def _setup_stochastic_policy(self, obs, action, reuse=False, scope="pi"):
"""Create the variables of a stochastic policy.
Parameters
----------
obs : tf.compat.v1.placeholder
the input observation placeholder
action : tf.compat.v1.placeholder
the input action placeholder
reuse : bool
whether or not to reuse parameters
scope : str
the scope name of the policy
"""
with tf.compat.v1.variable_scope(scope, reuse=reuse):
pi_h = obs
# create the hidden layers
for i, layer_size in enumerate(self.layers):
pi_h = layer(pi_h,
layer_size,
'fc{}'.format(i),
act_fun=self.act_fun,
layer_norm=self.layer_norm)
# create the output mean
policy_mean = layer(
pi_h,
self.ac_space.shape[0],
'mean',
act_fun=None,
kernel_initializer=tf.random_uniform_initializer(minval=-3e-3,
maxval=3e-3))
# create the output log_std
log_std = layer(
pi_h,
self.ac_space.shape[0],
'log_std',
act_fun=None,
)
# OpenAI Variation to cap the standard deviation
log_std = tf.clip_by_value(log_std, LOG_STD_MIN, LOG_STD_MAX)
std = tf.exp(log_std)
# Reparameterization trick
policy = policy_mean + tf.random.normal(tf.shape(policy_mean)) * std
logp_pi = gaussian_likelihood(policy, policy_mean, log_std)
logp_ac = gaussian_likelihood(action, policy_mean, log_std)
# Apply squashing and account for it in the probability
_, _, logp_ac = apply_squashing_func(policy_mean, action, logp_ac)
_, policy, _ = apply_squashing_func(policy_mean, policy, logp_pi)
# Store the variables under their respective parameters.
self.policy = policy
self.logp_ac = logp_ac
def test_apply_squashing(self):
"""Check the functionality of the apply_squashing() method."""
# Some inputs
mu_ = tf.constant([[0, 0.5, 1, 2]], dtype=tf.float32)
pi_ = tf.constant([[0, 0.5, 1, 2]], dtype=tf.float32)
logp_pi = tf.constant([[0, 0.5, 1, 2]], dtype=tf.float32)
# Run the function.
det_policy, policy, logp_pi = apply_squashing_func(mu_, pi_, logp_pi)
# Initialize everything.
sess = tf.compat.v1.Session()
sess.run(tf.compat.v1.global_variables_initializer())
# Test the output from the deterministic squashed output.
np.testing.assert_almost_equal(
sess.run(det_policy), [[0., 0.4621172, 0.7615942, 0.9640276]])
# Clear the graph.
tf.compat.v1.reset_default_graph()
def make_actor(self, obs, action, reuse=False, scope="pi"):
"""Create the actor variables.
Parameters
----------
obs : tf.compat.v1.placeholder
the input observation placeholder
action : tf.compat.v1.placeholder
the input action placeholder
reuse : bool
whether or not to reuse parameters
scope : str
the scope name of the actor
Returns
-------
tf.Variable
the output from the deterministic actor
tf.Variable
the output from the stochastic actor
tf.Variable
the log-probability of a given observation given the output action
from the policy
tf.Variable
the log-probability of a given observation given a fixed action
"""
# Initial image pre-processing (for convolutional policies).
if self.model_params["model_type"] == "conv":
pi_h = create_conv(
obs=obs,
image_height=self.model_params["image_height"],
image_width=self.model_params["image_width"],
image_channels=self.model_params["image_channels"],
ignore_flat_channels=self.model_params["ignore_flat_channels"],
ignore_image=self.model_params["ignore_image"],
filters=self.model_params["filters"],
kernel_sizes=self.model_params["kernel_sizes"],
strides=self.model_params["strides"],
act_fun=self.model_params["act_fun"],
layer_norm=self.model_params["layer_norm"],
scope=scope,
reuse=reuse,
)
else:
pi_h = obs
# Create the model.
policy_mean, log_std = create_fcnet(
obs=pi_h,
layers=self.model_params["layers"],
num_output=self.ac_space.shape[0],
stochastic=True,
act_fun=self.model_params["act_fun"],
layer_norm=self.model_params["layer_norm"],
scope=scope,
reuse=reuse,
)
# OpenAI Variation to cap the standard deviation
log_std = tf.clip_by_value(log_std, LOG_STD_MIN, LOG_STD_MAX)
std = tf.exp(log_std)
# Reparameterization trick
policy = policy_mean + tf.random.normal(tf.shape(policy_mean)) * std
logp_pi = gaussian_likelihood(policy, policy_mean, log_std)
logp_ac = gaussian_likelihood(action, policy_mean, log_std)
# Apply squashing and account for it in the probability
_, _, logp_ac = apply_squashing_func(policy_mean, action, logp_ac)
deterministic_policy, policy, logp_pi = apply_squashing_func(
policy_mean, policy, logp_pi)
return deterministic_policy, policy, logp_pi, logp_ac
def make_actor(self, obs, action, reuse=False, scope="pi"):
"""Create the actor variables.
Parameters
----------
obs : tf.compat.v1.placeholder
the input observation placeholder
action : tf.compat.v1.placeholder
the input action placeholder
reuse : bool
whether or not to reuse parameters
scope : str
the scope name of the actor
Returns
-------
tf.Variable
the output from the deterministic actor
tf.Variable
the output from the stochastic actor
tf.Variable
the log-probability of a given observation given the output action
from the policy
tf.Variable
the log-probability of a given observation given a fixed action
"""
with tf.compat.v1.variable_scope(scope, reuse=reuse):
pi_h = obs
# zero out the fingerprint observations for the worker policy
if self.zero_fingerprint:
pi_h = self._remove_fingerprint(
pi_h,
self.ob_space.shape[0],
self.fingerprint_dim,
self.co_space.shape[0]
)
# create the hidden layers
for i, layer_size in enumerate(self.layers):
pi_h = layer(
pi_h, layer_size, 'fc{}'.format(i),
act_fun=self.act_fun,
layer_norm=self.layer_norm
)
# create the output mean
policy_mean = layer(
pi_h, self.ac_space.shape[0], 'mean',
act_fun=None,
kernel_initializer=tf.random_uniform_initializer(
minval=-3e-3, maxval=3e-3)
)
# create the output log_std
log_std = layer(
pi_h, self.ac_space.shape[0], 'log_std',
act_fun=None,
)
# OpenAI Variation to cap the standard deviation
log_std = tf.clip_by_value(log_std, LOG_STD_MIN, LOG_STD_MAX)
std = tf.exp(log_std)
# Reparameterization trick
policy = policy_mean + tf.random.normal(tf.shape(policy_mean)) * std
logp_pi = gaussian_likelihood(policy, policy_mean, log_std)
logp_ac = gaussian_likelihood(action, policy_mean, log_std)
# Apply squashing and account for it in the probability
_, _, logp_ac = apply_squashing_func(
policy_mean, action, logp_ac)
deterministic_policy, policy, logp_pi = apply_squashing_func(
policy_mean, policy, logp_pi)
return deterministic_policy, policy, logp_pi, logp_ac