|
1
|
|
|
"""Model - heavily annotated computation graph. |
|
2
|
|
|
|
|
3
|
|
|
A model in Blocks is simply an annotated computation graph. The class |
|
4
|
|
|
:class:`Model` extends :class:`blocks.graph.ComputationGraph` :class:, |
|
5
|
|
|
which is able to handle annotations and roles in general, but is |
|
6
|
|
|
deliberately made unaware of specific annotations that a Theano graph |
|
7
|
|
|
created by Blocks typically has, such as bricks and application calls. The |
|
8
|
|
|
:class:`Model` adds this functionality. Using :class:`Model` you can do |
|
9
|
|
|
things like query all the bricks used to build the computation graph, |
|
10
|
|
|
request "hierarchical names" of the parameters (a hierarchical name is a |
|
11
|
|
|
path-like string which in addition to the parameter's name contains names |
|
12
|
|
|
of the bricks on the path from a root brick to the brick that owns the |
|
13
|
|
|
parameters, e.g. ``/mlp/linear/W``). |
|
14
|
|
|
|
|
15
|
|
|
For more information, see :class:`Model` docstring. |
|
16
|
|
|
|
|
17
|
|
|
""" |
|
18
|
|
|
import logging |
|
19
|
|
|
from collections import OrderedDict, Counter |
|
20
|
|
|
from itertools import chain |
|
21
|
|
|
|
|
22
|
|
|
from blocks.graph import ComputationGraph |
|
23
|
|
|
from blocks.select import Selector |
|
24
|
|
|
from blocks.filter import get_brick |
|
25
|
|
|
|
|
26
|
|
|
logger = logging.getLogger(__name__) |
|
27
|
|
|
|
|
28
|
|
|
|
|
29
|
|
|
class Model(ComputationGraph): |
|
30
|
|
|
"""Handles annotations in Blocks-built computation graphs. |
|
31
|
|
|
|
|
32
|
|
|
Use this class to handle your Blocks-created computation graph. |
|
33
|
|
|
|
|
34
|
|
|
Examples |
|
35
|
|
|
-------- |
|
36
|
|
|
>>> from theano import tensor |
|
37
|
|
|
>>> from blocks.bricks import MLP, Tanh |
|
38
|
|
|
>>> x = tensor.matrix('x') |
|
39
|
|
|
>>> mlp = MLP([Tanh(), Tanh()], [10, 10, 10]) |
|
40
|
|
|
>>> y = mlp.apply(x) |
|
41
|
|
|
>>> model = Model(y) |
|
42
|
|
|
|
|
43
|
|
|
With :class:`Model` you can get access to the brick hierarchy. The |
|
44
|
|
|
brick hierarchy is defined by ``children`` attributes that every brick |
|
45
|
|
|
has. The bricks that are not children of other bricks are called top |
|
46
|
|
|
bricks. It is often useful to have access to top bricks of a brick |
|
47
|
|
|
hierarchy used to build a computation graph, and here is how you can do |
|
48
|
|
|
it: |
|
49
|
|
|
|
|
50
|
|
|
>>> model.get_top_bricks() #doctest: +ELLIPSIS |
|
51
|
|
|
[<blocks.bricks.sequences.MLP object at ...] |
|
52
|
|
|
|
|
53
|
|
|
You can also get "hierarchical" names for the parameters, |
|
54
|
|
|
which encode the position of the owning brick in the |
|
55
|
|
|
brick hierarchy. |
|
56
|
|
|
|
|
57
|
|
|
>>> model.get_parameter_dict() #doctest: +NORMALIZE_WHITESPACE |
|
58
|
|
|
OrderedDict([('/mlp/linear_1.b', b), ('/mlp/linear_0.b', b), |
|
59
|
|
|
('/mlp/linear_0.W', W), ('/mlp/linear_1.W', W)]) |
|
60
|
|
|
|
|
61
|
|
|
""" |
|
62
|
|
|
def __init__(self, *args, **kwargs): |
|
63
|
|
|
super(Model, self).__init__(*args, **kwargs) |
|
64
|
|
|
bricks = [get_brick(var) for var |
|
65
|
|
|
in self.variables + self.scan_variables if get_brick(var)] |
|
66
|
|
|
children = set(chain(*(brick.children for brick in bricks))) |
|
67
|
|
|
# Quadratic complexity: we should not have thousands of |
|
68
|
|
|
# top-level bricks. |
|
69
|
|
|
self.top_bricks = [] |
|
70
|
|
|
for brick in bricks: |
|
71
|
|
|
if brick not in children and brick not in self.top_bricks: |
|
72
|
|
|
self.top_bricks.append(brick) |
|
73
|
|
|
names = Counter([brick.name for brick in self.top_bricks]) |
|
74
|
|
|
