def __init__(
self,
env_spec,
name='CategoricalMLPPolicy',
hidden_sizes=(32, 32),
hidden_nonlinearity=tf.nn.tanh,
prob_network=None,
):
"""
CategoricalMLPPolicy.
A policy that uses a MLP to estimate a categorical distribution.
Args:
env_spec (garage.envs.env_spec.EnvSpec): Environment specification.
hidden_sizes (list[int]): Output dimension of dense layer(s).
For example, (32, 32) means the MLP of this policy consists
of two hidden layers, each with 32 hidden units.
hidden_nonlinearity: Activation function for
intermediate dense layer(s).
prob_network (tf.Tensor): manually specified network for this
policy. If None, a MLP with the network parameters will be
created. If not None, other network params are ignored.
"""
assert isinstance(env_spec.action_space, akro.Discrete)
Serializable.quick_init(self, locals())
self.name = name
self._prob_network_name = 'prob_network'
with tf.variable_scope(name, 'CategoricalMLPPolicy'):
if prob_network is None:
prob_network = MLP(
input_shape=(env_spec.observation_space.flat_dim, ),
output_dim=env_spec.action_space.n,
hidden_sizes=hidden_sizes,
hidden_nonlinearity=hidden_nonlinearity,
output_nonlinearity=tf.nn.softmax,
name=self._prob_network_name,
)
self._l_prob = prob_network.output_layer
self._l_obs = prob_network.input_layer
with tf.name_scope(self._prob_network_name):
prob_network_outputs = L.get_output(prob_network.output_layer)
self._f_prob = tensor_utils.compile_function(
[prob_network.input_layer.input_var], prob_network_outputs)
self._dist = Categorical(env_spec.action_space.n)
super(CategoricalMLPPolicy, self).__init__(env_spec)
LayersPowered.__init__(self, [prob_network.output_layer])
def __init__(
self,
env_spec,
name="CategoricalMLPPolicy",
hidden_sizes=(32, 32),
hidden_nonlinearity=tf.nn.tanh,
prob_network=None,
):
"""
:param env_spec: A spec for the mdp.
:param hidden_sizes: list of sizes for the fully connected
hidden layers
:param hidden_nonlinearity: nonlinearity used for each hidden layer
:param prob_network: manually specified network for this policy, other
network params
are ignored
:return:
"""
assert isinstance(env_spec.action_space, Discrete)
Serializable.quick_init(self, locals())
self.name = name
self._prob_network_name = "prob_network"
with tf.variable_scope(name, "CategoricalMLPPolicy"):
if prob_network is None:
prob_network = MLP(
input_shape=(env_spec.observation_space.flat_dim, ),
output_dim=env_spec.action_space.n,
hidden_sizes=hidden_sizes,
hidden_nonlinearity=hidden_nonlinearity,
output_nonlinearity=tf.nn.softmax,
name=self._prob_network_name,
)
self._l_prob = prob_network.output_layer
self._l_obs = prob_network.input_layer
with tf.name_scope(self._prob_network_name):
prob_network_outputs = L.get_output(prob_network.output_layer)
self._f_prob = tensor_utils.compile_function(
[prob_network.input_layer.input_var], prob_network_outputs)
self._dist = Categorical(env_spec.action_space.n)
super(CategoricalMLPPolicy, self).__init__(env_spec)
LayersPowered.__init__(self, [prob_network.output_layer])
def __init__(self,
env_spec,
name='DeterministicMLPPolicy',
hidden_sizes=(32, 32),
hidden_nonlinearity=tf.nn.relu,
output_nonlinearity=tf.nn.tanh,
prob_network=None,
bn=False):
assert isinstance(env_spec.action_space, akro.Box)
Serializable.quick_init(self, locals())
self._prob_network_name = 'prob_network'
with tf.compat.v1.variable_scope(name, 'DeterministicMLPPolicy'):
if prob_network is None:
prob_network = MLP(
input_shape=(env_spec.observation_space.flat_dim, ),
output_dim=env_spec.action_space.flat_dim,
hidden_sizes=hidden_sizes,
hidden_nonlinearity=hidden_nonlinearity,
output_nonlinearity=output_nonlinearity,
# batch_normalization=True,
name='mlp_prob_network',
)
with tf.name_scope(self._prob_network_name):
prob_network_output = L.get_output(
prob_network.output_layer, deterministic=True)
self._l_prob = prob_network.output_layer
self._l_obs = prob_network.input_layer
self._f_prob = tensor_utils.compile_function(
[prob_network.input_layer.input_var], prob_network_output)
