forked from mindspore-Ecosystem/mindspore
added Cauchy distribution
This commit is contained in:
parent
e65c68a723
commit
1a68ccb40b
|
|
@ -175,6 +175,8 @@ class Bijector(Cell):
|
|||
"""
|
||||
Calculate batch_shape based on parameters.
|
||||
"""
|
||||
if 'param_dict' not in self.parameters.keys():
|
||||
return None
|
||||
param_dict = self.parameters['param_dict']
|
||||
broadcast_shape_tensor = None
|
||||
for value in param_dict.values():
|
||||
|
|
@ -191,6 +193,8 @@ class Bijector(Cell):
|
|||
"""
|
||||
Check if the parameters used during initialization are scalars.
|
||||
"""
|
||||
if 'param_dict' not in self.parameters.keys():
|
||||
return False
|
||||
param_dict = self.parameters['param_dict']
|
||||
for value in param_dict.values():
|
||||
if value is None:
|
||||
|
|
|
|||
|
|
@ -27,6 +27,7 @@ from .categorical import Categorical
|
|||
from .log_normal import LogNormal
|
||||
from .logistic import Logistic
|
||||
from .gumbel import Gumbel
|
||||
from .cauchy import Cauchy
|
||||
|
||||
__all__ = ['Distribution',
|
||||
'TransformedDistribution',
|
||||
|
|
@ -39,4 +40,5 @@ __all__ = ['Distribution',
|
|||
'LogNormal',
|
||||
'Logistic',
|
||||
'Gumbel',
|
||||
'Cauchy',
|
||||
]
|
||||
|
|
|
|||
|
|
@ -234,6 +234,11 @@ def raise_type_error(name, cur_type, required_type):
|
|||
raise TypeError(
|
||||
f"For {name} , the type should be or be subclass of {required_type}, but got {cur_type}")
|
||||
|
||||
@constexpr
|
||||
def raise_not_defined(func_name, obj, *args, **kwargs):
|
||||
raise ValueError(
|
||||
f"{func_name} is undefined for {obj} distribution.")
|
||||
|
||||
|
||||
@constexpr
|
||||
def check_distribution_name(name, expected_name):
|
||||
|
|
|
|||
|
|
@ -0,0 +1,345 @@
|
|||
# Copyright 2020 Huawei Technologies Co., Ltd
|
||||
#
|
||||
# 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.
|
||||
# ============================================================================
|
||||
"""Cauchy Distribution"""
|
||||
import numpy as np
|
||||
from mindspore.ops import operations as P
|
||||
from mindspore.ops import composite as C
|
||||
from mindspore._checkparam import Validator
|
||||
from mindspore.common import dtype as mstype
|
||||
from .distribution import Distribution
|
||||
from ._utils.utils import check_greater_zero, check_distribution_name, raise_not_defined
|
||||
from ._utils.custom_ops import exp_generic, log_generic, log1p_generic
|
||||
|
||||
|
||||
class Cauchy(Distribution):
|
||||
"""
|
||||
Cauchy distribution.
|
||||
|
||||
Args:
|
||||
loc (int, float, list, numpy.ndarray, Tensor, Parameter): The location of the Cauchy distribution.
|
||||
scale (int, float, list, numpy.ndarray, Tensor, Parameter): The scale of the Cauchy distribution.
|
||||
seed (int): The seed used in sampling. The global seed is used if it is None. Default: None.
|
||||
dtype (mindspore.dtype): The type of the event samples. Default: mstype.float32.
|
||||
name (str): The name of the distribution. Default: 'Cauchy'.
|
||||
|
||||
Note:
|
||||
`scale` must be greater than zero.
|
||||
`dist_spec_args` are `loc` and `scale`.
|
||||
`dtype` must be a float type because Cauchy distributions are continuous.
|
||||
Cauchy distribution is not supported on GPU backend.
|
||||
|
||||
Examples:
|
||||
>>> # To initialize a Cauchy distribution of loc 3.0 and scale 4.0.
|
||||
>>> import mindspore.nn.probability.distribution as msd
|
||||
>>> cauchy = msd.Cauchy(3.0, 4.0, dtype=mstype.float32)
|
||||
>>>
|
||||
>>> # The following creates two independent Cauchy distributions.
|
||||
>>> cauchy = msd.Cauchy([3.0, 3.0], [4.0, 4.0], dtype=mstype.float32)
|
||||
>>>
|
||||
>>> # A Cauchy distribution can be initilize without arguments.
|
||||
>>> # In this case, 'loc' and `scale` must be passed in through arguments.
|
||||
>>> cauchy = msd.Cauchy(dtype=mstype.float32)
|
||||
>>>
|
||||
>>> # To use a Cauchy distribution in a network.
|
||||
>>> class net(Cell):
|
||||
>>> def __init__(self):
|
||||
>>> super(net, self).__init__():
|
||||
>>> self.cau1 = msd.Cauchy(0.0, 1.0, dtype=mstype.float32)
|
||||
>>> self.cau2 = msd.Cauchy(dtype=mstype.float32)
|
||||
>>>
|
||||
>>> # The following calls are valid in construct.
|
||||
>>> def construct(self, value, loc_b, scale_b, loc_a, scale_a):
|
||||
>>>
|
||||
>>> # Private interfaces of probability functions corresponding to public interfaces, including
|
||||
>>> # `prob`, `log_prob`, `cdf`, `log_cdf`, `survival_function`, and `log_survival`, have the same arguments as follows.
|
||||
>>> # Args:
|
||||
>>> # value (Tensor): the value to be evaluated.
