|
1
|
|
|
"""Bricks that compose together other bricks in linear sequences.""" |
|
2
|
|
|
import copy |
|
3
|
|
|
from toolz import interleave |
|
4
|
|
|
from picklable_itertools.extras import equizip |
|
5
|
|
|
|
|
6
|
|
|
from ..utils import pack |
|
7
|
|
|
from .base import Brick, application, lazy |
|
8
|
|
|
from .interfaces import Feedforward, Initializable |
|
9
|
|
|
from .simple import Linear |
|
10
|
|
|
|
|
11
|
|
|
|
|
12
|
|
|
class Sequence(Brick): |
|
13
|
|
|
"""A sequence of bricks. |
|
14
|
|
|
|
|
15
|
|
|
This brick applies a sequence of bricks, assuming that their in- and |
|
16
|
|
|
outputs are compatible. |
|
17
|
|
|
|
|
18
|
|
|
Parameters |
|
19
|
|
|
---------- |
|
20
|
|
|
application_methods : list |
|
21
|
|
|
List of :class:`.BoundApplication` to apply |
|
22
|
|
|
|
|
23
|
|
|
""" |
|
24
|
|
|
def __init__(self, application_methods, **kwargs): |
|
25
|
|
|
super(Sequence, self).__init__(**kwargs) |
|
26
|
|
|
self.application_methods = application_methods |
|
27
|
|
|
|
|
28
|
|
|
seen = set() |
|
29
|
|
|
self.children = [app.brick for app in application_methods |
|
30
|
|
|
if not (app.brick in seen or seen.add(app.brick))] |
|
31
|
|
|
|
|
32
|
|
|
@application |
|
33
|
|
|
def apply(self, *args): |
|
34
|
|
|
child_input = args |
|
35
|
|
|
for application_method in self.application_methods: |
|
36
|
|
|
output = application_method(*pack(child_input)) |
|
37
|
|
|
child_input = output |
|
38
|
|
|
return output |
|
39
|
|
|
|
|
40
|
|
|
@apply.property('inputs') |
|
41
|
|
|
def apply_inputs(self): |
|
42
|
|
|
return self.application_methods[0].inputs |
|
43
|
|
|
|
|
44
|
|
|
@apply.property('outputs') |
|
45
|
|
|
def apply_outputs(self): |
|
46
|
|
|
return self.application_methods[-1].outputs |
|
47
|
|
|
|
|
48
|
|
|
|
|
49
|
|
|
class FeedforwardSequence(Sequence, Feedforward): |
|
50
|
|
|
"""A sequence where the first and last bricks are feedforward. |
|
51
|
|
|
|
|
52
|
|
|
Parameters |
|
53
|
|
|
---------- |
|
54
|
|
|
application_methods : list |
|
55
|
|
|
List of :class:`.BoundApplication` to apply. The first and last |
|
56
|
|
|
application method should belong to a :class:`Feedforward` brick. |
|
57
|
|
|
|
|
58
|
|
|
""" |
|
59
|
|
|
@property |
|
60
|
|
|
def input_dim(self): |
|
61
|
|
|
return self.children[0].input_dim |
|
62
|
|
|
|
|
63
|
|
|
@input_dim.setter |
|
64
|
|
|
def input_dim(self, value): |
|
65
|
|
|
self.children[0].input_dim = value |
|
66
|
|
|
|
|
67
|
|
|
@property |
|
68
|
|
|
def output_dim(self): |
|
69
|
|
|
return self.children[-1].output_dim |
|
70
|
|
|
|
|
71
|
|
|
@output_dim.setter |
|
72
|
|
|
def output_dim(self, value): |
|
73
|
|
|
self.children[-1].output_dim = value |
|
74
|
|
|
|
|
75
|
|
|
|
|
76
|
|
|
class MLP(Sequence, Initializable, Feedforward): |
|
77
|
|
|
"""A simple multi-layer perceptron. |
|
78
|
|
|
|
|
79
|
|
|
Parameters |
|
80
|
|
|
---------- |
|
81
|
|
|
activations : list of :class:`.Brick`, :class:`.BoundApplication`, |
|
82
|
|
|
or ``None`` |
|
83
|
|
|
A list of activations to apply after each linear transformation. |
|
84
|
|
|
Give ``None`` to not apply any activation. It is assumed that the |
|
85
|
|
|
application method to use is ``apply``. Required for |
|
86
|
|
|
:meth:`__init__`. |
|
87
|
|
|
dims : list of ints |
|
88
|
|
|
A list of input dimensions, as well as the output dimension of the |
