def record_demo(action_sizes, num_visual=0, num_vector=1):
env = RecordEnvironment(
[BRAIN_NAME],
action_sizes=action_sizes,
num_visual=num_visual,
num_vector=num_vector,
n_demos=100,
)
# If we want to use true demos, we can solve the env in the usual way
# Otherwise, we can just call solve to execute the optimal policy
env.solve()
agent_info_protos = env.demonstration_protos[BRAIN_NAME]
meta_data_proto = DemonstrationMetaProto()
continuous_action_size, discrete_action_size = action_sizes
action_spec_proto = ActionSpecProto(
num_continuous_actions=continuous_action_size,
num_discrete_actions=discrete_action_size,
discrete_branch_sizes=[2] if discrete_action_size > 0 else None,
)
brain_param_proto = BrainParametersProto(brain_name=BRAIN_NAME,
is_training=True,
action_spec=action_spec_proto)
action_type = "Discrete" if action_sizes else "Continuous"
demo_path_name = "1DTest" + action_type + ".demo"
demo_path = str(tmpdir_factory.mktemp("tmp_demo").join(demo_path_name))
write_demo(demo_path, meta_data_proto, brain_param_proto,
agent_info_protos)
return demo_path
def initialize(self, inputs: UnityInputProto) -> UnityOutputProto:
if self.is_discrete:
action_spec = ActionSpecProto(num_discrete_actions=2,
discrete_branch_sizes=[3, 2])
else:
action_spec = ActionSpecProto(num_continuous_actions=2)
bp = BrainParametersProto(brain_name=self.brain_name,
is_training=True,
action_spec=action_spec)
rl_init = UnityRLInitializationOutputProto(
name="RealFakeAcademy",
communication_version=UnityEnvironment.API_VERSION,
package_version="mock_package_version",
log_path="",
brain_parameters=[bp],
)
output = UnityRLOutputProto(agentInfos=self._get_agent_infos())
return UnityOutputProto(rl_initialization_output=rl_init,
rl_output=output)