self.prob_network = prob_network
self.name = name
# Note the deterministic=True argument. It makes sure that when getting
# actions from single observations, we do not update params in the
# batch normalization layers.
# TODO: this doesn't currently work properly in the tf version so we
# leave out batch_norm
super(DeterministicMLPPolicy, self).__init__(env_spec)
LayersPowered.__init__(self, [prob_network.output_layer])
def __init__(self,
input_shape,
output_dim,
name="GaussianMLPRegressor",
mean_network=None,
hidden_sizes=(32, 32),
hidden_nonlinearity=tf.nn.tanh,
optimizer=None,
optimizer_args=None,
use_trust_region=True,
max_kl_step=0.01,
learn_std=True,
init_std=1.0,
adaptive_std=False,
std_share_network=False,
std_hidden_sizes=(32, 32),
std_nonlinearity=None,
normalize_inputs=True,
normalize_outputs=True,
subsample_factor=1.0):
"""
:param input_shape: Shape of the input data.
:param output_dim: Dimension of output.
:param hidden_sizes: Number of hidden units of each layer of the mean
network.
:param hidden_nonlinearity: Non-linearity used for each layer of the
mean network.
:param optimizer: Optimizer for minimizing the negative log-likelihood.
:param use_trust_region: Whether to use trust region constraint.
:param max_kl_step: KL divergence constraint for each iteration
:param learn_std: Whether to learn the standard deviations. Only
effective if adaptive_std is False. If adaptive_std is True, this
parameter is ignored, and the weights for the std network are always
earned.
:param adaptive_std: Whether to make the std a function of the states.
:param std_share_network: Whether to use the same network as the mean.
:param std_hidden_sizes: Number of hidden units of each layer of the
std network. Only used if `std_share_network` is False. It defaults to
the same architecture as the mean.
:param std_nonlinearity: Non-linearity used for each layer of the std
network. Only used if `std_share_network` is False. It defaults to the
same non-linearity as the mean.
"""
Parameterized.__init__(self)
Serializable.quick_init(self, locals())
self._mean_network_name = "mean_network"
self._std_network_name = "std_network"
with tf.variable_scope(name):
if optimizer_args is None:
optimizer_args = dict()
if optimizer is None:
if use_trust_region:
optimizer = PenaltyLbfgsOptimizer(**optimizer_args)
else:
optimizer = LbfgsOptimizer(**optimizer_args)
else:
optimizer = optimizer(**optimizer_args)
self._optimizer = optimizer
self._subsample_factor = subsample_factor
if mean_network is None:
if std_share_network:
mean_network = MLP(
name="mean_network",
input_shape=input_shape,
output_dim=2 * output_dim,
hidden_sizes=hidden_sizes,
hidden_nonlinearity=hidden_nonlinearity,
output_nonlinearity=None,
)
l_mean = L.SliceLayer(
mean_network.output_layer,
slice(output_dim),
name="mean_slice",
)
else:
mean_network = MLP(
name="mean_network",
input_shape=input_shape,
output_dim=output_dim,
hidden_sizes=hidden_sizes,
hidden_nonlinearity=hidden_nonlinearity,
output_nonlinearity=None,
)
l_mean = mean_network.output_layer
if adaptive_std:
l_log_std = MLP(
name="log_std_network",
input_shape=input_shape,
input_var=mean_network.input_layer.input_var,
output_dim=output_dim,
hidden_sizes=std_hidden_sizes,
hidden_nonlinearity=std_nonlinearity,
output_nonlinearity=None,
).output_layer
elif std_share_network:
l_log_std = L.SliceLayer(
mean_network.output_layer,
slice(output_dim, 2 * output_dim),
name="log_std_slice",
)
else:
l_log_std = L.ParamLayer(
