# 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 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
[docs]class SquashedGaussian(ContinuosDistribution):
"""Gaussian distribution which its output is squashed with tanh.
:math:`z \\sim \\mathcal{N}(\\mu,\\,\\sigma^{2})`. :math:`out = \\tanh{z}`.
Args:
mean (nn.Variable): mean :math:`\\mu` of underlying gaussian distribution.
ln_var (nn.Variable): logarithm of the variance :math:`\\sigma^{2}`. (i.e. ln_var is :math:`\\log{\\sigma^{2}}`)
Note:
The log probability and kl_divergence of this distribution is different from
:py:class:`Gaussian distribution <nnabla_rl.distributions.Gaussian>` because the output is squashed.
"""
def __init__(self, mean, ln_var):
super(SquashedGaussian, self).__init__()
if not isinstance(mean, nn.Variable):
mean = nn.Variable.from_numpy_array(mean)
if not isinstance(ln_var, nn.Variable):
ln_var = nn.Variable.from_numpy_array(ln_var)
self._mean = mean
self._var = NF.exp(ln_var)
self._ln_var = ln_var
self._ndim = mean.shape[-1]
@property
def ndim(self):
return self._ndim
[docs] def sample(self, noise_clip=None):
x = RF.sample_gaussian(mean=self._mean,
ln_var=self._ln_var,
noise_clip=noise_clip)
return NF.tanh(x)
[docs] def sample_multiple(self, num_samples, noise_clip=None):
x = RF.sample_gaussian_multiple(self._mean,
self._ln_var,
num_samples,
noise_clip=noise_clip)
return NF.tanh(x)
[docs] def sample_and_compute_log_prob(self, noise_clip=None):
"""NOTE: In order to avoid sampling different random values for sample
and log_prob, you'll need to use nnabla.forward_all(sample, log_prob)
If you forward the two variables independently, you'll get a log_prob
for different sample, since different random variables are sampled
internally."""
x = RF.sample_gaussian(
mean=self._mean, ln_var=self._ln_var, noise_clip=noise_clip)
log_prob = self._log_prob_internal(
x, self._mean, self._var, self._ln_var)
return NF.tanh(x), log_prob
[docs] def sample_multiple_and_compute_log_prob(self, num_samples, noise_clip=None):
"""NOTE: In order to avoid sampling different random values for sample
and log_prob, you'll need to use nnabla.forward_all(sample, log_prob)
If you forward the two variables independently, you'll get a log_prob
for different sample, since different random variables are sampled
internally."""
x = RF.sample_gaussian_multiple(self._mean,
self._ln_var,
num_samples=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 == (x.shape[0], 1, x.shape[-1])
assert var.shape == mean.shape
assert ln_var.shape == mean.shape
log_prob = self._log_prob_internal(x, mean, var, ln_var)
return NF.tanh(x), log_prob
[docs] def choose_probable(self):
return NF.tanh(self._mean)
[docs] def log_prob(self, x):
x = NF.atanh(x)
return self._log_prob_internal(x, self._mean, self._var, self._ln_var)
def _log_prob_internal(self, x, mean, var, ln_var):
axis = len(x.shape) - 1
gaussian_part = common_utils.gaussian_log_prob(x, mean, var, ln_var)
adjust_part = NF.sum(self._log_determinant_jacobian(x), axis=axis, keepdims=True)
return gaussian_part - adjust_part
def _log_determinant_jacobian(self, x):
# arctanh(y)' = 1/(1 - y^2) (y=tanh(x))
# Below computes log(1 - tanh(x)^2)
# For derivation see:
# https://github.com/tensorflow/probability/blob/master/tensorflow_probability/python/bijectors/tanh.py
return 2.0 * (np.log(2.0) - x - NF.softplus(-2.0 * x))