Coverage for /pythoncovmergedfiles/medio/medio/usr/local/lib/python3.8/site-packages/tensorflow/python/ops/distributions/student_t.py: 48%
100 statements
« prev ^ index » next coverage.py v7.4.0, created at 2024-01-03 07:57 +0000
« prev ^ index » next coverage.py v7.4.0, created at 2024-01-03 07:57 +0000
1# Copyright 2016 The TensorFlow Authors. All Rights Reserved.
2#
3# Licensed under the Apache License, Version 2.0 (the "License");
4# you may not use this file except in compliance with the License.
5# You may obtain a copy of the License at
6#
7# http://www.apache.org/licenses/LICENSE-2.0
8#
9# Unless required by applicable law or agreed to in writing, software
10# distributed under the License is distributed on an "AS IS" BASIS,
11# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12# See the License for the specific language governing permissions and
13# limitations under the License.
14# ==============================================================================
15"""Student's t distribution class."""
17import numpy as np
19from tensorflow.python.framework import constant_op
20from tensorflow.python.framework import dtypes
21from tensorflow.python.framework import ops
22from tensorflow.python.framework import tensor_shape
23from tensorflow.python.ops import array_ops
24from tensorflow.python.ops import check_ops
25from tensorflow.python.ops import control_flow_ops
26from tensorflow.python.ops import math_ops
27from tensorflow.python.ops import nn
28from tensorflow.python.ops import random_ops
29from tensorflow.python.ops import special_math_ops
30from tensorflow.python.ops.distributions import distribution
31from tensorflow.python.ops.distributions import util as distribution_util
32from tensorflow.python.util import deprecation
33from tensorflow.python.util.tf_export import tf_export
36__all__ = [
37 "StudentT",
38 "StudentTWithAbsDfSoftplusScale",
39]
42@tf_export(v1=["distributions.StudentT"])
43class StudentT(distribution.Distribution):
44 """Student's t-distribution.
46 This distribution has parameters: degree of freedom `df`, location `loc`,
47 and `scale`.
49 #### Mathematical details
51 The probability density function (pdf) is,
53 ```none
54 pdf(x; df, mu, sigma) = (1 + y**2 / df)**(-0.5 (df + 1)) / Z
55 where,
56 y = (x - mu) / sigma
57 Z = abs(sigma) sqrt(df pi) Gamma(0.5 df) / Gamma(0.5 (df + 1))
58 ```
60 where:
61 * `loc = mu`,
62 * `scale = sigma`, and,
63 * `Z` is the normalization constant, and,
64 * `Gamma` is the [gamma function](
65 https://en.wikipedia.org/wiki/Gamma_function).
67 The StudentT distribution is a member of the [location-scale family](
68 https://en.wikipedia.org/wiki/Location-scale_family), i.e., it can be
69 constructed as,
71 ```none
72 X ~ StudentT(df, loc=0, scale=1)
73 Y = loc + scale * X
74 ```
76 Notice that `scale` has semantics more similar to standard deviation than
77 variance. However it is not actually the std. deviation; the Student's
78 t-distribution std. dev. is `scale sqrt(df / (df - 2))` when `df > 2`.
80 Samples of this distribution are reparameterized (pathwise differentiable).
81 The derivatives are computed using the approach described in
82 (Figurnov et al., 2018).
84 #### Examples
86 Examples of initialization of one or a batch of distributions.
88 ```python
89 import tensorflow_probability as tfp
90 tfd = tfp.distributions
92 # Define a single scalar Student t distribution.
93 single_dist = tfd.StudentT(df=3)
95 # Evaluate the pdf at 1, returning a scalar Tensor.
96 single_dist.prob(1.)
98 # Define a batch of two scalar valued Student t's.
99 # The first has degrees of freedom 2, mean 1, and scale 11.
100 # The second 3, 2 and 22.
101 multi_dist = tfd.StudentT(df=[2, 3], loc=[1, 2.], scale=[11, 22.])