mean_network.input_layer,
num_units=output_dim,
param=tf.constant_initializer(np.log(init_std)),
name="output_log_std",
trainable=learn_std,
)
LayersPowered.__init__(self, [l_mean, l_log_std])
xs_var = mean_network.input_layer.input_var
ys_var = tf.placeholder(dtype=tf.float32,
name="ys",
shape=(None, output_dim))
old_means_var = tf.placeholder(dtype=tf.float32,
name="ys",
shape=(None, output_dim))
old_log_stds_var = tf.placeholder(dtype=tf.float32,
name="old_log_stds",
shape=(None, output_dim))
x_mean_var = tf.Variable(
np.zeros((1, ) + input_shape, dtype=np.float32),
name="x_mean",
)
x_std_var = tf.Variable(
np.ones((1, ) + input_shape, dtype=np.float32),
name="x_std",
)
y_mean_var = tf.Variable(
np.zeros((1, output_dim), dtype=np.float32),
name="y_mean",
)
y_std_var = tf.Variable(
np.ones((1, output_dim), dtype=np.float32),
name="y_std",
)
normalized_xs_var = (xs_var - x_mean_var) / x_std_var
normalized_ys_var = (ys_var - y_mean_var) / y_std_var
with tf.name_scope(self._mean_network_name,
values=[normalized_xs_var]):
normalized_means_var = L.get_output(
l_mean, {mean_network.input_layer: normalized_xs_var})
with tf.name_scope(self._std_network_name,
values=[normalized_xs_var]):
normalized_log_stds_var = L.get_output(
l_log_std, {mean_network.input_layer: normalized_xs_var})
means_var = normalized_means_var * y_std_var + y_mean_var
log_stds_var = normalized_log_stds_var + tf.log(y_std_var)
normalized_old_means_var = (old_means_var - y_mean_var) / y_std_var
normalized_old_log_stds_var = old_log_stds_var - tf.log(y_std_var)
dist = self._dist = DiagonalGaussian(output_dim)
normalized_dist_info_vars = dict(mean=normalized_means_var,
log_std=normalized_log_stds_var)
mean_kl = tf.reduce_mean(
dist.kl_sym(
dict(mean=normalized_old_means_var,
log_std=normalized_old_log_stds_var),
normalized_dist_info_vars,
))
loss = -tf.reduce_mean(
dist.log_likelihood_sym(normalized_ys_var,
normalized_dist_info_vars))
self._f_predict = tensor_utils.compile_function([xs_var],
means_var)
self._f_pdists = tensor_utils.compile_function(
[xs_var], [means_var, log_stds_var])
self._l_mean = l_mean
self._l_log_std = l_log_std
optimizer_args = dict(
loss=loss,
target=self,
network_outputs=[
normalized_means_var, normalized_log_stds_var
],
)
if use_trust_region:
optimizer_args["leq_constraint"] = (mean_kl, max_kl_step)
optimizer_args["inputs"] = [
xs_var, ys_var, old_means_var, old_log_stds_var
]
else:
optimizer_args["inputs"] = [xs_var, ys_var]
self._optimizer.update_opt(**optimizer_args)
self._use_trust_region = use_trust_region
self._name = name
self._normalize_inputs = normalize_inputs
self._normalize_outputs = normalize_outputs
self._mean_network = mean_network
self._x_mean_var = x_mean_var
self._x_std_var = x_std_var
self._y_mean_var = y_mean_var
self._y_std_var = y_std_var
# Optionally create assign operations for normalization
if self._normalize_inputs:
self._x_mean_var_ph = tf.placeholder(
shape=(1, ) + input_shape,
dtype=tf.float32,
)
self._x_std_var_ph = tf.placeholder(
shape=(1, ) + input_shape,
dtype=tf.float32,
)
self._assign_x_mean = tf.assign(self._x_mean_var,
self._x_mean_var_ph)
self._assign_x_std = tf.assign(self._x_std_var,
self._x_std_var_ph)
if self._normalize_outputs:
self._y_mean_var_ph = tf.placeholder(
shape=(1, output_dim),
dtype=tf.float32,
)
self._y_std_var_ph = tf.placeholder(
shape=(1, output_dim),
dtype=tf.float32,
)
self._assign_y_mean = tf.assign(self._y_mean_var,
self._y_mean_var_ph)
