1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
|
# Copyright 2016 The TensorFlow Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# ==============================================================================
"""A Conditional Transformed Distribution class."""
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
from tensorflow.contrib.distributions.python.ops import conditional_distribution
from tensorflow.python.framework import dtypes
from tensorflow.python.framework import ops
from tensorflow.python.framework import tensor_util
from tensorflow.python.ops import array_ops
from tensorflow.python.ops import math_ops
from tensorflow.python.ops.distributions import transformed_distribution
from tensorflow.python.ops.distributions import util as distribution_util
# pylint: disable=protected-access
_concat_vectors = transformed_distribution._concat_vectors
# pylint: enable=protected-access
__all__ = [
"ConditionalTransformedDistribution",
]
_condition_kwargs_dict = {
"bijector_kwargs": ("Python dictionary of arg names/values "
"forwarded to the bijector."),
"distribution_kwargs": ("Python dictionary of arg names/values "
"forwarded to the distribution."),
}
class ConditionalTransformedDistribution(
conditional_distribution.ConditionalDistribution,
transformed_distribution.TransformedDistribution):
"""A TransformedDistribution that allows intrinsic conditioning."""
@distribution_util.AppendDocstring(kwargs_dict=_condition_kwargs_dict)
def _sample_n(self, n, seed=None,
bijector_kwargs=None,
distribution_kwargs=None):
sample_shape = _concat_vectors(
distribution_util.pick_vector(self._needs_rotation, self._empty, [n]),
self._override_batch_shape,
self._override_event_shape,
distribution_util.pick_vector(self._needs_rotation, [n], self._empty))
distribution_kwargs = distribution_kwargs or {}
x = self.distribution.sample(sample_shape=sample_shape,
seed=seed,
**distribution_kwargs)
x = self._maybe_rotate_dims(x)
# We'll apply the bijector in the `_call_sample_n` function.
return x
def _call_sample_n(self, sample_shape, seed, name,
bijector_kwargs=None,
distribution_kwargs=None):
# We override `_call_sample_n` rather than `_sample_n` so we can ensure that
# the result of `self.bijector.forward` is not modified (and thus caching
# works).
with self._name_scope(name, values=[sample_shape]):
sample_shape = ops.convert_to_tensor(
sample_shape, dtype=dtypes.int32, name="sample_shape")
sample_shape, n = self._expand_sample_shape_to_vector(
sample_shape, "sample_shape")
# First, generate samples. We will possibly generate extra samples in the
# event that we need to reinterpret the samples as part of the
# event_shape.
x = self._sample_n(n, seed, bijector_kwargs, distribution_kwargs)
# Next, we reshape `x` into its final form. We do this prior to the call
# to the bijector to ensure that the bijector caching works.
batch_event_shape = array_ops.shape(x)[1:]
final_shape = array_ops.concat([sample_shape, batch_event_shape], 0)
x = array_ops.reshape(x, final_shape)
# Finally, we apply the bijector's forward transformation. For caching to
# work, it is imperative that this is the last modification to the
# returned result.
bijector_kwargs = bijector_kwargs or {}
y = self.bijector.forward(x, **bijector_kwargs)
y = self._set_sample_static_shape(y, sample_shape)
return y
@distribution_util.AppendDocstring(kwargs_dict=_condition_kwargs_dict)
def _log_prob(self, y, bijector_kwargs=None, distribution_kwargs=None):
# For caching to work, it is imperative that the bijector is the first to
# modify the input.
bijector_kwargs = bijector_kwargs or {}
distribution_kwargs = distribution_kwargs or {}
x = self.bijector.inverse(y, **bijector_kwargs)
event_ndims = self._maybe_get_event_ndims_statically()
ildj = self.bijector.inverse_log_det_jacobian(
y, event_ndims=event_ndims, **bijector_kwargs)
if self.bijector._is_injective: # pylint: disable=protected-access
return self._finish_log_prob_for_one_fiber(y, x, ildj,
distribution_kwargs)
lp_on_fibers = [
self._finish_log_prob_for_one_fiber(y, x_i, ildj_i, distribution_kwargs)
for x_i, ildj_i in zip(x, ildj)]
return math_ops.reduce_logsumexp(array_ops.stack(lp_on_fibers), axis=0)
def _finish_log_prob_for_one_fiber(self, y, x, ildj, distribution_kwargs):
"""Finish computation of log_prob on one element of the inverse image."""
