if sample_shape_static_val is None:
if ndims is None or not sample_shape.get_shape().is_fully_defined(
):
ndims = array_ops.rank(sample_shape)
expanded_shape = distribution_util.pick_vector(
math_ops.equal(ndims, 0),
np.array((1, ), dtype=dtypes.int32.as_numpy_dtype()),
array_ops.shape(sample_shape))
sample_shape = array_ops.reshape(sample_shape, expanded_shape)
total = math_ops.reduce_prod(sample_shape) # reduce_prod([]) == 1
else:
if ndims is None:
raise ValueError(
"Shouldn't be here; ndims cannot be none when we have a "
"tf.constant shape.")
if ndims == 0:
sample_shape_static_val = np.reshape(sample_shape_static_val,
[1])
sample_shape = ops.convert_to_tensor(sample_shape_static_val,
dtype=dtypes.int32,
name="sample_shape")
total = np.prod(sample_shape_static_val,
dtype=dtypes.int32.as_numpy_dtype())
return sample_shape, total
distribution_util.append_class_fun_doc(BaseDistribution.sample_n,
doc_str=Distribution.sample_n.__doc__)
distribution_util.append_class_fun_doc(BaseDistribution.log_prob,
doc_str=Distribution.log_prob.__doc__)
ndims = sample_shape.get_shape().ndims
if sample_shape_static_val is None:
if ndims is None or not sample_shape.get_shape().is_fully_defined():
ndims = array_ops.rank(sample_shape)
expanded_shape = distribution_util.pick_vector(
math_ops.equal(ndims, 0),
np.array((1,), dtype=dtypes.int32.as_numpy_dtype()),
array_ops.shape(sample_shape))
sample_shape = array_ops.reshape(sample_shape, expanded_shape)
total = math_ops.reduce_prod(sample_shape) # reduce_prod([]) == 1
else:
if ndims is None:
raise ValueError(
"Shouldn't be here; ndims cannot be none when we have a "
"tf.constant shape.")
if ndims == 0:
sample_shape_static_val = np.reshape(sample_shape_static_val, [1])
sample_shape = ops.convert_to_tensor(
sample_shape_static_val,
dtype=dtypes.int32,
name="sample_shape")
total = np.prod(sample_shape_static_val,
dtype=dtypes.int32.as_numpy_dtype())
return sample_shape, total
distribution_util.append_class_fun_doc(BaseDistribution.sample_n,
doc_str=Distribution.sample_n.__doc__)
distribution_util.append_class_fun_doc(BaseDistribution.log_prob,
doc_str=Distribution.log_prob.__doc__)
message="mode not defined for components of alpha <= 1")
], mode)
def _assert_valid_sample(self, x):
if not self.validate_args: return x
return control_flow_ops.with_dependencies([
check_ops.assert_positive(x),
distribution_util.assert_close(
array_ops.ones((), dtype=self.dtype),
math_ops.reduce_sum(x, reduction_indices=[-1])),
], x)
_prob_note = """
Note that the input must be a non-negative tensor with dtype `dtype` and whose
shape can be broadcast with `self.alpha`. For fixed leading dimensions, the
last dimension represents counts for the corresponding Dirichlet distribution
in `self.alpha`. `x` is only legal if it sums up to one.
"""
distribution_util.append_class_fun_doc(Dirichlet.log_prob, doc_str=_prob_note)
distribution_util.append_class_fun_doc(Dirichlet.prob, doc_str=_prob_note)
distribution_util.append_class_fun_doc(Dirichlet.mode, doc_str="""
Note that the mode for the Dirichlet distribution is only defined
when `alpha > 1`. This returns the mode when `alpha > 1`,
and NaN otherwise. If `self.allow_nan_stats` is `False`, an exception
will be raised rather than returning `NaN`.