|
||||
>>> # loc (Tensor): the location of the distribution. Default: self.loc.
|
||||
>>> # scale (Tensor): the scale of the distribution. Default: self.scale.
|
||||
>>>
|
||||
>>> # Examples of `prob`.
|
||||
>>> # Similar calls can be made to other probability functions
|
||||
>>> # by replacing 'prob' by the name of the function
|
||||
>>> ans = self.cau1.prob(value)
|
||||
>>> # Evaluate with respect to distribution b.
|
||||
>>> ans = self.cau1.prob(value, loc_b, scale_b)
|
||||
>>> # `loc` and `scale` must be passed in during function calls
|
||||
>>> ans = self.cau2.prob(value, loc_a, scale_a)
|
||||
>>>
|
||||
>>> # Functions `mode` and `entropy` have the same arguments.
|
||||
>>> # Args:
|
||||
>>> # loc (Tensor): the location of the distribution. Default: self.loc.
|
||||
>>> # scale (Tensor): the scale of the distribution. Default: self.scale.
|
||||
>>>
|
||||
>>> # Example of `mode`.
|
||||
>>> ans = self.cau1.mode() # return 0.0
|
||||
>>> ans = self.cau1.mode(loc_b, scale_b) # return loc_b
|
||||
>>> # `loc` and `scale` must be passed in during function calls.
|
||||
>>> ans = self.cau2.mode(loc_a, scale_a)
|
||||
>>>
|
||||
>>> # Interfaces of 'kl_loss' and 'cross_entropy' are the same:
|
||||
>>> # Args:
|
||||
>>> # dist (str): the type of the distributions. Only "Cauchy" is supported.
|
||||
>>> # loc_b (Tensor): the loc of distribution b.
|
||||
>>> # scale_b (Tensor): the scale distribution b.
|
||||
>>> # loc (Tensor): the loc of distribution a. Default: self.loc.
|
||||
>>> # scale (Tensor): the scale distribution a. Default: self.scale.
|
||||
>>>
|
||||
>>> # Examples of `kl_loss`. `cross_entropy` is similar.
|
||||
>>> ans = self.cau1.kl_loss('Cauchy', loc_b, scale_b)
|
||||
>>> ans = self.cau1.kl_loss('Cauchy', loc_b, scale_b, loc_a, scale_a)
|
||||
>>> # Additional `loc` and `scale` must be passed in.
|
||||
>>> ans = self.cau2.kl_loss('Cauchy', loc_b, scale_b, loc_a, scale_a)
|
||||
>>>
|
||||
>>> # Examples of `sample`.
|
||||
>>> # Args:
|
||||
>>> # shape (tuple): the shape of the sample. Default: ()
|
||||
>>> # loc (Tensor): the location of the distribution. Default: self.loc.
|
||||
>>> # scale (Tensor): the scale of the distribution. Default: self.scale.
|
||||
>>> ans = self.cau1.sample()
|
||||
>>> ans = self.cau1.sample((2,3))
|
||||
>>> ans = self.cau1.sample((2,3), loc_b, s_b)
|
||||
>>> ans = self.cau2.sample((2,3), loc_a, s_a)
|
||||
"""
|
||||
|
||||
def __init__(self,
|
||||
loc=None,
|
||||
scale=None,
|
||||
seed=None,
|
||||
dtype=mstype.float32,
|
||||
name="Cauchy"):
|
||||
"""
|
||||
Constructor of Cauchy.
|
||||
"""
|
||||
param = dict(locals())
|
||||
param['param_dict'] = {'loc': loc, 'scale': scale}
|
||||
valid_dtype = mstype.float_type
|
||||
Validator.check_type_name("dtype", dtype, valid_dtype, type(self).__name__)
|
||||
super(Cauchy, self).__init__(seed, dtype, name, param)
|
||||
|
||||
self._loc = self._add_parameter(loc, 'loc')
|
||||
self._scale = self._add_parameter(scale, 'scale')
|
||||
if self._scale is not None:
|
||||
check_greater_zero(self._scale, "scale")
|
||||
|
||||
# ops needed for the class
|
||||
self.atan = P.Atan()
|
||||
self.cast = P.Cast()
|
||||
self.const = P.ScalarToArray()
|
||||
self.dtypeop = P.DType()
|
||||
self.exp = exp_generic
|
||||
self.fill = P.Fill()
|
||||
self.less = P.Less()
|
||||
self.log = log_generic
|
||||
self.log1p = log1p_generic
|
||||
self.squeeze = P.Squeeze(0)
|
||||
self.shape = P.Shape()
|
||||
self.sq = P.Square()
|
||||
self.sqrt = P.Sqrt()
|
||||
self.tan = P.Tan()
|
||||
self.uniform = C.uniform
|
||||
|
||||
|
||||
def extend_repr(self):
|
||||
if self.is_scalar_batch:
|
||||
str_info = f'location = {self._loc}, scale = {self._scale}'
|
||||
else:
|
||||
str_info = f'batch_shape = {self._broadcast_shape}'
|
||||
return str_info
|
||||
|
||||
@property
|
||||
def loc(self):
|
||||
"""
|
||||
Return the location of the distribution.
|
||||
"""
|
||||
return self._loc
|
||||
|
||||
@property
|
||||
def scale(self):
|
||||
"""
|
||||
Return the scale of the distribution.