x = self._maybe_rotate_dims(x, rotate_right=True)
log_prob = self.distribution.log_prob(x, **distribution_kwargs)
if self._is_maybe_event_override:
log_prob = math_ops.reduce_sum(log_prob, self._reduce_event_indices)
return math_ops.cast(ildj, log_prob.dtype) + log_prob
@distribution_util.AppendDocstring(kwargs_dict=_condition_kwargs_dict)
def _prob(self, y, bijector_kwargs=None, distribution_kwargs=None):
bijector_kwargs = bijector_kwargs or {}
distribution_kwargs = distribution_kwargs or {}
x = self.bijector.inverse(y, **bijector_kwargs)
event_ndims = self._maybe_get_event_ndims_statically()
ildj = self.bijector.inverse_log_det_jacobian(
y, event_ndims=event_ndims, **bijector_kwargs)
if self.bijector._is_injective: # pylint: disable=protected-access
return self._finish_prob_for_one_fiber(y, x, ildj, distribution_kwargs)
prob_on_fibers = [
self._finish_prob_for_one_fiber(y, x_i, ildj_i, distribution_kwargs)
for x_i, ildj_i in zip(x, ildj)]
return sum(prob_on_fibers)
def _finish_prob_for_one_fiber(self, y, x, ildj, distribution_kwargs):
"""Finish computation of prob on one element of the inverse image."""
x = self._maybe_rotate_dims(x, rotate_right=True)
prob = self.distribution.prob(x, **distribution_kwargs)
if self._is_maybe_event_override:
prob = math_ops.reduce_prod(prob, self._reduce_event_indices)
return math_ops.exp(math_ops.cast(ildj, prob.dtype)) * prob
@distribution_util.AppendDocstring(kwargs_dict=_condition_kwargs_dict)
def _log_cdf(self, y, bijector_kwargs=None, distribution_kwargs=None):
if self._is_maybe_event_override:
raise NotImplementedError("log_cdf is not implemented when overriding "
"event_shape")
if not self.bijector._is_injective: # pylint: disable=protected-access
raise NotImplementedError("log_cdf is not implemented when "
"bijector is not injective.")
bijector_kwargs = bijector_kwargs or {}
distribution_kwargs = distribution_kwargs or {}
x = self.bijector.inverse(y, **bijector_kwargs)
return self.distribution.log_cdf(x, **distribution_kwargs)
@distribution_util.AppendDocstring(kwargs_dict=_condition_kwargs_dict)
def _cdf(self, y, bijector_kwargs=None, distribution_kwargs=None):
if self._is_maybe_event_override:
raise NotImplementedError("cdf is not implemented when overriding "
"event_shape")
if not self.bijector._is_injective: # pylint: disable=protected-access
raise NotImplementedError("cdf is not implemented when "
"bijector is not injective.")
bijector_kwargs = bijector_kwargs or {}
distribution_kwargs = distribution_kwargs or {}
x = self.bijector.inverse(y, **bijector_kwargs)
return self.distribution.cdf(x, **distribution_kwargs)
@distribution_util.AppendDocstring(kwargs_dict=_condition_kwargs_dict)
def _log_survival_function(self, y,
bijector_kwargs=None, distribution_kwargs=None):
if self._is_maybe_event_override:
raise NotImplementedError("log_survival_function is not implemented when "
"overriding event_shape")
if not self.bijector._is_injective: # pylint: disable=protected-access
raise NotImplementedError("log_survival_function is not implemented when "
"bijector is not injective.")
bijector_kwargs = bijector_kwargs or {}
distribution_kwargs = distribution_kwargs or {}
x = self.bijector.inverse(y, **bijector_kwargs)
return self.distribution.log_survival_function(x, **distribution_kwargs)
@distribution_util.AppendDocstring(kwargs_dict=_condition_kwargs_dict)
def _survival_function(self, y,
bijector_kwargs=None, distribution_kwargs=None):
if self._is_maybe_event_override:
raise NotImplementedError("survival_function is not implemented when "
"overriding event_shape")
if not self.bijector._is_injective: # pylint: disable=protected-access
raise NotImplementedError("survival_function is not implemented when "
"bijector is not injective.")
bijector_kwargs = bijector_kwargs or {}
distribution_kwargs = distribution_kwargs or {}
x = self.bijector.inverse(y, **bijector_kwargs)
return self.distribution.survival_function(x, **distribution_kwargs)
@distribution_util.AppendDocstring(kwargs_dict=_condition_kwargs_dict)
def _quantile(self, value, bijector_kwargs=None, distribution_kwargs=None):
if self._is_maybe_event_override:
raise NotImplementedError("quantile is not implemented when overriding "
"event_shape")
if not self.bijector._is_injective: # pylint: disable=protected-access
raise NotImplementedError("quantile is not implemented when "
"bijector is not injective.")
bijector_kwargs = bijector_kwargs or {}
distribution_kwargs = distribution_kwargs or {}
# x_q is the "qth quantile" of X iff q = P[X <= x_q]. Now, since X =
# g^{-1}(Y), q = P[X <= x_q] = P[g^{-1}(Y) <= x_q] = P[Y <= g(x_q)],
# implies the qth quantile of Y is g(x_q).
inv_cdf = self.distribution.quantile(value, **distribution_kwargs)
return self.bijector.forward(inv_cdf, **bijector_kwargs)
def _maybe_get_event_ndims_statically(self):
if self.event_shape.ndims is not None:
return self.event_shape.ndims
event_ndims = array_ops.size(self.event_shape_tensor())
static_event_ndims = tensor_util.constant_value(event_ndims)
if static_event_ndims is not None:
return static_event_ndims
return event_ndims
|