|
89
|
|
|
last layer. Required for :meth:`~.Brick.allocate`. |
|
90
|
|
|
prototype : :class:`.Brick`, optional |
|
91
|
|
|
The transformation prototype. A copy will be created for every |
|
92
|
|
|
activation. If not provided, an instance of :class:`~simple.Linear` |
|
93
|
|
|
will be used. |
|
94
|
|
|
|
|
95
|
|
|
Notes |
|
96
|
|
|
----- |
|
97
|
|
|
See :class:`Initializable` for initialization parameters. |
|
98
|
|
|
|
|
99
|
|
|
Note that the ``weights_init``, ``biases_init`` and ``use_bias`` |
|
100
|
|
|
configurations will overwrite those of the layers each time the |
|
101
|
|
|
:class:`MLP` is re-initialized. For more fine-grained control, push the |
|
102
|
|
|
configuration to the child layers manually before initialization. |
|
103
|
|
|
|
|
104
|
|
|
>>> from blocks.bricks import Tanh |
|
105
|
|
|
>>> from blocks.initialization import IsotropicGaussian, Constant |
|
106
|
|
|
>>> mlp = MLP(activations=[Tanh(), None], dims=[30, 20, 10], |
|
107
|
|
|
... weights_init=IsotropicGaussian(), |
|
108
|
|
|
... biases_init=Constant(1)) |
|
109
|
|
|
>>> mlp.push_initialization_config() # Configure children |
|
110
|
|
|
>>> mlp.children[0].weights_init = IsotropicGaussian(0.1) |
|
111
|
|
|
>>> mlp.initialize() |
|
112
|
|
|
|
|
113
|
|
|
""" |
|
114
|
|
|
@lazy(allocation=['dims']) |
|
115
|
|
|
def __init__(self, activations, dims, prototype=None, **kwargs): |
|
116
|
|
|
self.activations = activations |
|
117
|
|
|
self.prototype = Linear() if prototype is None else prototype |
|
118
|
|
|
self.linear_transformations = [] |
|
119
|
|
|
for i in range(len(activations)): |
|
120
|
|
|
linear = copy.deepcopy(self.prototype) |
|
121
|
|
|
name = self.prototype.__class__.__name__.lower() |
|
122
|
|
|
linear.name = '{}_{}'.format(name, i) |
|
123
|
|
|
self.linear_transformations.append(linear) |
|
124
|
|
|
# Interleave the transformations and activations |
|
125
|
|
|
application_methods = [] |
|
126
|
|
|
for entity in interleave([self.linear_transformations, activations]): |
|
127
|
|
|
if entity is None: |
|
128
|
|
|
continue |
|
129
|
|
|
if isinstance(entity, Brick): |
|
130
|
|
|
application_methods.append(entity.apply) |
|
131
|
|
|
else: |
|
132
|
|
|
application_methods.append(entity) |
|
133
|
|
|
if not dims: |
|
134
|
|
|
dims = [None] * (len(activations) + 1) |
|
135
|
|
|
self.dims = dims |
|
136
|
|
|
super(MLP, self).__init__(application_methods, **kwargs) |
|
137
|
|
|
|
|
138
|
|
|
@property |
|
139
|
|
|
def input_dim(self): |
|
140
|
|
|
return self.dims[0] |
|
141
|
|
|
|
|
142
|
|
|
@input_dim.setter |
|
143
|
|
|
def input_dim(self, value): |
|
144
|
|
|
self.dims[0] = value |
|
145
|
|
|
|
|
146
|
|
|
@property |
|
147
|
|
|
def output_dim(self): |
|
148
|
|
|
return self.dims[-1] |
|
149
|
|
|
|
|
150
|
|
|
@output_dim.setter |
|
151
|
|
|
def output_dim(self, value): |
|
152
|
|
|
self.dims[-1] = value |
|
153
|
|
|
|
|
154
|
|
|
def _push_allocation_config(self): |
|
155
|
|
|
if not len(self.dims) - 1 == len(self.linear_transformations): |
|
156
|
|
|
raise ValueError |
|
157
|
|
|
for input_dim, output_dim, layer in \ |
|
158
|
|
|
equizip(self.dims[:-1], self.dims[1:], |
|
159
|
|
|
self.linear_transformations): |
|
160
|
|
|
layer.input_dim = input_dim |
|
161
|
|
|
layer.output_dim = output_dim |
|
162
|
|
|
layer.use_bias = self.use_bias |
|
163
|
|
|
|
mila-udem /
blocks