|
||||
"""
|
||||
return self._scale
|
||||
|
||||
def _get_dist_type(self):
|
||||
return "Cauchy"
|
||||
|
||||
def _get_dist_args(self, loc=None, scale=None):
|
||||
if loc is not None:
|
||||
self.checktensor(loc, 'loc')
|
||||
else:
|
||||
loc = self.loc
|
||||
if scale is not None:
|
||||
self.checktensor(scale, 'scale')
|
||||
else:
|
||||
scale = self.scale
|
||||
return loc, scale
|
||||
|
||||
def _mode(self, loc=None, scale=None):
|
||||
"""
|
||||
The mode of the distribution.
|
||||
"""
|
||||
loc, scale = self._check_param_type(loc, scale)
|
||||
return loc
|
||||
|
||||
def _mean(self, *args, **kwargs):
|
||||
return raise_not_defined('mean', 'Cauchy', *args, **kwargs)
|
||||
|
||||
def _sd(self, *args, **kwargs):
|
||||
return raise_not_defined('standard deviation', 'Cauchy', *args, **kwargs)
|
||||
|
||||
def _var(self, *args, **kwargs):
|
||||
return raise_not_defined('variance', 'Cauchy', *args, **kwargs)
|
||||
|
||||
def _entropy(self, loc=None, scale=None):
|
||||
r"""
|
||||
Evaluate entropy.
|
||||
|
||||
.. math::
|
||||
H(X) = \log(4 * \Pi * scale)
|
||||
"""
|
||||
loc, scale = self._check_param_type(loc, scale)
|
||||
return self.log(4 * np.pi * scale)
|
||||
|
||||
def _log_prob(self, value, loc=None, scale=None):
|
||||
r"""
|
||||
Evaluate log probability.
|
||||
|
||||
Args:
|
||||
value (Tensor): The value to be evaluated.
|
||||
loc (Tensor): The location of the distribution. Default: self.loc.
|
||||
scale (Tensor): The scale of the distribution. Default: self.scale.
|
||||
|
||||
.. math::
|
||||
L(x) = \log(\frac{1}{\pi * scale} * \frac{scale^{2}}{(x - loc)^{2} + scale^{2}})
|
||||
"""
|
||||
value = self._check_value(value, 'value')
|
||||
value = self.cast(value, self.dtype)
|
||||
loc, scale = self._check_param_type(loc, scale)
|
||||
z = (value - loc) / scale
|
||||
log_unnormalized_prob = - self.log1p(self.sq(z))
|
||||
log_normalization = self.log(np.pi * scale)
|
||||
return log_unnormalized_prob - log_normalization
|
||||
|
||||
def _cdf(self, value, loc=None, scale=None):
|
||||
r"""
|
||||
Evaluate the cumulative distribution function on the given value.
|
||||
|
||||
Args:
|
||||
value (Tensor): The value to be evaluated.
|
||||
loc (Tensor): The location of the distribution. Default: self.loc.
|
||||
scale (Tensor): The scale the distribution. Default: self.scale.
|
||||
|
||||
.. math::
|
||||
cdf(x) = \frac{\arctan{(x - loc) / scale}}{\pi} + 0.5
|
||||
"""
|
||||
value = self._check_value(value, 'value')
|
||||
value = self.cast(value, self.dtype)
|
||||
loc, scale = self._check_param_type(loc, scale)
|
||||
z = (value - loc) / scale
|
||||
return self.atan(z) / np.pi + 0.5
|
||||
|
||||
def _log_cdf(self, value, loc=None, scale=None):
|
||||
r"""
|
||||
Evaluate the log cumulative distribution function on the given value.
|
||||
|
||||
Args:
|
||||
value (Tensor): The value to be evaluated.
|
||||
loc (Tensor): The location of the distribution. Default: self.loc.
|
||||
scale (Tensor): The scale the distribution. Default: self.scale.
|
||||
|
||||
.. math::
|
||||
log_cdf(x) = \log(\frac{\arctan(\frac{x-loc}{scale})}{\pi} + 0.5)
|
||||
= \log {\arctan(\frac{x-loc}{scale}) + 0.5pi}{pi}
|
||||
= \log1p \frac{2 * arctan(\frac{x-loc}{scale})}{pi} - \log2
|
||||
"""
|
||||
value = self._check_value(value, 'value')
|
||||
value = self.cast(value, self.dtype)
|
||||
loc, scale = self._check_param_type(loc, scale)
|
||||
z = (value - loc) / scale
|
||||
return self.log1p(2. * self.atan(z) / np.pi) - self.log(self.const(2.))
|
||||
|
||||
def _quantile(self, p, loc=None, scale=None):
|
||||
loc, scale = self._check_param_type(loc, scale)
|
||||
return loc + scale * self.tan(np.pi * (p - 0.5))
|
||||
|
||||
def _kl_loss(self, dist, loc_b, scale_b, loc=None, scale=None):
|
||||
r"""
|
||||
Evaluate Cauchy-Cauchy kl divergence, i.e. KL(a||b).
|
||||
|
||||
Args:
|
||||
dist (str): The type of the distributions. Should be "Cauchy" in this case.
|
||||
loc_b (Tensor): The loc of distribution b.
|
||||
scale_b (Tensor): The scale of distribution b.
|
||||
loc (Tensor): The loc of distribution a. Default: self.loc.
|
||||
scale (Tensor): The scale of distribution a. Default: self.scale.
