/
softplus.py
144 lines (121 loc) · 5.27 KB
/
softplus.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
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
# 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.
# ==============================================================================
"""Softplus bijector."""
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
from tensorflow.python.framework import ops
from tensorflow.python.ops import check_ops
from tensorflow.python.ops import control_flow_ops
from tensorflow.python.ops import math_ops
from tensorflow.python.ops import nn_ops
from tensorflow.python.ops.distributions import bijector
from tensorflow.python.ops.distributions import util as distribution_util
__all__ = [
"Softplus",
]
class Softplus(bijector.Bijector):
"""Bijector which computes `Y = g(X) = Log[1 + exp(X)]`.
The softplus `Bijector` has the following two useful properties:
* The domain is the positive real numbers
* `softplus(x) approx x`, for large `x`, so it does not overflow as easily as
the `Exp` `Bijector`.
The optional nonzero `hinge_softness` parameter changes the transition at
zero. With `hinge_softness = c`, the bijector is:
```f_c(x) := c * g(x / c) = c * Log[1 + exp(x / c)].```
For large `x >> 1`, `c * Log[1 + exp(x / c)] approx c * Log[exp(x / c)] = x`,
so the behavior for large `x` is the same as the standard softplus.
As `c > 0` approaches 0 from the right, `f_c(x)` becomes less and less soft,
approaching `max(0, x)`.
* `c = 1` is the default.
* `c > 0` but small means `f(x) approx ReLu(x) = max(0, x)`.
* `c < 0` flips sign and reflects around the `y-axis`: `f_{-c}(x) = -f_c(-x)`.
* `c = 0` results in a non-bijective transformation and triggers an exception.
Example Use:
```python
# Create the Y=g(X)=softplus(X) transform which works only on Tensors with 1
# batch ndim and 2 event ndims (i.e., vector of matrices).
softplus = Softplus(event_ndims=2)
x = [[[1., 2],
[3, 4]],
[[5, 6],
[7, 8]]]
log(1 + exp(x)) == softplus.forward(x)
log(exp(x) - 1) == softplus.inverse(x)
```
Note: log(.) and exp(.) are applied element-wise but the Jacobian is a
reduction over the event space.
"""
@distribution_util.AppendDocstring(
kwargs_dict={
"hinge_softness": (
"Nonzero floating point `Tensor`. Controls the softness of what "
"would otherwise be a kink at the origin. Default is 1.0")})
def __init__(self,
event_ndims=0,
hinge_softness=None,
validate_args=False,
name="softplus"):
with ops.name_scope(name, values=[hinge_softness]):
if hinge_softness is not None:
self._hinge_softness = ops.convert_to_tensor(
hinge_softness, name="hinge_softness")
else:
self._hinge_softness = None
if validate_args:
nonzero_check = check_ops.assert_none_equal(
ops.convert_to_tensor(
0, dtype=self.hinge_softness.dtype),
self.hinge_softness,
message="hinge_softness must be non-zero")
self._hinge_softness = control_flow_ops.with_dependencies(
[nonzero_check], self.hinge_softness)
super(Softplus, self).__init__(
event_ndims=event_ndims,
validate_args=validate_args,
name=name)
def _forward(self, x):
if self.hinge_softness is None:
return nn_ops.softplus(x)
hinge_softness = math_ops.cast(self.hinge_softness, x.dtype)
return hinge_softness * nn_ops.softplus(x / hinge_softness)
def _inverse(self, y):
if self.hinge_softness is None:
return distribution_util.softplus_inverse(y)
hinge_softness = math_ops.cast(self.hinge_softness, y.dtype)
return hinge_softness * distribution_util.softplus_inverse(
y / hinge_softness)
def _inverse_log_det_jacobian(self, y):
# Could also do:
# ildj = math_ops.reduce_sum(y - distribution_util.softplus_inverse(y),
# axis=event_dims)
# but the following is more numerically stable. Ie,
# Y = Log[1 + exp{X}] ==> X = Log[exp{Y} - 1]
# ==> dX/dY = exp{Y} / (exp{Y} - 1)
# = 1 / (1 - exp{-Y}),
# which is the most stable for large Y > 0. For small Y, we use
# 1 - exp{-Y} approx Y.
if self.hinge_softness is not None:
y /= math_ops.cast(self.hinge_softness, y.dtype)
return -math_ops.reduce_sum(math_ops.log(-math_ops.expm1(-y)),
axis=self._event_dims_tensor(y))
def _forward_log_det_jacobian(self, x):
if self.hinge_softness is not None:
x /= math_ops.cast(self.hinge_softness, x.dtype)
return -math_ops.reduce_sum(nn_ops.softplus(-x),
axis=self._event_dims_tensor(x))
@property
def hinge_softness(self):
return self._hinge_softness