self._assign_y_std = tf.assign(self._y_std_var,
self._y_std_var_ph)
def __init__(
self,
input_shape,
output_dim,
name='CategoricalMLPRegressor',
prob_network=None,
hidden_sizes=(32, 32),
hidden_nonlinearity=tf.nn.tanh,
optimizer=None,
tr_optimizer=None,
use_trust_region=True,
max_kl_step=0.01,
normalize_inputs=True,
no_initial_trust_region=True,
):
"""
:param input_shape: Shape of the input data.
:param output_dim: Dimension of output.
:param hidden_sizes: Number of hidden units of each layer of the mean
network.
:param hidden_nonlinearity: Non-linearity used for each layer of the
mean network.
:param optimizer: Optimizer for minimizing the negative log-likelihood.
:param use_trust_region: Whether to use trust region constraint.
:param max_kl_step: KL divergence constraint for each iteration
"""
Parameterized.__init__(self)
Serializable.quick_init(self, locals())
with tf.compat.v1.variable_scope(name, 'CategoricalMLPRegressor'):
if optimizer is None:
optimizer = LbfgsOptimizer()
if tr_optimizer is None:
tr_optimizer = ConjugateGradientOptimizer()
self.output_dim = output_dim
self.optimizer = optimizer
self.tr_optimizer = tr_optimizer
self._prob_network_name = 'prob_network'
if prob_network is None:
prob_network = MLP(input_shape=input_shape,
output_dim=output_dim,
hidden_sizes=hidden_sizes,
hidden_nonlinearity=hidden_nonlinearity,
output_nonlinearity=tf.nn.softmax,
name=self._prob_network_name)
l_prob = prob_network.output_layer
LayersPowered.__init__(self, [l_prob])
xs_var = prob_network.input_layer.input_var
ys_var = tf.compat.v1.placeholder(dtype=tf.float32,
shape=[None, output_dim],
name='ys')
old_prob_var = tf.compat.v1.placeholder(dtype=tf.float32,
shape=[None, output_dim],
name='old_prob')
x_mean_var = tf.compat.v1.get_variable(
name='x_mean',
shape=(1, ) + input_shape,
initializer=tf.constant_initializer(0., dtype=tf.float32))
x_std_var = tf.compat.v1.get_variable(
name='x_std',
shape=(1, ) + input_shape,
initializer=tf.constant_initializer(1., dtype=tf.float32))
normalized_xs_var = (xs_var - x_mean_var) / x_std_var
with tf.name_scope(self._prob_network_name,
values=[normalized_xs_var]):
prob_var = L.get_output(
l_prob, {prob_network.input_layer: normalized_xs_var})
old_info_vars = dict(prob=old_prob_var)
info_vars = dict(prob=prob_var)
dist = self._dist = Categorical(output_dim)
mean_kl = tf.reduce_mean(dist.kl_sym(old_info_vars, info_vars))
loss = -tf.reduce_mean(dist.log_likelihood_sym(ys_var, info_vars))
predicted = tf.one_hot(tf.argmax(prob_var, axis=1),
depth=output_dim)
self.prob_network = prob_network
self.f_predict = tensor_utils.compile_function([xs_var], predicted)
self.f_prob = tensor_utils.compile_function([xs_var], prob_var)
self.l_prob = l_prob
self.optimizer.update_opt(loss=loss,
target=self,
network_outputs=[prob_var],
inputs=[xs_var, ys_var])
self.tr_optimizer.update_opt(loss=loss,
target=self,
network_outputs=[prob_var],
inputs=[xs_var, ys_var, old_prob_var],
leq_constraint=(mean_kl, max_kl_step))
self.use_trust_region = use_trust_region
self.name = name
self.normalize_inputs = normalize_inputs
self.x_mean_var = x_mean_var
self.x_std_var = x_std_var
self.first_optimized = not no_initial_trust_region
def __init__(
self,
input_shape,
output_dim,
name="BernoulliMLPRegressor",
hidden_sizes=(32, 32),
hidden_nonlinearity=tf.nn.relu,
optimizer=None,
tr_optimizer=None,
use_trust_region=True,
step_size=0.01,
normalize_inputs=True,
no_initial_trust_region=True,
):
"""
:param input_shape: Shape of the input data.