|
||||
|
||||
.. math::
|
||||
KL(a||b) = \log(\frac{(scale_a + scale_b)^{2} + (loc_a - loc_b)^{2}}
|
||||
{4 * scale_a * scale_b})
|
||||
"""
|
||||
check_distribution_name(dist, 'Cauchy')
|
||||
loc, scale = self._check_param_type(loc, scale)
|
||||
loc_b = self._check_value(loc_b, 'loc_b')
|
||||
loc_b = self.cast(loc_b, self.parameter_type)
|
||||
scale_b = self._check_value(scale_b, 'scale_b')
|
||||
scale_b = self.cast(scale_b, self.parameter_type)
|
||||
sum_square = self.sq(scale + scale_b)
|
||||
square_diff = self.sq(loc - loc_b)
|
||||
return self.log(sum_square + square_diff) - \
|
||||
self.log(self.const(4.0)) - self.log(scale) - self.log(scale_b)
|
||||
|
||||
def _cross_entropy(self, dist, loc_b, scale_b, loc=None, scale=None):
|
||||
r"""
|
||||
Evaluate cross entropy between Cauchy distributions.
|
||||
|
||||
Args:
|
||||
dist (str): The type of the distributions. Should be "Cauchy" in this case.
|
||||
loc_b (Tensor): The loc of distribution b.
|
||||
scale_b (Tensor): The scale of distribution b.
|
||||
loc (Tensor): The loc of distribution a. Default: self.loc.
|
||||
scale (Tensor): The scale of distribution a. Default: self.scale.
|
||||
"""
|
||||
check_distribution_name(dist, 'Cauchy')
|
||||
return self._entropy(loc, scale) + self._kl_loss(dist, loc_b, scale_b, loc, scale)
|
||||
|
||||
def _sample(self, shape=(), loc=None, scale=None):
|
||||
"""
|
||||
Sampling.
|
||||
|
||||
Args:
|
||||
shape (tuple): The shape of the sample. Default: ().
|
||||
loc (Tensor): The location of the samples. Default: self.loc.
|
||||
scale (Tensor): The scale of the samples. Default: self.scale.
|
||||
|
||||
Returns:
|
||||
Tensor, with the shape being shape + batch_shape.
|
||||
"""
|
||||
shape = self.checktuple(shape, 'shape')
|
||||
loc, scale = self._check_param_type(loc, scale)
|
||||
batch_shape = self.shape(loc + scale)
|
||||
origin_shape = shape + batch_shape
|
||||
if origin_shape == ():
|
||||
sample_shape = (1,)
|
||||
else:
|
||||
sample_shape = origin_shape
|
||||
l_zero = self.const(0.0)
|
||||
h_one = self.const(1.0)
|
||||
sample_uniform = self.uniform(sample_shape, l_zero, h_one, self.seed)
|
||||
sample = self._quantile(sample_uniform, loc, scale)
|
||||
value = self.cast(sample, self.dtype)
|
||||
if origin_shape == ():
|
||||
value = self.squeeze(value)
|
||||
return value
|
||||
|
|
@ -21,7 +21,7 @@ import mindspore.nn as nn
|
|||
import mindspore.nn.probability.bijector as msb
|
||||
import mindspore.nn.probability.distribution as msd
|
||||
from .transformed_distribution import TransformedDistribution
|
||||
from ._utils.utils import check_distribution_name, raise_not_implemented_util
|
||||
from ._utils.utils import check_distribution_name
|
||||
from ._utils.custom_ops import exp_generic, expm1_generic, log_generic
|
||||
|
||||
class Gumbel(TransformedDistribution):
|
||||
|
|
@ -39,6 +39,7 @@ class Gumbel(TransformedDistribution):
|
|||
`scale` must be greater than zero.
|
||||
`dist_spec_args` are `loc` and `scale`.
|
||||
`dtype` must be a float type because Gumbel distributions are continuous.
|
||||
`kl_loss` and `cross_entropy` are not supported on GPU backend.
|
||||
|
||||
Examples:
|
||||
>>> # To initialize a Gumbel distribution of `loc` 3.0 and `scale` 4.0.
|
||||
|
|
@ -219,8 +220,6 @@ class Gumbel(TransformedDistribution):
|
|||
loc_b (Tensor): The loc of distribution b.
|
||||
scale_b (Tensor): The scale of distribution b.
|
||||
"""
|
||||
if self.device_target == 'GPU':
|
||||
raise_not_implemented_util('On GPU backend, cross_entropy', self.name)
|
||||
check_distribution_name(dist, 'Gumbel')
|
||||
return self._entropy() + self._kl_loss(dist, loc_b, scale_b)
|
||||
|
||||
|
|
@ -237,8 +236,6 @@ class Gumbel(TransformedDistribution):
|
|||
KL(a||b) = \log(scale_b / scale_a) + Euler-Mascheroni_constant * (scale_a / scale_b - 1.) +
|
||||
\exp(\frac{(loc_b - loc_a)}{scale_b}) * \Gamma(scale_a / scale_b + 1.) - 1.
|
||||
"""
|
||||
if self.device_target == 'GPU':
|
||||
raise_not_implemented_util('On GPU backend, kl_loss', self.name)
|
||||
check_distribution_name(dist, 'Gumbel')
|
||||
loc_b = self._check_value(loc_b, 'loc_b')
|
||||
scale_b = self._check_value(scale_b, 'scale_b')
|
||||
|
|
|
|||
|
|
@ -0,0 +1,282 @@
|
|||
# Copyright 2019 Huawei Technologies Co., Ltd
|
||||
#
|
||||
# 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.
|
||||
# ============================================================================
|
||||
"""test cases for Cauchy distribution"""
|
||||
import numpy as np
|
||||
from scipy import stats
|
||||
import mindspore.context as context
|
||||
import mindspore.nn as nn
|
||||
import mindspore.nn.probability.distribution as msd
|
||||
from mindspore import Tensor
|
||||
from mindspore import dtype
|
||||
|
||||
context.set_context(mode=context.GRAPH_MODE, device_target="Ascend")
|
||||
|
||||
class Prob(nn.Cell):
|
||||
"""
|
||||
Test class: probability of Cauchy distribution.