""")
return control_flow_ops.with_dependencies([
check_ops.assert_non_negative(
counts, message="counts has negative components."),
check_ops.assert_less_equal(
counts, self._n, message="counts are not less than or equal to n."),
distribution_util.assert_integer_form(
counts, message="counts have non-integer components.")], counts)
_prob_note = """
For each batch member of counts `k`, `P[counts]` is the probability that
after sampling `n` draws from this Binomial distribution, the number of
successes is `k`. Note that different sequences of draws can result in the
same counts, thus the probability includes a combinatorial coefficient.
counts: Non-negative tensor with dtype `dtype` and whose shape can be
broadcast with `self.p` and `self.n`. `counts` is only legal if it is
less than or equal to `n` and its components are equal to integer
values.
"""
distribution_util.append_class_fun_doc(Binomial.log_prob, doc_str=_prob_note)
distribution_util.append_class_fun_doc(Binomial.prob, doc_str=_prob_note)
distribution_util.append_class_fun_doc(Binomial.mode, doc_str="""
Note that when `(n + 1) * p` is an integer, there are actually two modes.
Namely, `(n + 1) * p` and `(n + 1) * p - 1` are both modes. Here we return
only the larger of the two modes.
""")
accurate for large values of `y`, and in this case the `survival_function` must
also be defined on `y - 1`.
"""
)
_log_prob_note = (
_prob_base_note
+ """
The base distribution's `log_cdf` method must be defined on `y - 1`. If the
base distribution has a `log_survival_function` method results will be more
accurate for large values of `y`, and in this case the `log_survival_function`
must also be defined on `y - 1`.
"""
)
distribution_util.append_class_fun_doc(QuantizedDistribution.prob, doc_str=_prob_note)
distribution_util.append_class_fun_doc(QuantizedDistribution.log_prob, doc_str=_log_prob_note)
_cdf_base_note = """
For whole numbers `y`,
```
cdf(y) := P[Y <= y]
= 1, if y >= upper_cutoff,
= 0, if y < lower_cutoff,
= P[X <= y], otherwise.
```
Since `Y` only has mass at whole numbers, `P[Y <= y] = P[Y <= floor(y)]`.
nan = np.array(np.nan, dtype=self.dtype.as_numpy_dtype())
return math_ops.select(
self.alpha >= 1.,
mode,
array_ops.fill(self.batch_shape(), nan, name="nan"))
else:
return control_flow_ops.with_dependencies([
check_ops.assert_less(
array_ops.ones((), self.dtype),
self.alpha,
message="mode not defined for components of alpha <= 1"),
], mode)
distribution_util.append_class_fun_doc(Gamma.sample_n, doc_str="""
See the documentation for tf.random_gamma for more details.
""")
distribution_util.append_class_fun_doc(Gamma.entropy, doc_str="""
This is defined to be
```
entropy = alpha - log(beta) + log(Gamma(alpha))
+ (1-alpha)digamma(alpha)
```
where digamma(alpha) is the digamma function.
""")
distribution_util.append_class_fun_doc(Gamma.mode, doc_str="""
check_ops.assert_positive(
x,
message="Negative events lie outside Beta distribution support."
),
check_ops.assert_less(
x,
array_ops.ones((), self.dtype),
message="Event>=1 lies outside Beta distribution support."),
], x)
_prob_note = """
Note that the argument `x` must be a non-negative floating point tensor
whose shape can be broadcast with `self.a` and `self.b`. For fixed leading
dimensions, the last dimension represents counts for the corresponding Beta
distribution in `self.a` and `self.b`. `x` is only legal if `0 < x < 1`.
"""
distribution_util.append_class_fun_doc(Beta.log_prob, doc_str=_prob_note)
distribution_util.append_class_fun_doc(Beta.prob, doc_str=_prob_note)
distribution_util.append_class_fun_doc(Beta.mode,
doc_str="""
Note that the mode for the Beta distribution is only defined
when `a > 1`, `b > 1`. This returns the mode when `a > 1` and `b > 1`,
and `NaN` otherwise. If `self.allow_nan_stats` is `False`, an exception
will be raised rather than returning `NaN`.
""")
def _mean(self):
return array_ops.identity(self.p)
def _variance(self):
return self.q * self.p
def _std(self):
return math_ops.sqrt(self._variance())
def _mode(self):
return math_ops.cast(self.p > self.q, self.dtype)
distribution_util.append_class_fun_doc(Bernoulli.mode,
doc_str="""
Specific notes:
1 if p > 1-p. 0 otherwise.
