def __str__(self):
result = []
for network in range(self.num_networks):
network_structure = []
# embedder
for embedder in self.input_embedders:
network_structure.append("Input Embedder: {}".format(
embedder.name))
network_structure.append(indent_string(str(embedder)))
if len(self.input_embedders) > 1:
network_structure.append("{} ({})".format(
self.network_parameters.embedding_merger_type.name,
", ".join([
"{} embedding".format(e.name)
for e in self.input_embedders
])))
# middleware
network_structure.append("Middleware:")
network_structure.append(indent_string(str(self.middleware)))
# head
if self.network_parameters.use_separate_networks_per_head:
heads = range(network, network + 1)
else:
heads = range(0, len(self.output_heads))
for head_idx in heads:
head = self.output_heads[head_idx]
head_params = self.network_parameters.heads_parameters[
head_idx]
if head_params.num_output_head_copies > 1:
network_structure.append(
"Output Head: {} (num copies = {})".format(
head.name, head_params.num_output_head_copies))
else:
network_structure.append("Output Head: {}".format(
head.name))
network_structure.append(indent_string(str(head)))
# finalize network
if self.num_networks > 1:
result.append("Sub-network for head: {}".format(
self.output_heads[network].name))
result.append(indent_string('\n'.join(network_structure)))
else:
result.append('\n'.join(network_structure))
result = '\n'.join(result)
return result
def __str__(self):
action_spaces = [self.spaces.action]
if isinstance(self.spaces.action, CompoundActionSpace):
action_spaces = self.spaces.action.sub_action_spaces
result = []
for action_space_idx, action_space in enumerate(action_spaces):
action_head_mean_result = []
if isinstance(action_space, DiscreteActionSpace):
# create a discrete action network (softmax probabilities output)
action_head_mean_result.append(
"Dense (num outputs = {})".format(len(
action_space.actions)))
action_head_mean_result.append("Softmax")
elif isinstance(action_space, BoxActionSpace):
# create a continuous action network (bounded mean and stdev outputs)
action_head_mean_result.append(
"Dense (num outputs = {})".format(action_space.shape))
if np.all(action_space.max_abs_range < np.inf):
# bounded actions
action_head_mean_result.append(
"Activation (type = {})".format(
self.activation_function.__name__))
action_head_mean_result.append(
"Multiply (factor = {})".format(
action_space.max_abs_range))
action_head_stdev_result = []
if isinstance(self.exploration_policy,
ContinuousEntropyParameters):
action_head_stdev_result.append(
"Dense (num outputs = {})".format(action_space.shape))
action_head_stdev_result.append("Softplus")
action_head_result = []
if action_head_stdev_result:
action_head_result.append("Mean Stream")
action_head_result.append(
indent_string('\n'.join(action_head_mean_result)))
action_head_result.append("Stdev Stream")
action_head_result.append(
indent_string('\n'.join(action_head_stdev_result)))
else:
action_head_result.append('\n'.join(action_head_mean_result))
if len(action_spaces) > 1:
result.append("Action head {}".format(action_space_idx))
result.append(indent_string('\n'.join(action_head_result)))
else:
result.append('\n'.join(action_head_result))
return '\n'.join(result)