|
||||
"""
|
||||
def __init__(self):
|
||||
super(Prob, self).__init__()
|
||||
self.c = msd.Cauchy(np.array([3.0]), np.array([[2.0], [4.0]]), dtype=dtype.float32)
|
||||
|
||||
def construct(self, x_):
|
||||
return self.c.prob(x_)
|
||||
|
||||
def test_pdf():
|
||||
"""
|
||||
Test pdf.
|
||||
"""
|
||||
cauchy_benchmark = stats.cauchy(np.array([3.0]), np.array([[2.0], [4.0]]))
|
||||
expect_pdf = cauchy_benchmark.pdf([1.0, 2.0]).astype(np.float32)
|
||||
pdf = Prob()
|
||||
output = pdf(Tensor([1.0, 2.0], dtype=dtype.float32))
|
||||
tol = 1e-6
|
||||
assert (np.abs(output.asnumpy() - expect_pdf) < tol).all()
|
||||
|
||||
class LogProb(nn.Cell):
|
||||
"""
|
||||
Test class: log probability of Cauchy distribution.
|
||||
"""
|
||||
def __init__(self):
|
||||
super(LogProb, self).__init__()
|
||||
self.c = msd.Cauchy(np.array([3.0]), np.array([[2.0], [4.0]]), dtype=dtype.float32)
|
||||
|
||||
def construct(self, x_):
|
||||
return self.c.log_prob(x_)
|
||||
|
||||
def test_log_likelihood():
|
||||
"""
|
||||
Test log_pdf.
|
||||
"""
|
||||
cauchy_benchmark = stats.cauchy(np.array([3.0]), np.array([[2.0], [4.0]]))
|
||||
expect_logpdf = cauchy_benchmark.logpdf([1.0, 2.0]).astype(np.float32)
|
||||
logprob = LogProb()
|
||||
output = logprob(Tensor([1.0, 2.0], dtype=dtype.float32))
|
||||
tol = 1e-6
|
||||
assert (np.abs(output.asnumpy() - expect_logpdf) < tol).all()
|
||||
|
||||
class KL(nn.Cell):
|
||||
"""
|
||||
Test class: kl_loss of Cauchy distribution.
|
||||
"""
|
||||
def __init__(self):
|
||||
super(KL, self).__init__()
|
||||
self.c = msd.Cauchy(np.array([3.]), np.array([4.]), dtype=dtype.float32)
|
||||
|
||||
def construct(self, mu, s):
|
||||
return self.c.kl_loss('Cauchy', mu, s)
|
||||
|
||||
def test_kl_loss():
|
||||
"""
|
||||
Test kl_loss.
|
||||
"""
|
||||
loc_b = np.array([0.]).astype(np.float32)
|
||||
scale_b = np.array([1.]).astype(np.float32)
|
||||
|
||||
loc_a = np.array([3.0]).astype(np.float32)
|
||||
scale_a = np.array([4.0]).astype(np.float32)
|
||||
|
||||
sum_square = np.square(scale_a + scale_b)
|
||||
square_diff = np.square(loc_a - loc_b)
|
||||
expect_kl_loss = np.log(sum_square + square_diff) - \
|
||||
np.log(4.0 * scale_a * scale_b)
|
||||
|
||||
kl_loss = KL()
|
||||
loc = Tensor(loc_b, dtype=dtype.float32)
|
||||
scale = Tensor(scale_b, dtype=dtype.float32)
|
||||
output = kl_loss(loc, scale)
|
||||
tol = 1e-6
|
||||
assert (np.abs(output.asnumpy() - expect_kl_loss) < tol).all()
|
||||
|
||||
class Basics(nn.Cell):
|
||||
"""
|
||||
Test class: mode of Cauchy distribution.
|
||||
"""
|
||||
def __init__(self):
|
||||
super(Basics, self).__init__()
|
||||
self.c = msd.Cauchy(np.array([3.0]), np.array([2.0, 4.0]), dtype=dtype.float32)
|
||||
|
||||
def construct(self):
|
||||
return self.c.mode()
|
||||
|
||||
def test_basics():
|
||||
"""
|
||||
Test mode.
|
||||
"""
|
||||
basics = Basics()
|
||||
mode = basics()
|
||||
expect_mode = np.array([3.0, 3.0])
|
||||
tol = 1e-6
|
||||
assert (np.abs(mode.asnumpy() - expect_mode) < tol).all()
|
||||
|
||||
class Sampling(nn.Cell):
|
||||
"""
|
||||
Test class: sample of Cauchy distribution.
|
||||
"""
|
||||
def __init__(self, shape, seed=0):
|
||||
super(Sampling, self).__init__()
|
||||
self.c = msd.Cauchy(np.array([3.0]), np.array([[2.0], [4.0]]), seed=seed, dtype=dtype.float32)
|
||||
self.shape = shape
|
||||
|
||||
def construct(self, mean=None, sd=None):
|
||||
return self.c.sample(self.shape, mean, sd)
|
||||
|
||||
def test_sample():
|
||||
"""
|
||||
Test sample.
