def _create_entropy(
self, encoded: "GaussianDistribution.MuSigmaTensors") -> tf.Tensor:
single_dim_entropy = 0.5 * tf.reduce_mean(
tf.log(2 * np.pi * np.e) + tf.square(encoded.log_sigma))
# Make entropy the right shape
return tf.ones_like(tf.reshape(encoded.mu[:, 0],
[-1])) * single_dim_entropy
def create_loss(self, learning_rate: float) -> None:
"""
Creates the loss and update nodes for the GAIL reward generator
:param learning_rate: The learning rate for the optimizer
"""
self.mean_expert_estimate = tf.reduce_mean(self.expert_estimate)
self.mean_policy_estimate = tf.reduce_mean(self.policy_estimate)
if self.use_vail:
self.beta = tf.get_variable(
"gail_beta",
[],
trainable=False,
dtype=tf.float32,
initializer=tf.ones_initializer(),
)
self.discriminator_loss = -tf.reduce_mean(
tf.log(self.expert_estimate + EPSILON)
+ tf.log(1.0 - self.policy_estimate + EPSILON)
)
if self.use_vail:
# KL divergence loss (encourage latent representation to be normal)
self.kl_loss = tf.reduce_mean(
-tf.reduce_sum(
1
+ self.z_log_sigma_sq
- 0.5 * tf.square(self.z_mean_expert)
- 0.5 * tf.square(self.z_mean_policy)
- tf.exp(self.z_log_sigma_sq),
1,
)
)
self.loss = (
self.beta * (self.kl_loss - self.mutual_information)
+ self.discriminator_loss
)
else:
self.loss = self.discriminator_loss
if self.gradient_penalty_weight > 0.0:
self.loss += self.gradient_penalty_weight * self.create_gradient_magnitude()
optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate)
self.update_batch = optimizer.minimize(self.loss)
def create_cc_actor_critic(self, h_size: int, num_layers: int,
vis_encode_type: EncoderType) -> None:
"""
Creates Continuous control actor-critic model.
:param h_size: Size of hidden linear layers.
:param num_layers: Number of hidden linear layers.
"""
hidden_streams = self.create_observation_streams(
2, h_size, num_layers, vis_encode_type)
if self.use_recurrent:
self.memory_in = tf.placeholder(shape=[None, self.m_size],
dtype=tf.float32,
name="recurrent_in")
_half_point = int(self.m_size / 2)
hidden_policy, memory_policy_out = self.create_recurrent_encoder(
hidden_streams[0],
self.memory_in[:, :_half_point],
self.sequence_length,
name="lstm_policy",
)
hidden_value, memory_value_out = self.create_recurrent_encoder(
hidden_streams[1],
self.memory_in[:, _half_point:],
self.sequence_length,
name="lstm_value",
)
self.memory_out = tf.concat([memory_policy_out, memory_value_out],
axis=1,
name="recurrent_out")
else:
hidden_policy = hidden_streams[0]
hidden_value = hidden_streams[1]
mu = tf.layers.dense(
hidden_policy,
self.act_size[0],
activation=None,
kernel_initializer=LearningModel.scaled_init(0.01),
reuse=tf.AUTO_REUSE,
)
self.log_sigma_sq = tf.get_variable(
"log_sigma_squared",
[self.act_size[0]],
dtype=tf.float32,
initializer=tf.zeros_initializer(),
)
sigma_sq = tf.exp(self.log_sigma_sq)
self.epsilon = tf.placeholder(shape=[None, self.act_size[0]],
dtype=tf.float32,
name="epsilon")
# Clip and scale output to ensure actions are always within [-1, 1] range.
self.output_pre = mu + tf.sqrt(sigma_sq) * self.epsilon
output_post = tf.clip_by_value(self.output_pre, -3, 3) / 3
self.output = tf.identity(output_post, name="action")
self.selected_actions = tf.stop_gradient(output_post)
# Compute probability of model output.
all_probs = (-0.5 * tf.square(tf.stop_gradient(self.output_pre) - mu) /
sigma_sq - 0.5 * tf.log(2.0 * np.pi) -
0.5 * self.log_sigma_sq)
self.all_log_probs = tf.identity(all_probs, name="action_probs")
self.entropy = 0.5 * tf.reduce_mean(
tf.log(2 * np.pi * np.e) + self.log_sigma_sq)
self.create_value_heads(self.stream_names, hidden_value)
self.all_old_log_probs = tf.placeholder(shape=[None, self.act_size[0]],
dtype=tf.float32,
name="old_probabilities")
# We keep these tensors the same name, but use new nodes to keep code parallelism with discrete control.
self.log_probs = tf.reduce_sum((tf.identity(self.all_log_probs)),
axis=1,
keepdims=True)
self.old_log_probs = tf.reduce_sum(
(tf.identity(self.all_old_log_probs)), axis=1, keepdims=True)
