# Copyright 2020,2021 Sony Corporation.
# Copyright 2021,2022,2023 Sony Group Corporation.
#
# 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.
import warnings
from typing import Union
import numpy as np
import nnabla as nn
import nnabla.functions as NF
import nnabla_rl.functions as RF
from nnabla_rl.distributions import common_utils
from nnabla_rl.distributions.distribution import ContinuosDistribution, Distribution
[docs]class Gaussian(ContinuosDistribution):
"""Gaussian distribution.
:math:`\\mathcal{N}(\\mu,\\,\\sigma^{2})`
Args:
mean (nn.Variable): mean :math:`\\mu` of gaussian distribution.
ln_var (nn.Variable): logarithm of the variance :math:`\\sigma^{2}`. (i.e. ln_var is :math:`\\log{\\sigma^{2}}`)
"""
_delegate: Union["NumpyGaussian", "NnablaGaussian"]
def __init__(self, mean: Union[nn.Variable, np.ndarray], ln_var: Union[nn.Variable, np.ndarray]):
super(Gaussian, self).__init__()
if isinstance(mean, np.ndarray) and isinstance(ln_var, np.ndarray):
warnings.warn(
"Numpy ndarrays are given as mean and ln_var.\n"
"From v0.12.0, if numpy.ndarray is given, "
"all Gaussian class methods return numpy.ndarray not nnabla.Variable")
self._delegate = NumpyGaussian(mean, ln_var)
elif isinstance(mean, nn.Variable) and isinstance(ln_var, nn.Variable):
self._delegate = NnablaGaussian(mean, ln_var)
else:
raise ValueError(
f"Invalid type or a pair of types, mean type is {type(mean)} and ln type is {type(ln_var)}")
@property
def ndim(self):
return self._delegate.ndim
[docs] def sample(self, noise_clip=None):
return self._delegate.sample(noise_clip)
[docs] def sample_multiple(self, num_samples, noise_clip=None):
return self._delegate.sample_multiple(num_samples, noise_clip)
[docs] def sample_and_compute_log_prob(self, noise_clip=None):
return self._delegate.sample_and_compute_log_prob(noise_clip)
def sample_multiple_and_compute_log_prob(self, num_samples, noise_clip=None):
return self._delegate.sample_multiple_and_compute_log_prob(num_samples, noise_clip)
[docs] def choose_probable(self):
return self._delegate.choose_probable()
[docs] def mean(self):
return self._delegate.mean()
def var(self):
return self._delegate.var()
[docs] def log_prob(self, x):
return self._delegate.log_prob(x)
[docs] def entropy(self):
return self._delegate.entropy()
[docs] def kl_divergence(self, q):
assert isinstance(q, Gaussian)
return self._delegate.kl_divergence(q._delegate)
class NnablaGaussian(Distribution):
_mean: nn.Variable
_var: nn.Variable
def __init__(self, mean: nn.Variable, ln_var: nn.Variable):
super(Distribution, self).__init__()
assert mean.shape == ln_var.shape
self._mean = mean
self._var = NF.exp(ln_var)
self._ln_var = ln_var
self._batch_size = mean.shape[0]
self._data_dim = mean.shape[1:]
self._ndim = mean.shape[-1]
@property
def ndim(self):
return self._ndim
def sample(self, noise_clip=None):
return RF.sample_gaussian(self._mean,
self._ln_var,
noise_clip=noise_clip)
def sample_multiple(self, num_samples, noise_clip=None):
return RF.sample_gaussian_multiple(self._mean,
self._ln_var,
num_samples,
noise_clip=noise_clip)
def sample_and_compute_log_prob(self, noise_clip=None):
x = RF.sample_gaussian(mean=self._mean,
ln_var=self._ln_var,
noise_clip=noise_clip)
return x, self.log_prob(x)
def sample_multiple_and_compute_log_prob(self, num_samples, noise_clip=None):
x = RF.sample_gaussian_multiple(self._mean,
self._ln_var,
num_samples,
noise_clip=noise_clip)
mean = RF.expand_dims(self._mean, axis=1)
var = RF.expand_dims(self._var, axis=1)
ln_var = RF.expand_dims(self._ln_var, axis=1)
assert mean.shape == (self._batch_size, 1, ) + self._data_dim
assert var.shape == mean.shape
assert ln_var.shape == mean.shape
return x, common_utils.gaussian_log_prob(x, mean, var, ln_var)
def choose_probable(self):
return self._mean
def mean(self):
return self._mean
def var(self):
return self._var
def log_prob(self, x):
return common_utils.gaussian_log_prob(x, self._mean, self._var, self._ln_var)
def entropy(self):
return NF.sum(0.5 + 0.5 * np.log(2.0 * np.pi) + 0.5 * self._ln_var, axis=1, keepdims=True)
def kl_divergence(self, q):
assert isinstance(q, NnablaGaussian)
p = self
return 0.5 * NF.sum(q._ln_var - p._ln_var + (p._var + (p._mean - q._mean) ** 2.0) / q._var - 1,
axis=1,
keepdims=True)
class NumpyGaussian(Distribution):
_mean: np.ndarray
_var: np.ndarray
def __init__(self, mean: np.ndarray, ln_var: np.ndarray) -> None:
super(Distribution, self).__init__()
self._dim = mean.shape[0]
assert (self._dim, ) == mean.shape
assert (self._dim, self._dim) == ln_var.shape
self._mean = mean
self._var = np.exp(ln_var)
self._inv_var = np.linalg.inv(self._var)
def log_prob(self, x):
log_det_term = np.log(np.linalg.det(2.0 * np.pi * self._var))
diff = self._mean - x
quadratic_term = diff.T.dot(self._inv_var).dot(diff)
return -0.5 * (log_det_term + quadratic_term)
def mean(self):
return self._mean
def var(self):
return self._var
def sample(self, noise_clip=None):
if noise_clip is not None:
raise NotImplementedError
return np.random.multivariate_normal(self._mean, self._var)
def sample_and_compute_log_prob(self, noise_clip=None):
raise NotImplementedError
def sample_multiple(self, num_samples, noise_clip=None):
raise NotImplementedError
def sample_multiple_and_compute_log_prob(self, num_samples, noise_clip=None):
raise NotImplementedError
def kl_divergence(self, q: 'Distribution'):
if not isinstance(q, NumpyGaussian):
raise NotImplementedError
p_mean = self._mean
p_var = self._var
q_mean = q.mean()
q_var = q.var()
q_var_inv = np.linalg.inv(q.var())
trace_term = np.trace(q_var_inv.dot(p_var))
diff = q_mean - p_mean
quadratic_term = diff.T.dot(q_var_inv).dot(diff)
dimension = self._dim
log_det_term = np.log(np.linalg.det(q_var)) - np.log(np.linalg.det(p_var))
return 0.5 * (trace_term + quadratic_term - dimension + log_det_term)