""")
@kullback_leibler.RegisterKL(Bernoulli, Bernoulli)
def _kl_bernoulli_bernoulli(a, b, name=None):
"""Calculate the batched KL divergence KL(a || b) with a and b Bernoulli.
Args:
a: instance of a Bernoulli distribution object.
b: instance of a Bernoulli distribution object.
name: (optional) Name to use for created operations.
default is "kl_bernoulli_bernoulli".
Returns:
counts,
self._n,
message="counts are not less than or equal to n."),
distribution_util.assert_integer_form(
counts, message="counts have non-integer components.")
], counts)
_prob_note = """
For each batch member of counts `k`, `P[counts]` is the probability that
after sampling `n` draws from this Binomial distribution, the number of
successes is `k`. Note that different sequences of draws can result in the
same counts, thus the probability includes a combinatorial coefficient.
counts: Non-negative tensor with dtype `dtype` and whose shape can be
broadcast with `self.p` and `self.n`. `counts` is only legal if it is
less than or equal to `n` and its components are equal to integer
values.
"""
distribution_util.append_class_fun_doc(Binomial.log_prob, doc_str=_prob_note)
distribution_util.append_class_fun_doc(Binomial.prob, doc_str=_prob_note)
distribution_util.append_class_fun_doc(Binomial.mode,
doc_str="""
Note that when `(n + 1) * p` is an integer, there are actually two modes.
Namely, `(n + 1) * p` and `(n + 1) * p - 1` are both modes. Here we return
only the larger of the two modes.
""")
message="variance not defined for components of df <= 1"),
], result_where_defined)
def _std(self):
return math_ops.sqrt(self.variance())
def _mode(self):
return array_ops.identity(self.mu)
def _ones(self):
return array_ops.ones(self.batch_shape(), dtype=self.dtype)
distribution_util.append_class_fun_doc(StudentT.mean, doc_str="""
The mean of Student's T equals `mu` if `df > 1`, otherwise it is `NaN`. If
`self.allow_nan_stats=False`, then an exception will be raised rather than
returning `NaN`.
""")
distribution_util.append_class_fun_doc(StudentT.variance, doc_str="""
Variance for Student's T equals
```
df / (df - 2), when df > 2
infinity, when 1 < df <= 2
NaN, when df <= 1
```
""")
return math_ops.sqrt(self.variance())
def _mode(self):
return math_ops.floor(self.lam)
def _assert_valid_sample(self, x, check_integer=True):
if not self.validate_args: return x
with ops.name_scope('check_x', values=[x]):
dependencies = [check_ops.assert_non_negative(x)]
if check_integer:
dependencies += [distribution_util.assert_integer_form(
x, message="x has non-integer components.")]
return control_flow_ops.with_dependencies(dependencies, x)
_prob_note = """
Note thet the input value must be a non-negative floating point tensor with
dtype `dtype` and whose shape can be broadcast with `self.lam`. `x` is only
legal if it is non-negative and its components are equal to integer values.
"""
distribution_util.append_class_fun_doc(Poisson.log_prob, doc_str=_prob_note)
distribution_util.append_class_fun_doc(Poisson.prob, doc_str=_prob_note)
distribution_util.append_class_fun_doc(Poisson.mode, doc_str="""
Note that when `lam` is an integer, there are actually two modes.
Namely, `lam` and `lam - 1` are both modes. Here we return
only the larger of the two modes.
""")
_prob_note = """
For each batch of counts `[n_1,...,n_k]`, `P[counts]` is the probability
that after sampling `n` draws from this Dirichlet Multinomial
distribution, the number of draws falling in class `j` is `n_j`. Note that
different sequences of draws can result in the same counts, thus the
probability includes a combinatorial coefficient.
Note that input, "counts", must be a non-negative tensor with dtype `dtype`
and whose shape can be broadcast with `self.alpha`. For fixed leading
dimensions, the last dimension represents counts for the corresponding
Dirichlet Multinomial distribution in `self.alpha`. `counts` is only legal if
it sums up to `n` and its components are equal to integer values.