|
||||
"""
|
||||
shape = (2, 3)
|
||||
seed = 10
|
||||
mean = Tensor([2.0], dtype=dtype.float32)
|
||||
sd = Tensor([2.0, 2.0, 2.0], dtype=dtype.float32)
|
||||
sample = Sampling(shape, seed=seed)
|
||||
output = sample(mean, sd)
|
||||
assert output.shape == (2, 3, 3)
|
||||
|
||||
class CDF(nn.Cell):
|
||||
"""
|
||||
Test class: cdf of Cauchy distribution.
|
||||
"""
|
||||
def __init__(self):
|
||||
super(CDF, self).__init__()
|
||||
self.c = msd.Cauchy(np.array([3.0]), np.array([[2.0], [4.0]]), dtype=dtype.float32)
|
||||
|
||||
def construct(self, x_):
|
||||
return self.c.cdf(x_)
|
||||
|
||||
|
||||
def test_cdf():
|
||||
"""
|
||||
Test cdf.
|
||||
"""
|
||||
cauchy_benchmark = stats.cauchy(np.array([3.0]), np.array([[2.0], [4.0]]))
|
||||
expect_cdf = cauchy_benchmark.cdf([1.0, 2.0]).astype(np.float32)
|
||||
cdf = CDF()
|
||||
output = cdf(Tensor([1.0, 2.0], dtype=dtype.float32))
|
||||
tol = 2e-5
|
||||
assert (np.abs(output.asnumpy() - expect_cdf) < tol).all()
|
||||
|
||||
class LogCDF(nn.Cell):
|
||||
"""
|
||||
Test class: log_cdf of Cauchy distribution.
|
||||
"""
|
||||
def __init__(self):
|
||||
super(LogCDF, self).__init__()
|
||||
self.c = msd.Cauchy(np.array([3.0]), np.array([[2.0], [4.0]]), dtype=dtype.float32)
|
||||
|
||||
def construct(self, x_):
|
||||
return self.c.log_cdf(x_)
|
||||
|
||||
def test_log_cdf():
|
||||
"""
|
||||
Test log cdf.
|
||||
"""
|
||||
cauchy_benchmark = stats.cauchy(np.array([3.0]), np.array([[2.0], [4.0]]))
|
||||
expect_logcdf = cauchy_benchmark.logcdf([1.0, 2.0]).astype(np.float32)
|
||||
logcdf = LogCDF()
|
||||
output = logcdf(Tensor([1.0, 2.0], dtype=dtype.float32))
|
||||
tol = 5e-5
|
||||
assert (np.abs(output.asnumpy() - expect_logcdf) < tol).all()
|
||||
|
||||
class SF(nn.Cell):
|
||||
"""
|
||||
Test class: survival function of Cauchy distribution.
|
||||
"""
|
||||
def __init__(self):
|
||||
super(SF, self).__init__()
|
||||
self.c = msd.Cauchy(np.array([3.0]), np.array([[2.0], [4.0]]), dtype=dtype.float32)
|
||||
|
||||
def construct(self, x_):
|
||||
return self.c.survival_function(x_)
|
||||
|
||||
def test_survival():
|
||||
"""
|
||||
Test log_survival.
|
||||
"""
|
||||
cauchy_benchmark = stats.cauchy(np.array([3.0]), np.array([[2.0], [4.0]]))
|
||||
expect_survival = cauchy_benchmark.sf([1.0, 2.0]).astype(np.float32)
|
||||
survival_function = SF()
|
||||
output = survival_function(Tensor([1.0, 2.0], dtype=dtype.float32))
|
||||
tol = 2e-5
|
||||
assert (np.abs(output.asnumpy() - expect_survival) < tol).all()
|
||||
|
||||
class LogSF(nn.Cell):
|
||||
"""
|
||||
Test class: log survival function of Cauchy distribution.
|
||||
"""
|
||||
def __init__(self):
|
||||
super(LogSF, self).__init__()
|
||||
self.c = msd.Cauchy(np.array([3.0]), np.array([[2.0], [4.0]]), dtype=dtype.float32)
|
||||
|
||||
def construct(self, x_):
|
||||
return self.c.log_survival(x_)
|
||||
|
||||
def test_log_survival():
|
||||
"""
|
||||
Test log_survival.
|
||||
"""
|
||||
cauchy_benchmark = stats.cauchy(np.array([3.0]), np.array([[2.0], [4.0]]))
|
||||
expect_log_survival = cauchy_benchmark.logsf([1.0, 2.0]).astype(np.float32)
|
||||
log_survival = LogSF()
|
||||
output = log_survival(Tensor([1.0, 2.0], dtype=dtype.float32))
|
||||
tol = 2e-5
|
||||
assert (np.abs(output.asnumpy() - expect_log_survival) < tol).all()
|
||||
|
||||
class EntropyH(nn.Cell):
|
||||
"""
|
||||
Test class: entropy of Cauchy distribution.
|
||||
"""
|
||||
def __init__(self):
|
||||
super(EntropyH, self).__init__()
|
||||
self.c = msd.Cauchy(np.array([3.0]), np.array([[2.0], [4.0]]), dtype=dtype.float32)
|
||||
|
||||
def construct(self):
|
||||
return self.c.entropy()
|
||||
|
||||
def test_entropy():
|
||||
"""
|
||||
Test entropy.
|
||||
"""
|
||||
expect_entropy = np.log(4 * np.pi * np.array([[2.0], [4.0]]))
|
||||
entropy = EntropyH()
|
||||
output = entropy()
|
||||
tol = 1e-6
|
||||
assert (np.abs(output.asnumpy() - expect_entropy) < tol).all()
|
||||
|
||||
class CrossEntropy(nn.Cell):
|
||||
"""
|
||||
Test class: cross entropy between Cauchy distributions.
