension of tensor, 0 indicates that the last dimension is flexible |
|
76
|
|
|
:rtype: deepy.core.neural_var.NeuralVariable |
|
77
|
|
|
""" |
|
78
|
|
|
# Create tensor |
|
79
|
|
|
from deepy.core.neural_var import NeuralVariable |
|
80
|
|
|
from deepy.core.env import env |
|
81
|
|
|
from theano.tensor.var import TensorVariable |
|
82
|
|
|
if isinstance(tensor_type, NeuralVariable): |
|
83
|
|
|
var = tensor_type |
|
84
|
|
|
if last_dim != 0: |
|
85
|
|
|
var.output_dim = last_dim |
|
86
|
|
|
elif isinstance(tensor_type, TensorVariable): |
|
87
|
|
|
var = NeuralVariable(tensor_type, dim=last_dim) |
|
88
|
|
|
elif isinstance(tensor_type, str): |
|
89
|
|
|
theano_tensor = getattr(TT, tensor_type)() |
|
90
|
|
|
var = NeuralVariable(theano_tensor, dim=last_dim) |
|
91
|
|
|
else: |
|
92
|
|
|
raise Exception("tensor_type shall be a string or a NeuralVariable") |
|
93
|
|
|
# Set test value |
|
94
|
|
|
if test_shape: |
|
95
|
|
|
if type(test_shape) != list and type(test_shape) != tuple: |
|
96
|
|
|
# May be it's a value |
|
97
|
|
|
var.set_test_value(test_shape) |
|
98
|
|
|
else: |
|
99
|
|
|
test_val = env.numpy_rand.rand(*test_shape) |
|
100
|
|
|
if len(test_shape) > 0: |
|
101
|
|
|
test_val = test_val.astype(var.tensor.dtype) |
|
102
|
|
|
elif var.tensor.dtype.startswith("int"): |
|
103
|
|
|
test_val = 1 |
|
104
|
|
|
var.set_test_value(test_val) |
|
105
|
|
|
else: |
|
106
|
|
|
# Create a general test_shape |
|
107
|
|
|
dims = [(d + 1) * 3 for d in range(var.tensor.ndim)] |
|
108
|
|
|
if var.dim() != 0: |
|
109
|
|
|
dims[-1] = var.dim() |
|
110
|
|
|
test_val = env.numpy_rand.rand(*dims) |
|
111
|
|
|
if len(dims) > 0: |
|
112
|
|
|
test_val = test_val.astype(var.tensor.dtype) |
|
113
|
|
|
elif var.tensor.dtype.startswith("int"): |
|
114
|
|
|
test_val = 1 |
|
115
|
|
|
var.set_test_value(test_val) |
|
116
|
|
|
return var |
|
117
|
|
|
|
|
118
|
|
|
|
|
119
|
|
|
def is_neural_var(var): |
|
120
|
|
|
from deepy.core.neural_var import NeuralVariable |
|
121
|
|
|
return isinstance(var, NeuralVariable) |
|
122
|
|
|
|
|
123
|
|
|
def is_theano_var(var): |
|
124
|
|
|
from theano.tensor.var import TensorVariable |
|
125
|
|
|
return isinstance(var, TensorVariable) |
zomux /
deepy