"""
distribution_util.append_class_fun_doc(DirichletMultinomial.log_prob,
doc_str=_prob_note)
distribution_util.append_class_fun_doc(DirichletMultinomial.prob,
doc_str=_prob_note)
distribution_util.append_class_fun_doc(DirichletMultinomial.variance,
doc_str="""
The variance for each batch member is defined as the following:
```
Var(X_j) = n * alpha_j / alpha_0 * (1 - alpha_j / alpha_0) *
(n + alpha_0) / (1 + alpha_0)
```
where `alpha_0 = sum_j alpha_j`.
], result_where_defined)
def _std(self):
return math_ops.sqrt(self.variance())
def _mode(self):
return array_ops.identity(self.mu)
def _ones(self):
return array_ops.ones(self.batch_shape(), dtype=self.dtype)
distribution_util.append_class_fun_doc(StudentT.mean,
doc_str="""
The mean of Student's T equals `mu` if `df > 1`, otherwise it is `NaN`. If
`self.allow_nan_stats=False`, then an exception will be raised rather than
returning `NaN`.
""")
distribution_util.append_class_fun_doc(StudentT.variance,
doc_str="""
Variance for Student's T equals
```
df / (df - 2), when df > 2
infinity, when 1 < df <= 2
NaN, when df <= 1
```
check_ops.assert_positive(x, message="Negative events lie outside Beta distribution support."),
check_ops.assert_less(
x, array_ops.ones((), self.dtype), message="Event>=1 lies outside Beta distribution support."
),
],
x,
)
_prob_note = """
Note that the argument `x` must be a non-negative floating point tensor
whose shape can be broadcast with `self.a` and `self.b`. For fixed leading
dimensions, the last dimension represents counts for the corresponding Beta
distribution in `self.a` and `self.b`. `x` is only legal if `0 < x < 1`.
"""
distribution_util.append_class_fun_doc(Beta.log_prob, doc_str=_prob_note)
distribution_util.append_class_fun_doc(Beta.prob, doc_str=_prob_note)
distribution_util.append_class_fun_doc(
Beta.mode,
doc_str="""
Note that the mode for the Beta distribution is only defined
when `a > 1`, `b > 1`. This returns the mode when `a > 1` and `b > 1`,
and `NaN` otherwise. If `self.allow_nan_stats` is `False`, an exception
will be raised rather than returning `NaN`.
""",
)
y = ops.convert_to_tensor(y, name="y")
with ops.name_scope("inverse"):
if y in self._inverse_cache:
x = self._inverse_cache[y]
elif self.inverse:
x = self.inverse(y)
else:
raise ValueError("No inverse function exists and input `y` was not " "returned from `sample`.")
return x
distribution_util.append_class_fun_doc(
TransformedDistribution.batch_shape,
doc_str="""
The product of the dimensions of the `batch_shape` is the number of
independent distributions of this kind the instance represents.
""",
)
distribution_util.append_class_fun_doc(
TransformedDistribution.sample_n,
doc_str="""
Samples from the base distribution and then passes through the transform.
""",
)
distribution_util.append_class_fun_doc(
TransformedDistribution.log_prob,
return math_ops.floor(self.lam)
def _assert_valid_sample(self, x, check_integer=True):
if not self.validate_args: return x
with ops.name_scope("check_x", values=[x]):
dependencies = [check_ops.assert_non_negative(x)]
if check_integer:
dependencies += [
distribution_util.assert_integer_form(
x, message="x has non-integer components.")
]
return control_flow_ops.with_dependencies(dependencies, x)
_prob_note = """
Note thet the input value must be a non-negative floating point tensor with
dtype `dtype` and whose shape can be broadcast with `self.lam`. `x` is only
legal if it is non-negative and its components are equal to integer values.
"""
distribution_util.append_class_fun_doc(Poisson.log_prob, doc_str=_prob_note)
distribution_util.append_class_fun_doc(Poisson.prob, doc_str=_prob_note)
distribution_util.append_class_fun_doc(Poisson.mode,
doc_str="""
Note that when `lam` is an integer, there are actually two modes.