103 # Evaluate the pdf of the first distribution on 0, and the second on 1.5,
104 # returning a length two tensor.
105 multi_dist.prob([0, 1.5])
107 # Get 3 samples, returning a 3 x 2 tensor.
108 multi_dist.sample(3)
109 ```
111 Arguments are broadcast when possible.
113 ```python
114 # Define a batch of two Student's t distributions.
115 # Both have df 2 and mean 1, but different scales.
116 dist = tfd.StudentT(df=2, loc=1, scale=[11, 22.])
118 # Evaluate the pdf of both distributions on the same point, 3.0,
119 # returning a length 2 tensor.
120 dist.prob(3.0)
121 ```
123 Compute the gradients of samples w.r.t. the parameters:
125 ```python
126 df = tf.constant(2.0)
127 loc = tf.constant(2.0)
128 scale = tf.constant(11.0)
129 dist = tfd.StudentT(df=df, loc=loc, scale=scale)
130 samples = dist.sample(5) # Shape [5]
131 loss = tf.reduce_mean(tf.square(samples)) # Arbitrary loss function
132 # Unbiased stochastic gradients of the loss function
133 grads = tf.gradients(loss, [df, loc, scale])
134 ```
136 References:
137 Implicit Reparameterization Gradients:
138 [Figurnov et al., 2018]
139 (http://papers.nips.cc/paper/7326-implicit-reparameterization-gradients)
140 ([pdf](http://papers.nips.cc/paper/7326-implicit-reparameterization-gradients.pdf))
141 """
143 @deprecation.deprecated(
144 "2019-01-01",
145 "The TensorFlow Distributions library has moved to "
146 "TensorFlow Probability "
147 "(https://github.com/tensorflow/probability). You "
148 "should update all references to use `tfp.distributions` "
149 "instead of `tf.distributions`.",
150 warn_once=True)
151 def __init__(self,
152 df,
153 loc,
154 scale,
155 validate_args=False,
156 allow_nan_stats=True,
157 name="StudentT"):
158 """Construct Student's t distributions.
160 The distributions have degree of freedom `df`, mean `loc`, and scale
161 `scale`.
163 The parameters `df`, `loc`, and `scale` must be shaped in a way that
164 supports broadcasting (e.g. `df + loc + scale` is a valid operation).
166 Args:
167 df: Floating-point `Tensor`. The degrees of freedom of the
168 distribution(s). `df` must contain only positive values.
169 loc: Floating-point `Tensor`. The mean(s) of the distribution(s).
170 scale: Floating-point `Tensor`. The scaling factor(s) for the
171 distribution(s). Note that `scale` is not technically the standard
172 deviation of this distribution but has semantics more similar to
173 standard deviation than variance.
174 validate_args: Python `bool`, default `False`. When `True` distribution
175 parameters are checked for validity despite possibly degrading runtime
176 performance. When `False` invalid inputs may silently render incorrect
177 outputs.
178 allow_nan_stats: Python `bool`, default `True`. When `True`,
179 statistics (e.g., mean, mode, variance) use the value "`NaN`" to
180 indicate the result is undefined. When `False`, an exception is raised
181 if one or more of the statistic's batch members are undefined.
182 name: Python `str` name prefixed to Ops created by this class.
184 Raises:
185 TypeError: if loc and scale are different dtypes.
186 """
187 parameters = dict(locals())
188 with ops.name_scope(name, values=[df, loc, scale]) as name:
189 with ops.control_dependencies([check_ops.assert_positive(df)]
190 if validate_args else []):
191 self._df = array_ops.identity(df, name="df")
192 self._loc = array_ops.identity(loc, name="loc")
193 self._scale = array_ops.identity(scale, name="scale")
194 check_ops.assert_same_float_dtype(
195 (self._df, self._loc, self._scale))
196 super(StudentT, self).__init__(
197 dtype=self._scale.dtype,
198 reparameterization_type=distribution.FULLY_REPARAMETERIZED,
199 validate_args=validate_args,
200 allow_nan_stats=allow_nan_stats,
201 parameters=parameters,
202 graph_parents=[self._df, self._loc, self._scale],
203 name=name)
205 @staticmethod
206 def _param_shapes(sample_shape):
207 return dict(
208 zip(("df", "loc", "scale"), (
209 [ops.convert_to_tensor(
210 sample_shape, dtype=dtypes.int32)] * 3)))
212 @property
213 def df(self):
214 """Degrees of freedom in these Student's t distribution(s)."""
