def load_demonstration(
file_path: str,
) -> Tuple[BehaviorSpec, List[AgentInfoActionPairProto], int]:
"""
Loads and parses a demonstration file.
:param file_path: Location of demonstration file (.demo).
:return: BrainParameter and list of AgentInfoActionPairProto containing demonstration data.
"""
# First 32 bytes of file dedicated to meta-data.
file_paths = get_demo_files(file_path)
behavior_spec = None
brain_param_proto = None
info_action_pairs = []
total_expected = 0
for _file_path in file_paths:
with open(_file_path, "rb") as fp:
with hierarchical_timer("read_file"):
data = fp.read()
next_pos, pos, obs_decoded = 0, 0, 0
while pos < len(data):
next_pos, pos = _DecodeVarint32(data, pos)
if obs_decoded == 0:
meta_data_proto = DemonstrationMetaProto()
meta_data_proto.ParseFromString(data[pos : pos + next_pos])
if (
meta_data_proto.api_version
not in SUPPORTED_DEMONSTRATION_VERSIONS
):
raise RuntimeError(
f"Can't load Demonstration data from an unsupported version ({meta_data_proto.api_version})"
)
total_expected += meta_data_proto.number_steps
pos = INITIAL_POS
if obs_decoded == 1:
brain_param_proto = BrainParametersProto()
brain_param_proto.ParseFromString(data[pos : pos + next_pos])
pos += next_pos
if obs_decoded > 1:
agent_info_action = AgentInfoActionPairProto()
agent_info_action.ParseFromString(data[pos : pos + next_pos])
if behavior_spec is None:
behavior_spec = behavior_spec_from_proto(
brain_param_proto, agent_info_action.agent_info
)
info_action_pairs.append(agent_info_action)
if len(info_action_pairs) == total_expected:
break
pos += next_pos
obs_decoded += 1
if not behavior_spec:
raise RuntimeError(
f"No BrainParameters found in demonstration file at {file_path}."
)
return behavior_spec, info_action_pairs, total_expected
def record_demo(use_discrete, num_visual=0, num_vector=1):
env = RecordEnvironment(
[BRAIN_NAME],
use_discrete=use_discrete,
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()
brain_param_proto = BrainParametersProto(
vector_action_size=[2] if use_discrete else [1],
vector_action_descriptions=[""],
vector_action_space_type=discrete if use_discrete else continuous,
brain_name=BRAIN_NAME,
is_training=True,
)
action_type = "Discrete" if use_discrete 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 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:
bp = BrainParametersProto(
vector_action_size=[2],
vector_action_descriptions=["", ""],
vector_action_space_type=discrete if self.is_discrete else continuous,
brain_name=self.brain_name,
is_training=True,
)
rl_init = UnityRLInitializationOutputProto(
name="RealFakeAcademy",
version=UnityEnvironment.API_VERSION,
log_path="",
brain_parameters=[bp],
)
output = UnityRLOutputProto(agentInfos=self._get_agent_infos())
return UnityOutputProto(rl_initialization_output=rl_init, rl_output=output)
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)
def test_agent_behavior_spec_from_proto():
agent_proto = generate_list_agent_proto(1, [(3, ), (4, )])[0]
bp = BrainParametersProto()
bp.vector_action_size.extend([5, 4])
bp.vector_action_space_type = 0
behavior_spec = behavior_spec_from_proto(bp, agent_proto)
assert behavior_spec.is_action_discrete()
assert not behavior_spec.is_action_continuous()
assert behavior_spec.observation_shapes == [(3, ), (4, )]
assert behavior_spec.discrete_action_branches == (5, 4)
assert behavior_spec.action_size == 2
bp = BrainParametersProto()
bp.vector_action_size.extend([6])
bp.vector_action_space_type = 1
behavior_spec = behavior_spec_from_proto(bp, agent_proto)
assert not behavior_spec.is_action_discrete()
assert behavior_spec.is_action_continuous()
assert behavior_spec.action_size == 6
def test_agent_behavior_spec_from_proto():
agent_proto = generate_list_agent_proto(1, [(3, ), (4, )])[0]
bp = BrainParametersProto()
bp.vector_action_size_deprecated.extend([5, 4])
bp.vector_action_space_type_deprecated = 0
behavior_spec = behavior_spec_from_proto(bp, agent_proto)
assert behavior_spec.action_spec.is_discrete()
assert not behavior_spec.action_spec.is_continuous()
assert [spec.shape for spec in behavior_spec.sensor_specs] == [(3, ),
(4, )]
assert behavior_spec.action_spec.discrete_branches == (5, 4)
assert behavior_spec.action_spec.discrete_size == 2
bp = BrainParametersProto()
bp.vector_action_size_deprecated.extend([6])
bp.vector_action_space_type_deprecated = 1
behavior_spec = behavior_spec_from_proto(bp, agent_proto)
assert not behavior_spec.action_spec.is_discrete()
assert behavior_spec.action_spec.is_continuous()
assert behavior_spec.action_spec.continuous_size == 6
def load_demonstration(
file_path: str
) -> Tuple[BrainParameters, List[AgentInfoActionPairProto], int]:
"""
Loads and parses a demonstration file.
:param file_path: Location of demonstration file (.demo).
:return: BrainParameter and list of AgentInfoActionPairProto containing demonstration data.
"""
# First 32 bytes of file dedicated to meta-data.
INITIAL_POS = 33
file_paths = []
if os.path.isdir(file_path):
all_files = os.listdir(file_path)
for _file in all_files:
if _file.endswith(".demo"):
file_paths.append(os.path.join(file_path, _file))
if not all_files:
raise ValueError("There are no '.demo' files in the provided directory.")
elif os.path.isfile(file_path):
file_paths.append(file_path)
file_extension = pathlib.Path(file_path).suffix
if file_extension != ".demo":
raise ValueError(
"The file is not a '.demo' file. Please provide a file with the "
"correct extension."
)
else:
raise FileNotFoundError(
"The demonstration file or directory {} does not exist.".format(file_path)
)
brain_params = None
brain_param_proto = None
info_action_pairs = []
total_expected = 0
for _file_path in file_paths:
with open(_file_path, "rb") as fp:
with hierarchical_timer("read_file"):
data = fp.read()
next_pos, pos, obs_decoded = 0, 0, 0
while pos < len(data):
next_pos, pos = _DecodeVarint32(data, pos)
if obs_decoded == 0:
meta_data_proto = DemonstrationMetaProto()
meta_data_proto.ParseFromString(data[pos : pos + next_pos])
total_expected += meta_data_proto.number_steps
pos = INITIAL_POS
if obs_decoded == 1:
brain_param_proto = BrainParametersProto()
brain_param_proto.ParseFromString(data[pos : pos + next_pos])
pos += next_pos
if obs_decoded > 1:
agent_info_action = AgentInfoActionPairProto()
agent_info_action.ParseFromString(data[pos : pos + next_pos])
if brain_params is None:
brain_params = BrainParameters.from_proto(
brain_param_proto, agent_info_action.agent_info
)
info_action_pairs.append(agent_info_action)
if len(info_action_pairs) == total_expected:
break
pos += next_pos
obs_decoded += 1
if not brain_params:
raise RuntimeError(
f"No BrainParameters found in demonstration file at {file_path}."
)
return brain_params, info_action_pairs, total_expected