def _get_masked_actions_probs(
self,
unmasked_log_probs: List[tf.Tensor],
act_size: List[int],
action_masks: tf.Tensor,
) -> Tuple[tf.Tensor, tf.Tensor, np.ndarray]:
output, _, all_log_probs = ModelUtils.create_discrete_action_masking_layer(
unmasked_log_probs, action_masks, act_size)
action_idx = [0] + list(np.cumsum(act_size))
return output, all_log_probs, action_idx
def _create_dc_critic(self, h_size: int, num_layers: int,
vis_encode_type: EncoderType) -> None:
"""
Creates Discrete control critic (value) network.
:param h_size: Size of hidden linear layers.
:param num_layers: Number of hidden linear layers.
:param vis_encode_type: The type of visual encoder to use.
"""
hidden_stream = ModelUtils.create_observation_streams(
self.policy.visual_in,
self.policy.processed_vector_in,
1,
h_size,
num_layers,
vis_encode_type,
)[0]
if self.policy.use_recurrent:
hidden_value, memory_value_out = ModelUtils.create_recurrent_encoder(
hidden_stream,
self.memory_in,
self.policy.sequence_length_ph,
name="lstm_value",
)
self.memory_out = memory_value_out
else:
hidden_value = hidden_stream
self.value_heads, self.value = ModelUtils.create_value_heads(
self.stream_names, hidden_value)
self.all_old_log_probs = tf.placeholder(
shape=[None, sum(self.policy.act_size)],
dtype=tf.float32,
name="old_probabilities",
)
# Break old log log_probs into separate branches
old_log_prob_branches = ModelUtils.break_into_branches(
self.all_old_log_probs, self.policy.act_size)
_, _, old_normalized_logits = ModelUtils.create_discrete_action_masking_layer(
old_log_prob_branches, self.policy.action_masks,
self.policy.act_size)
action_idx = [0] + list(np.cumsum(self.policy.act_size))
self.old_log_probs = tf.reduce_sum(
(tf.stack(
[
-tf.nn.softmax_cross_entropy_with_logits_v2(
labels=self.policy.
selected_actions[:, action_idx[i]:action_idx[i + 1]],
logits=old_normalized_logits[:, action_idx[i]:
action_idx[i + 1]],
) for i in range(len(self.policy.act_size))
],
axis=1,
)),
axis=1,
keepdims=True,
)