|
||||
"""
|
||||
def __init__(self):
|
||||
super(CrossEntropy, self).__init__()
|
||||
self.c = msd.Cauchy(np.array([3.]), np.array([[2.0], [4.0]]), dtype=dtype.float32)
|
||||
|
||||
def construct(self, mu, s):
|
||||
entropy = self.c.entropy()
|
||||
kl_loss = self.c.kl_loss('Cauchy', mu, s)
|
||||
h_sum_kl = entropy + kl_loss
|
||||
cross_entropy = self.c.cross_entropy('Cauchy', mu, s)
|
||||
return h_sum_kl - cross_entropy
|
||||
|
||||
def test_cross_entropy():
|
||||
"""
|
||||
Test cross_entropy.
|
||||
"""
|
||||
cross_entropy = CrossEntropy()
|
||||
mean = Tensor([1.0], dtype=dtype.float32)
|
||||
sd = Tensor([1.0], dtype=dtype.float32)
|
||||
diff = cross_entropy(mean, sd)
|
||||
tol = 1e-6
|
||||
assert (np.abs(diff.asnumpy() - np.zeros(diff.shape)) < tol).all()
|
||||
|
|
@ -0,0 +1,231 @@
|
|||
# Copyright 2020 Huawei Technologies Co., Ltd
|
||||
#
|
||||
# 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.
|
||||
# ============================================================================
|
||||
"""
|
||||
Test nn.probability.distribution.cauchy.
|
||||
"""
|
||||
import pytest
|
||||
|
||||
import mindspore.nn as nn
|
||||
import mindspore.nn.probability.distribution as msd
|
||||
from mindspore import dtype
|
||||
from mindspore import Tensor
|
||||
|
||||
def test_cauchy_shape_errpr():
|
||||
"""
|
||||
Invalid shapes.
|
||||
"""
|
||||
with pytest.raises(ValueError):
|
||||
msd.Cauchy([[2.], [1.]], [[2.], [3.], [4.]], dtype=dtype.float32)
|
||||
|
||||
def test_type():
|
||||
with pytest.raises(TypeError):
|
||||
msd.Cauchy(0., 1., dtype=dtype.int32)
|
||||
|
||||
def test_name():
|
||||
with pytest.raises(TypeError):
|
||||
msd.Cauchy(0., 1., name=1.0)
|
||||
|
||||
def test_seed():
|
||||
with pytest.raises(TypeError):
|
||||
msd.Cauchy(0., 1., seed='seed')
|
||||
|
||||
def test_scale():
|
||||
with pytest.raises(ValueError):
|
||||
msd.Cauchy(0., 0.)
|
||||
with pytest.raises(ValueError):
|
||||
msd.Cauchy(0., -1.)
|
||||
|
||||
def test_arguments():
|
||||
"""
|
||||
args passing during initialization.
|
||||
"""
|
||||
l = msd.Cauchy()
|
||||
assert isinstance(l, msd.Distribution)
|
||||
l = msd.Cauchy([3.0], [4.0], dtype=dtype.float32)
|
||||
assert isinstance(l, msd.Distribution)
|
||||
|
||||
|
||||
class CauchyProb(nn.Cell):
|
||||
"""
|
||||
Cauchy distribution: initialize with loc/scale.
|
||||
"""
|
||||
def __init__(self):
|
||||
super(CauchyProb, self).__init__()
|
||||
self.cauchy = msd.Cauchy(3.0, 4.0, dtype=dtype.float32)
|
||||
|
||||
def construct(self, value):
|
||||
prob = self.cauchy.prob(value)
|
||||
log_prob = self.cauchy.log_prob(value)
|
||||
cdf = self.cauchy.cdf(value)
|
||||
log_cdf = self.cauchy.log_cdf(value)
|
||||
sf = self.cauchy.survival_function(value)
|
||||
log_sf = self.cauchy.log_survival(value)
|
||||
return prob + log_prob + cdf + log_cdf + sf + log_sf
|
||||
|
||||
def test_cauchy_prob():
|
||||
"""
|
||||
Test probability functions: passing value through construct.
|
||||
"""
|
||||
net = CauchyProb()
|
||||
value = Tensor([0.5, 1.0], dtype=dtype.float32)
|
||||
ans = net(value)
|
||||
assert isinstance(ans, Tensor)
|
||||
|
||||
|
||||
class CauchyProb1(nn.Cell):
|
||||
"""
|
||||
Cauchy distribution: initialize without loc/scale.
|
||||
"""
|
||||
def __init__(self):
|
||||
super(CauchyProb1, self).__init__()
|
||||
self.cauchy = msd.Cauchy()
|
||||
|
||||
def construct(self, value, mu, s):
|
||||
prob = self.cauchy.prob(value, mu, s)
|
||||
log_prob = self.cauchy.log_prob(value, mu, s)
|
||||
cdf = self.cauchy.cdf(value, mu, s)
|
||||
log_cdf = self.cauchy.log_cdf(value, mu, s)
|
||||
sf = self.cauchy.survival_function(value, mu, s)
|
||||
log_sf = self.cauchy.log_survival(value, mu, s)
|
||||
return prob + log_prob + cdf + log_cdf + sf + log_sf
|
||||
|
||||
def test_cauchy_prob1():
|
||||
"""
|
||||
Test probability functions: passing loc/scale, value through construct.