215 return self._df
217 @property
218 def loc(self):
219 """Locations of these Student's t distribution(s)."""
220 return self._loc
222 @property
223 def scale(self):
224 """Scaling factors of these Student's t distribution(s)."""
225 return self._scale
227 def _batch_shape_tensor(self):
228 return array_ops.broadcast_dynamic_shape(
229 array_ops.shape(self.df),
230 array_ops.broadcast_dynamic_shape(
231 array_ops.shape(self.loc), array_ops.shape(self.scale)))
233 def _batch_shape(self):
234 return array_ops.broadcast_static_shape(
235 array_ops.broadcast_static_shape(self.df.get_shape(),
236 self.loc.get_shape()),
237 self.scale.get_shape())
239 def _event_shape_tensor(self):
240 return constant_op.constant([], dtype=math_ops.int32)
242 def _event_shape(self):
243 return tensor_shape.TensorShape([])
245 def _sample_n(self, n, seed=None):
246 # The sampling method comes from the fact that if:
247 # X ~ Normal(0, 1)
248 # Z ~ Chi2(df)
249 # Y = X / sqrt(Z / df)
250 # then:
251 # Y ~ StudentT(df).
252 shape = array_ops.concat([[n], self.batch_shape_tensor()], 0)
253 normal_sample = random_ops.random_normal(shape, dtype=self.dtype, seed=seed)
254 df = self.df * array_ops.ones(self.batch_shape_tensor(), dtype=self.dtype)
255 gamma_sample = random_ops.random_gamma(
256 [n],
257 0.5 * df,
258 beta=0.5,
259 dtype=self.dtype,
260 seed=distribution_util.gen_new_seed(seed, salt="student_t"))
261 samples = normal_sample * math_ops.rsqrt(gamma_sample / df)
262 return samples * self.scale + self.loc # Abs(scale) not wanted.
264 def _log_prob(self, x):
265 return self._log_unnormalized_prob(x) - self._log_normalization()
267 def _log_unnormalized_prob(self, x):
268 y = (x - self.loc) / self.scale # Abs(scale) superfluous.
269 return -0.5 * (self.df + 1.) * math_ops.log1p(y**2. / self.df)
271 def _log_normalization(self):
272 return (math_ops.log(math_ops.abs(self.scale)) +
273 0.5 * math_ops.log(self.df) +
274 0.5 * np.log(np.pi) +
275 math_ops.lgamma(0.5 * self.df) -
276 math_ops.lgamma(0.5 * (self.df + 1.)))
278 def _cdf(self, x):
279 # Take Abs(scale) to make subsequent where work correctly.
280 y = (x - self.loc) / math_ops.abs(self.scale)
281 x_t = self.df / (y**2. + self.df)
282 neg_cdf = 0.5 * math_ops.betainc(0.5 * self.df, 0.5, x_t)
283 return array_ops.where_v2(math_ops.less(y, 0.), neg_cdf, 1. - neg_cdf)
285 def _entropy(self):
286 v = array_ops.ones(self.batch_shape_tensor(),
287 dtype=self.dtype)[..., array_ops.newaxis]
288 u = v * self.df[..., array_ops.newaxis]
289 beta_arg = array_ops.concat([u, v], -1) / 2.
