def _create_merlin_algorithm(
encoder_fc_layers=(3, ),
latent_dim=3,
lstm_size=(4, ),
memory_size=20,
learning_rate=1e-3,
debug_summaries=True):
observation_spec = common.get_observation_spec()
action_spec = common.get_action_spec()
algorithm = MerlinAlgorithm(
observation_spec=observation_spec,
action_spec=action_spec,
encoders=EncodingNetwork(
input_tensor_spec=observation_spec,
fc_layer_params=encoder_fc_layers,
activation_fn=None,
name="ObsEncoder"),
decoders=DecodingAlgorithm(
decoder=EncodingNetwork(
input_tensor_spec=TensorSpec((latent_dim, ), dtype=tf.float32),
fc_layer_params=encoder_fc_layers,
activation_fn=None,
name="ObsDecoder"),
loss_weight=100.),
latent_dim=latent_dim,
lstm_size=lstm_size,
memory_size=memory_size,
optimizer=tf.optimizers.Adam(learning_rate=learning_rate),
debug_summaries=debug_summaries)
return algorithm
def _create_ddpg_algorithm():
observation_spec = common.get_observation_spec()
action_spec = common.get_action_spec()
actor_net = ActorNetwork(
observation_spec, action_spec, fc_layer_params=(16, 16))
critic_net = CriticNetwork((observation_spec, action_spec),
joint_fc_layer_params=(16, 16))
return DdpgAlgorithm(
observation_spec=observation_spec,
action_spec=action_spec,
actor_network=actor_net,
critic_network=critic_net,
actor_optimizer=tf.optimizers.Adam(learning_rate=5e-3),
critic_optimizer=tf.optimizers.Adam(learning_rate=1e-1),
debug_summaries=True)
def _create_ac_algorithm():
observation_spec = common.get_observation_spec()
action_spec = common.get_action_spec()
optimizer = tf.optimizers.Adam(learning_rate=5e-5)
actor_net = ActorDistributionNetwork(observation_spec,
action_spec,
fc_layer_params=(8, ))
value_net = ValueNetwork(observation_spec, fc_layer_params=(8, ))
return ActorCriticAlgorithm(action_spec=action_spec,
actor_network=actor_net,
value_network=value_net,
loss_class=ActorCriticLoss,
optimizer=optimizer,
debug_summaries=True)
def _create_ppo_algorithm():
observation_spec = common.get_observation_spec()
action_spec = common.get_action_spec()
optimizer = tf.optimizers.Adam(learning_rate=1e-3)
actor_net = ActorDistributionRnnNetwork(observation_spec,
action_spec,
input_fc_layer_params=(),
output_fc_layer_params=None)
value_net = ValueRnnNetwork(observation_spec,
input_fc_layer_params=(),
output_fc_layer_params=None)
return PPOAlgorithm(action_spec=action_spec,
actor_network=actor_net,
value_network=value_net,
loss_class=PPOLoss,
optimizer=optimizer,
debug_summaries=True)
def get_ac_networks(conv_layer_params=None,
num_embedding_dims=None,
fc_layer_params=None,
num_state_tiles=None,
num_sentence_tiles=None):
"""
Generate the actor and value networks
Args:
conv_layer_params (list[int 3 tuple]): optional convolution layers
parameters, where each item is a length-three tuple indicating
(filters, kernel_size, stride).
num_embedding_dims (int): optional number of dimensions of the
vocabulary embedding space.
fc_layer_params (list[int]): optional fully_connected parameters, where
each item is the number of units in the layer.
num_state_tiles (int): optional number of times to repeat the
internal state tensor before concatenation with other inputs.
The rationale is to match the number of dimentions of the image
input, so that the final concatenation will have roughly equal
representation from different sources of input. Without this,
typically image input, due to its large input size, will take over
and trump all other small dimensional inputs.
num_sentence_tiles (int): optional number of times to repeat the
sentence embedding tensor before concatenation with other inputs,
so that sentence input won't be trumped by other high dimensional
inputs like image observation.
"""
observation_spec = common.get_observation_spec()
action_spec = common.get_action_spec()
conv_layers = tf.keras.Sequential(
tf_agents.networks.utils.mlp_layers(
conv_layer_params=conv_layer_params))
preprocessing_layers = {
'image': conv_layers,
}
if common.get_states_shape():
state_layers = get_identity_layer()
# [image: (1, 12800), sentence: (1, 16 * 800), states: (1, 16 * 800)]
# Here, we tile along the last dimension of the input.
if num_state_tiles:
state_layers = tf.keras.Sequential([
tf.keras.layers.Lambda(
lambda x: tf.tile(x, multiples=[1, num_state_tiles]))
])
preprocessing_layers['states'] = state_layers
vocab_size = common.get_vocab_size()
if vocab_size:
sentence_layers = tf.keras.Sequential([
tf.keras.layers.Embedding(vocab_size, num_embedding_dims),
tf.keras.layers.GlobalAveragePooling1D()
])
if num_sentence_tiles:
sentence_layers.add(
tf.keras.layers.Lambda(
lambda x: tf.tile(x, multiples=[1, num_sentence_tiles])))
preprocessing_layers['sentence'] = sentence_layers
preprocessing_combiner = tf.keras.layers.Concatenate()
actor = ActorDistributionRnnNetwork(
input_tensor_spec=observation_spec,
output_tensor_spec=action_spec,
preprocessing_layers=preprocessing_layers,
preprocessing_combiner=preprocessing_combiner,
input_fc_layer_params=fc_layer_params)
value = ValueRnnNetwork(input_tensor_spec=observation_spec,
preprocessing_layers=preprocessing_layers,
preprocessing_combiner=preprocessing_combiner,
input_fc_layer_params=fc_layer_params)
return actor, value