|
||||
"""
|
||||
net = CauchyProb1()
|
||||
value = Tensor([0.5, 1.0], dtype=dtype.float32)
|
||||
mu = Tensor([0.0], dtype=dtype.float32)
|
||||
s = Tensor([1.0], dtype=dtype.float32)
|
||||
ans = net(value, mu, s)
|
||||
assert isinstance(ans, Tensor)
|
||||
|
||||
class KL(nn.Cell):
|
||||
"""
|
||||
Test kl_loss and cross entropy.
|
||||
"""
|
||||
def __init__(self):
|
||||
super(KL, self).__init__()
|
||||
self.cauchy = msd.Cauchy(3.0, 4.0)
|
||||
self.cauchy1 = msd.Cauchy()
|
||||
|
||||
def construct(self, mu, s, mu_a, s_a):
|
||||
kl = self.cauchy.kl_loss('Cauchy', mu, s)
|
||||
kl1 = self.cauchy1.kl_loss('Cauchy', mu, s, mu_a, s_a)
|
||||
cross_entropy = self.cauchy.cross_entropy('Cauchy', mu, s)
|
||||
cross_entropy1 = self.cauchy.cross_entropy('Cauchy', mu, s, mu_a, s_a)
|
||||
return kl + kl1 + cross_entropy + cross_entropy1
|
||||
|
||||
def test_kl_cross_entropy():
|
||||
"""
|
||||
Test kl_loss and cross_entropy.
|
||||
"""
|
||||
net = KL()
|
||||
mu = Tensor([0.0], dtype=dtype.float32)
|
||||
s = Tensor([1.0], dtype=dtype.float32)
|
||||
mu_a = Tensor([0.0], dtype=dtype.float32)
|
||||
s_a = Tensor([1.0], dtype=dtype.float32)
|
||||
ans = net(mu, s, mu_a, s_a)
|
||||
assert isinstance(ans, Tensor)
|
||||
|
||||
|
||||
class CauchyBasics(nn.Cell):
|
||||
"""
|
||||
Test class: basic loc/scale function.
|
||||
"""
|
||||
def __init__(self):
|
||||
super(CauchyBasics, self).__init__()
|
||||
self.cauchy = msd.Cauchy(3.0, 4.0, dtype=dtype.float32)
|
||||
|
||||
def construct(self):
|
||||
mode = self.cauchy.mode()
|
||||
entropy = self.cauchy.entropy()
|
||||
return mode + entropy
|
||||
|
||||
class CauchyMean(nn.Cell):
|
||||
"""
|
||||
Test class: basic loc/scale function.
|
||||
"""
|
||||
def __init__(self):
|
||||
super(CauchyMean, self).__init__()
|
||||
self.cauchy = msd.Cauchy(3.0, 4.0, dtype=dtype.float32)
|
||||
|
||||
def construct(self):
|
||||
return self.cauchy.mean()
|
||||
|
||||
class CauchyVar(nn.Cell):
|
||||
"""
|
||||
Test class: basic loc/scale function.
|
||||
"""
|
||||
def __init__(self):
|
||||
super(CauchyVar, self).__init__()
|
||||
self.cauchy = msd.Cauchy(3.0, 4.0, dtype=dtype.float32)
|
||||
|
||||
def construct(self):
|
||||
return self.cauchy.var()
|
||||
|
||||
class CauchySd(nn.Cell):
|
||||
"""
|
||||
Test class: basic loc/scale function.
|
||||
"""
|
||||
def __init__(self):
|
||||
super(CauchySd, self).__init__()
|
||||
self.cauchy = msd.Cauchy(3.0, 4.0, dtype=dtype.float32)
|
||||
|
||||
def construct(self):
|
||||
return self.cauchy.sd()
|
||||
|
||||
def test_bascis():
|
||||
"""
|
||||
Test mean/sd/var/mode/entropy functionality of Cauchy.
|
||||
"""
|
||||
net = CauchyBasics()
|
||||
ans = net()
|
||||
assert isinstance(ans, Tensor)
|
||||
with pytest.raises(ValueError):
|
||||
net = CauchyMean()
|
||||
ans = net()
|
||||
with pytest.raises(ValueError):
|
||||
net = CauchyVar()
|
||||
ans = net()
|
||||
with pytest.raises(ValueError):
|
||||
net = CauchySd()
|
||||
ans = net()
|
||||
|
||||
class CauchyConstruct(nn.Cell):
|
||||
"""
|
||||
Cauchy distribution: going through construct.
|
||||
"""
|
||||
def __init__(self):
|
||||
super(CauchyConstruct, self).__init__()
|
||||
self.cauchy = msd.Cauchy(3.0, 4.0)
|
||||
self.cauchy1 = msd.Cauchy()
|
||||
|
||||
def construct(self, value, mu, s):
|
||||
prob = self.cauchy('prob', value)
|
||||
prob1 = self.cauchy('prob', value, mu, s)
|
||||
prob2 = self.cauchy1('prob', value, mu, s)
|
||||
return prob + prob1 + prob2
|
||||
|
||||
def test_cauchy_construct():
|
||||
"""
|
||||
Test probability function going through construct.
|
||||
"""
|
||||
net = CauchyConstruct()
|
||||
value = Tensor([0.5, 1.0], dtype=dtype.float32)
|
||||
mu = Tensor([0.0], dtype=dtype.float32)
|
||||
s = Tensor([1.0], dtype=dtype.float32)
|
||||
ans = net(value, mu, s)
|
||||
assert isinstance(ans, Tensor)
|
||||
Loading…
Reference in New Issue