/
masked.py
117 lines (103 loc) · 4.45 KB
/
masked.py
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
# coding=utf-8
# Copyright 2020 The TF-Agents Authors.
#
# 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
#
# https://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.
"""Define distributions for spaces where not all actions are valid."""
import tensorflow as tf # pylint: disable=g-explicit-tensorflow-version-import
import tensorflow_probability as tfp
class MaskedCategorical(tfp.distributions.Categorical):
"""A categorical distribution which supports masks per step.
Masked values are replaced with -inf inside the logits. This means the values
will never be sampled.
When computing the log probability of a set of actions, each action is
assigned a probability under each sample. _log_prob is modified to only return
the probability of a sample under the distribution for the same timestep.
TODO(ddohan): Integrate entropy calculation from cl/207017752
"""
def __init__(
self,
logits,
mask,
probs=None,
dtype=tf.int32,
validate_args=False,
allow_nan_stats=True,
neg_inf=-1e10,
name='MaskedCategorical',
):
"""Initialize Categorical distributions using class log-probabilities.
Args:
logits: An N-D `Tensor`, `N >= 1`, representing the log probabilities of a
set of Categorical distributions. The first `N - 1` dimensions index
into a batch of independent distributions and the last dimension
represents a vector of logits for each class. Only one of `logits` or
`probs` should be passed in.
mask: A boolean mask. False/0 values mean a position should be masked out.
probs: Must be `None`. Required to conform with base class
`tfp.distributions.Categorical`.
dtype: The type of the event samples (default: int32).
validate_args: Python `bool`, default `False`. When `True` distribution
parameters are checked for validity despite possibly degrading runtime
performance. When `False` invalid inputs may silently render incorrect
outputs.
allow_nan_stats: Python `bool`, default `True`. When `True`, statistics
(e.g., mean, mode, variance) use the value "`NaN`" to indicate the
result is undefined. When `False`, an exception is raised if one or more
of the statistic's batch members are undefined.
neg_inf: None or Float. Value used to mask out invalid positions. If None,
use logits.dtype.min to get a large negative number. Otherwise use given
value.
name: Python `str` name prefixed to Ops created by this class.
"""
parameters = dict(locals())
logits = tf.convert_to_tensor(value=logits)
mask = tf.convert_to_tensor(value=mask)
self._mask = tf.cast(mask, tf.bool) # Nonzero values are True
self._neg_inf = neg_inf
if probs is not None:
raise ValueError(
'Must provide masked predictions as logits.'
' Probs are accepted for API compatibility with '
' Categorical distribution. Given `%s`.' % probs
)
if neg_inf is None:
neg_inf = logits.dtype.min
neg_inf = tf.cast(
tf.fill(dims=tf.shape(input=logits), value=neg_inf), logits.dtype
)
logits = tf.compat.v2.where(self._mask, logits, neg_inf)
super(MaskedCategorical, self).__init__(
logits=logits,
probs=None,
dtype=dtype,
validate_args=validate_args,
allow_nan_stats=allow_nan_stats,
name=name,
)
self._parameters = parameters
def _entropy(self):
entropy = tf.nn.log_softmax(self.logits) * self.probs_parameter()
# Replace the (potentially -inf) values with 0s before summing.
entropy = tf.compat.v1.where(self._mask, entropy, tf.zeros_like(entropy))
return -tf.reduce_sum(input_tensor=entropy, axis=-1)
@classmethod
def _parameter_properties(cls, dtype, num_classes=None):
return dict(logits=tfp.util.ParameterProperties(event_ndims=1))
@property
def mask(self):
return self._mask
@property
def parameters(self):
params = super(MaskedCategorical, self).parameters
params['mask'] = self.mask
return params