:param output_dim: Dimension of output.
:param hidden_sizes: Number of hidden units of each layer of the mean
network.
:param hidden_nonlinearity: Non-linearity used for each layer of the
mean network.
:param optimizer: Optimizer for minimizing the negative log-likelihood.
:param use_trust_region: Whether to use trust region constraint.
:param step_size: KL divergence constraint for each iteration
"""
Parameterized.__init__(self)
Serializable.quick_init(self, locals())
with tf.variable_scope(name):
if optimizer is None:
optimizer = LbfgsOptimizer()
if tr_optimizer is None:
tr_optimizer = ConjugateGradientOptimizer()
self.output_dim = output_dim
self.optimizer = optimizer
self.tr_optimizer = tr_optimizer
p_network = MLP(input_shape=input_shape,
output_dim=output_dim,
hidden_sizes=hidden_sizes,
hidden_nonlinearity=hidden_nonlinearity,
output_nonlinearity=tf.nn.sigmoid,
name="p_network")
l_p = p_network.output_layer
LayersPowered.__init__(self, [l_p])
xs_var = p_network.input_layer.input_var
ys_var = tf.placeholder(dtype=tf.float32,
shape=(None, output_dim),
name="ys")
old_p_var = tf.placeholder(dtype=tf.float32,
shape=(None, output_dim),
name="old_p")
x_mean_var = tf.get_variable(name="x_mean",
initializer=tf.zeros_initializer(),
shape=(1, ) + input_shape)
x_std_var = tf.get_variable(name="x_std",
initializer=tf.ones_initializer(),
shape=(1, ) + input_shape)
normalized_xs_var = (xs_var - x_mean_var) / x_std_var
p_var = L.get_output(l_p,
{p_network.input_layer: normalized_xs_var})
old_info_vars = dict(p=old_p_var)
info_vars = dict(p=p_var)
dist = self._dist = Bernoulli(output_dim)
mean_kl = tf.reduce_mean(dist.kl_sym(old_info_vars, info_vars))
loss = -tf.reduce_mean(dist.log_likelihood_sym(ys_var, info_vars))
predicted = p_var >= 0.5
self.f_predict = tensor_utils.compile_function([xs_var], predicted)
self.f_p = tensor_utils.compile_function([xs_var], p_var)
self.l_p = l_p
self.optimizer.update_opt(loss=loss,
target=self,
network_outputs=[p_var],
inputs=[xs_var, ys_var])
self.tr_optimizer.update_opt(loss=loss,
target=self,
network_outputs=[p_var],
inputs=[xs_var, ys_var, old_p_var],
leq_constraint=(mean_kl, step_size))
self.use_trust_region = use_trust_region
self.name = name
self.normalize_inputs = normalize_inputs
self.x_mean_var = x_mean_var
self.x_std_var = x_std_var
self.first_optimized = not no_initial_trust_region
def __init__(
self,
input_shape,
output_dim,
name='ContinuousMLPRegressor',
network=None,
hidden_sizes=(32, 32),
hidden_nonlinearity=tf.nn.tanh,
output_nonlinearity=None,
optimizer=None,
optimizer_args=None,
normalize_inputs=True,
):
"""
:param input_shape: Shape of the input data.