repeated_names = [name for name, count in names.items() if count > 1] |
|
75
|
|
|
if repeated_names: |
|
76
|
|
|
raise ValueError("top bricks with the same name:" |
|
77
|
|
|
" {}".format(', '.join(repeated_names))) |
|
78
|
|
|
brick_parameter_names = { |
|
79
|
|
|
v: k for k, v in Selector( |
|
80
|
|
|
self.top_bricks).get_parameters().items()} |
|
81
|
|
|
parameter_list = [] |
|
82
|
|
|
for parameter in self.parameters: |
|
83
|
|
|
if parameter in brick_parameter_names: |
|
84
|
|
|
parameter_list.append((brick_parameter_names[parameter], |
|
85
|
|
|
parameter)) |
|
86
|
|
|
else: |
|
87
|
|
|
parameter_list.append((parameter.name, parameter)) |
|
88
|
|
|
self._parameter_dict = OrderedDict(parameter_list) |
|
89
|
|
|
|
|
90
|
|
|
def get_parameter_dict(self): |
|
91
|
|
|
"""Returns parameters with their hierarchical names. |
|
92
|
|
|
|
|
93
|
|
|
The parameter names are formed from positions of their owner bricks |
|
94
|
|
|
in the bricks hierarchy. The variable names are used for the |
|
95
|
|
|
parameters that do not belong to any brick. |
|
96
|
|
|
|
|
97
|
|
|
Returns |
|
98
|
|
|
------- |
|
99
|
|
|
parameter_dict : dict |
|
100
|
|
|
A dictionary of (hierarchical name, shared variable) pairs. |
|
101
|
|
|
|
|
102
|
|
|
""" |
|
103
|
|
|
return self._parameter_dict |
|
104
|
|
|
|
|
105
|
|
|
def get_parameter_values(self): |
|
106
|
|
|
"""Return the values of model parameters. |
|
107
|
|
|
|
|
108
|
|
|
The same hierarhical names as in :meth:`get_parameter_dict` are |
|
109
|
|
|
used to uniquely identify parameters. |
|
110
|
|
|
|
|
111
|
|
|
Returns |
|
112
|
|
|
------- |
|
113
|
|
|
parameter_values : OrderedDict |
|
114
|
|
|
Dictionary of (hierarchical name, :class:`~numpy.ndarray`) |
|
115
|
|
|
pairs. |
|
116
|
|
|
|
|
117
|
|
|
""" |
|
118
|
|
|
return OrderedDict( |
|
119
|
|
|
(name, parameter.get_value()) |
|
120
|
|
|
for name, parameter in self.get_parameter_dict().items()) |
|
121
|
|
|
|
|
122
|
|
|
def set_parameter_values(self, parameter_values): |
|
123
|
|
|
"""Set the values of model parameters. |
|
124
|
|
|
|
|
125
|
|
|
The same hierarhical names as in :meth:`get_parameter_dict` are |
|
126
|
|
|
used to uniquely identify parameters. |
|
127
|
|
|
|
|
128
|
|
|
Parameters |
|
129
|
|
|
---------- |
|
130
|
|
|
parameter_values : OrderedDict |
|
131
|
|
|
Dictionary of (hierarchical name, :class:`~numpy.ndarray`) |
|
132
|
|
|
pairs. |
|
133
|
|
|
|
|
134
|
|
|
""" |
|
135
|
|
|
parameters = self.get_parameter_dict() |
|
136
|
|
|
|
|
137
|
|
|
unknown = set(parameter_values) - set(parameters) |
|
138
|
|
|
missing = set(parameters) - set(parameter_values) |
|
139
|
|
|
if len(unknown): |
|
140
|
|
|
logger.error("unknown parameter names: {}\n".format(unknown)) |
|
141
|
|
|
if len(missing): |
|
142
|
|
|
logger.error("missing values for parameters: {}\n".format(missing)) |
|
143
|
|
|
|
|
144
|
|
|
for name, value in parameter_values.items(): |
|
145
|
|
|
if name in parameters: |
|
146
|
|
|
model_shape = parameters[name].container.data.shape |
|
147
|
|
|
if model_shape != value.shape: |
|
148
|
|
|
raise ValueError("Shape mismatch for parameter: {}. " |
|
149
|
|
|
"Expected {}, got {}." |
|
150
|
|
|
.format(name, model_shape, value.shape)) |
|
151
|
|
|
parameters[name].set_value(value) |
|
152
|
|
|
|
|
153
|
|
|
def get_top_bricks(self): |
|
154
|
|
|
"""Get the bricks that do not have parents. |
|
155
|
|
|
|
|
156
|
|
|
Returns |
|
157
|
|
|
------- |
|
158
|
|
|
bricks : list of :class:`~blocks.bricks.base.Brick` |
|
159
|
|
|
|
|
160
|
|
|
""" |
|
161
|
|
|
return self.top_bricks |
|
162
|
|
|
|
mila-udem /
blocks