290 return (math_ops.log(math_ops.abs(self.scale)) +
291 0.5 * math_ops.log(self.df) +
292 special_math_ops.lbeta(beta_arg) +
293 0.5 * (self.df + 1.) *
294 (math_ops.digamma(0.5 * (self.df + 1.)) -
295 math_ops.digamma(0.5 * self.df)))
297 @distribution_util.AppendDocstring(
298 """The mean of Student's T equals `loc` if `df > 1`, otherwise it is
299 `NaN`. If `self.allow_nan_stats=True`, then an exception will be raised
300 rather than returning `NaN`.""")
301 def _mean(self):
302 mean = self.loc * array_ops.ones(self.batch_shape_tensor(),
303 dtype=self.dtype)
304 if self.allow_nan_stats:
305 nan = np.array(np.nan, dtype=self.dtype.as_numpy_dtype())
306 return array_ops.where_v2(
307 math_ops.greater(
308 self.df,
309 array_ops.ones(self.batch_shape_tensor(), dtype=self.dtype)),
310 mean, array_ops.fill(self.batch_shape_tensor(), nan, name="nan"))
311 else:
312 return control_flow_ops.with_dependencies(
313 [
314 check_ops.assert_less(
315 array_ops.ones([], dtype=self.dtype),
316 self.df,
317 message="mean not defined for components of df <= 1"),
318 ],
319 mean)
321 @distribution_util.AppendDocstring("""
322 The variance for Student's T equals
324 ```
325 df / (df - 2), when df > 2
326 infinity, when 1 < df <= 2
327 NaN, when df <= 1
328 ```
329 """)
330 def _variance(self):
331 # We need to put the tf.where inside the outer tf.where to ensure we never
332 # hit a NaN in the gradient.
333 denom = array_ops.where_v2(
334 math_ops.greater(self.df, 2.), self.df - 2.,
335 array_ops.ones_like(self.df))
336 # Abs(scale) superfluous.
337 var = (array_ops.ones(self.batch_shape_tensor(), dtype=self.dtype) *
338 math_ops.square(self.scale) * self.df / denom)
339 # When 1 < df <= 2, variance is infinite.
340 inf = np.array(np.inf, dtype=self.dtype.as_numpy_dtype())
341 result_where_defined = array_ops.where_v2(
342 self.df > array_ops.fill(self.batch_shape_tensor(), 2.), var,
343 array_ops.fill(self.batch_shape_tensor(), inf, name="inf"))
345 if self.allow_nan_stats:
346 nan = np.array(np.nan, dtype=self.dtype.as_numpy_dtype())
347 return array_ops.where_v2(
348 math_ops.greater(
349 self.df,
350 array_ops.ones(self.batch_shape_tensor(), dtype=self.dtype)),
351 result_where_defined,
352 array_ops.fill(self.batch_shape_tensor(), nan, name="nan"))
353 else:
354 return control_flow_ops.with_dependencies(
355 [
356 check_ops.assert_less(
357 array_ops.ones([], dtype=self.dtype),
358 self.df,
359 message="variance not defined for components of df <= 1"),
360 ],
361 result_where_defined)
363 def _mode(self):
364 return array_ops.identity(self.loc)
367class StudentTWithAbsDfSoftplusScale(StudentT):
368 """StudentT with `df = floor(abs(df))` and `scale = softplus(scale)`."""
370 @deprecation.deprecated(
371 "2019-01-01",
372 "Use `tfd.StudentT(tf.floor(tf.abs(df)), loc, "
373 "tf.nn.softplus(scale)) instead.",
374 warn_once=True)
375 def __init__(self,
376 df,
377 loc,
378 scale,
379 validate_args=False,
380 allow_nan_stats=True,
381 name="StudentTWithAbsDfSoftplusScale"):
382 parameters = dict(locals())
383 with ops.name_scope(name, values=[df, scale]) as name:
384 super(StudentTWithAbsDfSoftplusScale, self).__init__(
385 df=math_ops.floor(math_ops.abs(df)),
386 loc=loc,
387 scale=nn.softplus(scale, name="softplus_scale"),
388 validate_args=validate_args,
389 allow_nan_stats=allow_nan_stats,
390 name=name)
391 self._parameters = parameters