:param output_dim: Dimension of output.
:param hidden_sizes: Number of hidden units of each layer of the
mean network.
:param hidden_nonlinearity: Non-linearity used for each layer of the
mean network.
:param optimizer: Optimizer for minimizing the negative log-likelihood.
"""
Parameterized.__init__(self)
Serializable.quick_init(self, locals())
with tf.compat.v1.variable_scope(name, 'ContinuousMLPRegressor'):
if optimizer_args is None:
optimizer_args = dict()
if optimizer is None:
optimizer = LbfgsOptimizer(**optimizer_args)
else:
optimizer = optimizer(**optimizer_args)
self.output_dim = output_dim
self.optimizer = optimizer
self._network_name = 'network'
if network is None:
network = MLP(input_shape=input_shape,
output_dim=output_dim,
hidden_sizes=hidden_sizes,
hidden_nonlinearity=hidden_nonlinearity,
output_nonlinearity=output_nonlinearity,
name=self._network_name)
l_out = network.output_layer
LayersPowered.__init__(self, [l_out])
xs_var = network.input_layer.input_var
ys_var = tf.compat.v1.placeholder(dtype=tf.float32,
shape=[None, output_dim],
name='ys')
x_mean_var = tf.compat.v1.get_variable(
name='x_mean',
shape=(1, ) + input_shape,
initializer=tf.constant_initializer(0., dtype=tf.float32))
x_std_var = tf.compat.v1.get_variable(
name='x_std',
shape=(1, ) + input_shape,
initializer=tf.constant_initializer(1., dtype=tf.float32))
normalized_xs_var = (xs_var - x_mean_var) / x_std_var
with tf.name_scope(self._network_name, values=[normalized_xs_var]):
fit_ys_var = L.get_output(
l_out, {network.input_layer: normalized_xs_var})
loss = -tf.reduce_mean(tf.square(fit_ys_var - ys_var))
self.f_predict = tensor_utils.compile_function([xs_var],
fit_ys_var)
optimizer_args = dict(
loss=loss,
target=self,
network_outputs=[fit_ys_var],
)
optimizer_args['inputs'] = [xs_var, ys_var]
self.optimizer.update_opt(**optimizer_args)
self.name = name
self.l_out = l_out
self.normalize_inputs = normalize_inputs
self.x_mean_var = x_mean_var
self.x_std_var = x_std_var
def __init__(self,
env_spec,
name='GaussianMLPPolicy',
hidden_sizes=(32, 32),
learn_std=True,
init_std=1.0,
adaptive_std=False,
std_share_network=False,
std_hidden_sizes=(32, 32),
min_std=1e-6,
std_hidden_nonlinearity=tf.nn.tanh,
hidden_nonlinearity=tf.nn.tanh,
output_nonlinearity=None,
mean_network=None,
std_network=None,
std_parametrization='exp'):
"""
:param env_spec:
:param hidden_sizes: list of sizes for the fully-connected hidden
layers
:param learn_std: Is std trainable
:param init_std: Initial std
:param adaptive_std:
:param std_share_network:
:param std_hidden_sizes: list of sizes for the fully-connected layers
for std
:param min_std: whether to make sure that the std is at least some
threshold value, to avoid numerical issues
:param std_hidden_nonlinearity:
:param hidden_nonlinearity: nonlinearity used for each hidden layer
:param output_nonlinearity: nonlinearity for the output layer
:param mean_network: custom network for the output mean
:param std_network: custom network for the output log std
:param std_parametrization: how the std should be parametrized. There
are a few options:
- exp: the logarithm of the std will be stored, and applied a
exponential transformation
- softplus: the std will be computed as log(1+exp(x))
:return:
"""
assert isinstance(env_spec.action_space, Box)
Serializable.quick_init(self, locals())
self.name = name