Namely, `lam` and `lam - 1` are both modes. Here we return
only the larger of the two modes.
""")
def _mode(self):
return array_ops.identity(self._mu)
_prob_note = """
`x` is a batch vector with compatible shape if `x` is a `Tensor` whose
shape can be broadcast up to either:
````
self.batch_shape + self.event_shape
OR
[M1,...,Mm] + self.batch_shape + self.event_shape
```
"""
distribution_util.append_class_fun_doc(_MultivariateNormalOperatorPD.log_prob,
doc_str=_prob_note)
distribution_util.append_class_fun_doc(_MultivariateNormalOperatorPD.prob,
doc_str=_prob_note)
class MultivariateNormalDiag(_MultivariateNormalOperatorPD):
"""The multivariate normal distribution on `R^k`.
This distribution is defined by a 1-D mean `mu` and a 1-D diagonal
`diag_stdev`, representing the standard deviations. This distribution
assumes the random variables, `(X_1,...,X_k)` are independent, thus no
non-diagonal terms of the covariance matrix are needed.
This allows for `O(k)` pdf evaluation, sampling, and storage.
#### Mathematical details
def _mean(self):
return array_ops.identity(self.p)
def _variance(self):
return self.q * self.p
def _std(self):
return math_ops.sqrt(self._variance())
def _mode(self):
return math_ops.cast(self.p > self.q, self.dtype)
distribution_util.append_class_fun_doc(Bernoulli.mode, doc_str="""
Specific notes:
1 if p > 1-p. 0 otherwise.
""")
class BernoulliWithSigmoidP(Bernoulli):
"""Bernoulli with `p = sigmoid(p)`."""
def __init__(self,
p=None,
dtype=dtypes.int32,
validate_args=False,
allow_nan_stats=True,
name="BernoulliWithSigmoidP"):
with ops.name_scope(name) as ns:
super(BernoulliWithSigmoidP, self).__init__(
def _assert_valid_sample(self, counts):
"""Check counts for proper shape, values, then return tensor version."""
if not self.validate_args: return counts
return control_flow_ops.with_dependencies([
check_ops.assert_non_negative(
counts, message="counts has negative components."),
check_ops.assert_equal(
self.n, math_ops.reduce_sum(counts, reduction_indices=[-1]),
message="counts do not sum to n."),
distribution_util.assert_integer_form(
counts, message="counts have non-integer components.")
], counts)
_prob_note = """
For each batch of counts `[n_1,...,n_k]`, `P[counts]` is the probability
that after sampling `n` draws from this Multinomial distribution, the
number of draws falling in class `j` is `n_j`. Note that different
sequences of draws can result in the same counts, thus the probability
includes a combinatorial coefficient.
Note that input "counts" must be a non-negative tensor with dtype `dtype`
and whose shape can be broadcast with `self.p` and `self.n`. For fixed
leading dimensions, the last dimension represents counts for the
corresponding Multinomial distribution in `self.p`. `counts` is only legal
if it sums up to `n` and its components are equal to integer values.
"""
distribution_util.append_class_fun_doc(Multinomial.log_prob, doc_str=_prob_note)
distribution_util.append_class_fun_doc(Multinomial.prob, doc_str=_prob_note)
else:
raise ValueError(
"No inverse function exists and input `y` was not "
"returned from `sample`.")
with ops.name_scope("log_det_jacobian"):
log_det_jacobian = self.log_det_jacobian(x)
return self.base_distribution.log_prob(x) - log_det_jacobian
def _prob(self, y):
return math_ops.exp(self._log_prob(y))
distribution_util.append_class_fun_doc(TransformedDistribution.batch_shape,
doc_str="""
The product of the dimensions of the `batch_shape` is the number of
independent distributions of this kind the instance represents.
""")
distribution_util.append_class_fun_doc(TransformedDistribution.sample_n,
doc_str="""
Samples from the base distribution and then passes through the transform.
