def create_forward_model(
self, encoded_state: tf.Tensor, encoded_next_state: tf.Tensor
) -> None:
"""
Creates forward model TensorFlow ops for Curiosity module.
Predicts encoded future state based on encoded current state and given action.
:param encoded_state: Tensor corresponding to encoded current state.
:param encoded_next_state: Tensor corresponding to encoded next state.
"""
combined_input = tf.concat(
[encoded_state, self.policy.selected_actions], axis=1
)
hidden = tf.layers.dense(combined_input, 256, activation=ModelUtils.swish)
pred_next_state = tf.layers.dense(
hidden,
self.encoding_size
* (self.policy.vis_obs_size + int(self.policy.vec_obs_size > 0)),
activation=None,
)
squared_difference = 0.5 * tf.reduce_sum(
tf.squared_difference(pred_next_state, encoded_next_state), axis=1
)
self.intrinsic_reward = squared_difference
self.forward_loss = tf.reduce_mean(
tf.dynamic_partition(squared_difference, self.policy.mask, 2)[1]
)
def create_inverse_model(self, encoded_state: tf.Tensor,
encoded_next_state: tf.Tensor) -> None:
"""
Creates inverse model TensorFlow ops for Curiosity module.
Predicts action taken given current and future encoded states.
:param encoded_state: Tensor corresponding to encoded current state.
:param encoded_next_state: Tensor corresponding to encoded next state.
"""
combined_input = tf.concat([encoded_state, encoded_next_state], axis=1)
hidden = tf.layers.dense(combined_input,
256,
activation=LearningModel.swish)
if self.policy_model.brain.vector_action_space_type == "continuous":
pred_action = tf.layers.dense(hidden,
self.policy_model.act_size[0],
activation=None)
squared_difference = tf.reduce_sum(
tf.squared_difference(pred_action,
self.policy_model.selected_actions),
axis=1,
)
self.inverse_loss = tf.reduce_mean(
tf.dynamic_partition(squared_difference,
self.policy_model.mask, 2)[1])
else:
pred_action = tf.concat(
[
tf.layers.dense(hidden,
self.policy_model.act_size[i],
activation=tf.nn.softmax)
for i in range(len(self.policy_model.act_size))
],
axis=1,
)
cross_entropy = tf.reduce_sum(
-tf.log(pred_action + 1e-10) *
self.policy_model.selected_actions,
axis=1,
)
self.inverse_loss = tf.reduce_mean(
tf.dynamic_partition(cross_entropy, self.policy_model.mask,
2)[1])
def _create_losses(self, probs, old_probs, value_heads, entropy, beta,
epsilon, lr, max_step):
"""
Creates training-specific Tensorflow ops for PPO models.
:param probs: Current policy probabilities
:param old_probs: Past policy probabilities
:param value_heads: Value estimate tensors from each value stream
:param beta: Entropy regularization strength
:param entropy: Current policy entropy
:param epsilon: Value for policy-divergence threshold
:param lr: Learning rate
:param max_step: Total number of training steps.
"""
self.returns_holders = {}
self.old_values = {}
for name in value_heads.keys():
returns_holder = tf.placeholder(shape=[None],
dtype=tf.float32,
name="{}_returns".format(name))
old_value = tf.placeholder(shape=[None],
dtype=tf.float32,
name="{}_value_estimate".format(name))
self.returns_holders[name] = returns_holder
self.old_values[name] = old_value
self.advantage = tf.placeholder(shape=[None],
dtype=tf.float32,
name="advantages")
advantage = tf.expand_dims(self.advantage, -1)
decay_epsilon = tf.train.polynomial_decay(epsilon,
self.policy.global_step,
max_step,
0.1,
power=1.0)
decay_beta = tf.train.polynomial_decay(beta,
self.policy.global_step,
max_step,
1e-5,
power=1.0)
value_losses = []
for name, head in value_heads.items():
clipped_value_estimate = self.old_values[name] + tf.clip_by_value(
tf.reduce_sum(head, axis=1) - self.old_values[name],
-decay_epsilon,
decay_epsilon,
)
v_opt_a = tf.squared_difference(self.returns_holders[name],
tf.reduce_sum(head, axis=1))
v_opt_b = tf.squared_difference(self.returns_holders[name],
clipped_value_estimate)
value_loss = tf.reduce_mean(
tf.dynamic_partition(tf.maximum(v_opt_a, v_opt_b),
self.policy.mask, 2)[1])
value_losses.append(value_loss)
self.value_loss = tf.reduce_mean(value_losses)
r_theta = tf.exp(probs - old_probs)
p_opt_a = r_theta * advantage
p_opt_b = (tf.clip_by_value(r_theta, 1.0 - decay_epsilon,
1.0 + decay_epsilon) * advantage)
self.policy_loss = -tf.reduce_mean(
tf.dynamic_partition(tf.minimum(p_opt_a, p_opt_b),
self.policy.mask, 2)[1])
# For cleaner stats reporting
self.abs_policy_loss = tf.abs(self.policy_loss)
self.loss = (
self.policy_loss + 0.5 * self.value_loss -
decay_beta * tf.reduce_mean(
tf.dynamic_partition(entropy, self.policy.mask, 2)[1]))