self._mean_network_name = 'mean_network'
self._std_network_name = 'std_network'
with tf.variable_scope(name, 'GaussianMLPPolicy'):
obs_dim = env_spec.observation_space.flat_dim
action_dim = env_spec.action_space.flat_dim
# create network
if mean_network is None:
if std_share_network:
if std_parametrization == 'exp':
init_std_param = np.log(init_std)
elif std_parametrization == 'softplus':
init_std_param = np.log(np.exp(init_std) - 1)
else:
raise NotImplementedError
b = np.concatenate((np.zeros(action_dim),
np.full(action_dim, init_std_param)),
axis=0)
b = tf.constant_initializer(b)
with tf.variable_scope(self._mean_network_name):
mean_network = MLP(
name='mlp',
input_shape=(obs_dim, ),
output_dim=2 * action_dim,
hidden_sizes=hidden_sizes,
hidden_nonlinearity=hidden_nonlinearity,
output_nonlinearity=output_nonlinearity,
output_b_init=b,
)
l_mean = L.SliceLayer(
mean_network.output_layer,
slice(action_dim),
name='mean_slice',
)
else:
mean_network = MLP(
name=self._mean_network_name,
input_shape=(obs_dim, ),
output_dim=action_dim,
hidden_sizes=hidden_sizes,
hidden_nonlinearity=hidden_nonlinearity,
output_nonlinearity=output_nonlinearity,
)
l_mean = mean_network.output_layer
self._mean_network = mean_network
obs_var = mean_network.input_layer.input_var
if std_network is not None:
l_std_param = std_network.output_layer
else:
if std_parametrization == 'exp':
init_std_param = np.log(init_std)
elif std_parametrization == 'softplus':
init_std_param = np.log(np.exp(init_std) - 1)
else:
raise ValueError('Invalid argument for std_parametrization'
': {}'.format(std_parametrization))
if adaptive_std:
b = tf.constant_initializer(init_std_param)
std_network = MLP(
name=self._std_network_name,
input_shape=(obs_dim, ),
input_layer=mean_network.input_layer,
output_dim=action_dim,
hidden_sizes=std_hidden_sizes,
hidden_nonlinearity=std_hidden_nonlinearity,
output_nonlinearity=None,
output_b_init=b,
)
l_std_param = std_network.output_layer
elif std_share_network:
with tf.variable_scope(self._std_network_name):
l_std_param = L.SliceLayer(
mean_network.output_layer,
slice(action_dim, 2 * action_dim),
name='std_slice',
)
else:
with tf.variable_scope(self._std_network_name):
l_std_param = L.ParamLayer(
mean_network.input_layer,
num_units=action_dim,
param=tf.constant_initializer(init_std_param),
name='output_std_param',
trainable=learn_std,
)
self.std_parametrization = std_parametrization
if std_parametrization == 'exp':
min_std_param = np.log(min_std)
elif std_parametrization == 'softplus':
min_std_param = np.log(np.exp(min_std) - 1)
else:
raise NotImplementedError
self.min_std_param = min_std_param
# mean_var, log_std_var = L.get_output([l_mean, l_std_param])
#
# if self.min_std_param is not None:
# log_std_var = tf.maximum(log_std_var, np.log(min_std))
#
# self._mean_var, self._log_std_var = mean_var, log_std_var
self._l_mean = l_mean
self._l_std_param = l_std_param
self._dist = DiagonalGaussian(action_dim)
LayersPowered.__init__(self, [l_mean, l_std_param])
super(GaussianMLPPolicy, self).__init__(env_spec)
dist_info_sym = self.dist_info_sym(
mean_network.input_layer.input_var, dict())
mean_var = tf.identity(dist_info_sym['mean'], name='mean')
log_std_var = tf.identity(
dist_info_sym['log_std'], name='standard_dev')
self._f_dist = tensor_utils.compile_function(
inputs=[obs_var],
outputs=[mean_var, log_std_var],
)