""")
distribution_util.append_class_fun_doc(TransformedDistribution.log_prob,
doc_str="""
`(log o p o g)(y) - (log o det o J o g)(y)`,
where `g` is the inverse of `transform`.
nan = np.array(np.nan, dtype=self.dtype.as_numpy_dtype())
return math_ops.select(
self.alpha >= 1., mode,
array_ops.fill(self.batch_shape(), nan, name="nan"))
else:
return control_flow_ops.with_dependencies([
check_ops.assert_less(
array_ops.ones((), self.dtype),
self.alpha,
message="mode not defined for components of alpha <= 1"),
], mode)
distribution_util.append_class_fun_doc(Gamma.sample_n,
doc_str="""
See the documentation for tf.random_gamma for more details.
""")
distribution_util.append_class_fun_doc(Gamma.entropy,
doc_str="""
This is defined to be
```
entropy = alpha - log(beta) + log(Gamma(alpha))
+ (1-alpha)digamma(alpha)
```
where digamma(alpha) is the digamma function.
""")
_prob_note = """
For each batch of counts `[n_1,...,n_k]`, `P[counts]` is the probability
that after sampling `n` draws from this Dirichlet Multinomial
distribution, the number of draws falling in class `j` is `n_j`. Note that
different sequences of draws can result in the same counts, thus the
probability includes a combinatorial coefficient.
Note that input, "counts", must be a non-negative tensor with dtype `dtype`
and whose shape can be broadcast with `self.alpha`. For fixed leading
dimensions, the last dimension represents counts for the corresponding
Dirichlet Multinomial distribution in `self.alpha`. `counts` is only legal if
it sums up to `n` and its components are equal to integer values.
"""
distribution_util.append_class_fun_doc(DirichletMultinomial.log_prob,
doc_str=_prob_note)
distribution_util.append_class_fun_doc(DirichletMultinomial.prob,
doc_str=_prob_note)
distribution_util.append_class_fun_doc(DirichletMultinomial.variance,
doc_str="""
The variance for each batch member is defined as the following:
```
Var(X_j) = n * alpha_j / alpha_0 * (1 - alpha_j / alpha_0) *
(n + alpha_0) / (1 + alpha_0)
```
where `alpha_0 = sum_j alpha_j`.
def _mode(self):
return array_ops.identity(self._mu)
_prob_note = """
`x` is a batch vector with compatible shape if `x` is a `Tensor` whose
shape can be broadcast up to either:
````
self.batch_shape + self.event_shape
OR
[M1,...,Mm] + self.batch_shape + self.event_shape
```
"""
distribution_util.append_class_fun_doc(_MultivariateNormalOperatorPD.log_prob,
doc_str=_prob_note)
distribution_util.append_class_fun_doc(_MultivariateNormalOperatorPD.prob,
doc_str=_prob_note)
class MultivariateNormalDiag(_MultivariateNormalOperatorPD):
"""The multivariate normal distribution on `R^k`.
This distribution is defined by a 1-D mean `mu` and a 1-D diagonal
`diag_stdev`, representing the standard deviations. This distribution
assumes the random variables, `(X_1,...,X_k)` are independent, thus no
non-diagonal terms of the covariance matrix are needed.
This allows for `O(k)` pdf evaluation, sampling, and storage.
#### Mathematical details
_prob_note = _prob_base_note + """
The base distribution's `cdf` method must be defined on `y - 1`. If the
base distribution has a `survival_function` method, results will be more
accurate for large values of `y`, and in this case the `survival_function` must
also be defined on `y - 1`.
"""
_log_prob_note = _prob_base_note + """
The base distribution's `log_cdf` method must be defined on `y - 1`. If the
base distribution has a `log_survival_function` method results will be more
accurate for large values of `y`, and in this case the `log_survival_function`
must also be defined on `y - 1`.
"""
distribution_util.append_class_fun_doc(QuantizedDistribution.prob,
doc_str=_prob_note)
distribution_util.append_class_fun_doc(QuantizedDistribution.log_prob,
doc_str=_log_prob_note)
_cdf_base_note = """
For whole numbers `y`,
```
cdf(y) := P[Y <= y]
= 1, if y >= upper_cutoff,
= 0, if y < lower_cutoff,
= P[X <= y], otherwise.
```
