def create_mock_vector_steps(specs,
num_agents=1,
number_visual_observations=0):
"""
Creates a mock BatchedStepResult with vector observations. Imitates constant
vector observations, rewards, dones, and agents.
:BehaviorSpecs specs: The BehaviorSpecs for this mock
:int num_agents: Number of "agents" to imitate in your BatchedStepResult values.
"""
obs = [
np.array([num_agents * [1, 2, 3]],
dtype=np.float32).reshape(num_agents, 3)
]
if number_visual_observations:
obs += [np.zeros(shape=(num_agents, 8, 8, 3), dtype=np.float32)
] * number_visual_observations
rewards = np.array(num_agents * [1.0])
agents = np.array(range(0, num_agents))
group_id = np.array(num_agents * [0])
group_rewards = np.array(num_agents * [0.0])
return (
DecisionSteps(obs, rewards, agents, None, group_id, group_rewards),
TerminalSteps.empty(specs),
)
def _make_batched_step(
self, name: str, done: bool,
reward: float) -> Tuple[DecisionSteps, TerminalSteps]:
m_vector_obs = self._make_obs(self.goal[name])
m_reward = np.array([reward], dtype=np.float32)
m_agent_id = np.array([self.agent_id[name]], dtype=np.int32)
action_mask = self._generate_mask()
decision_step = DecisionSteps(m_vector_obs, m_reward, m_agent_id,
action_mask)
terminal_step = TerminalSteps.empty(self.behavior_spec)
if done:
self._reset_agent(name)
new_vector_obs = self._make_obs(self.goal[name])
(
new_reward,
new_done,
new_agent_id,
new_action_mask,
) = self._construct_reset_step(name)
decision_step = DecisionSteps(new_vector_obs, new_reward,
new_agent_id, new_action_mask)
terminal_step = TerminalSteps(m_vector_obs, m_reward,
np.array([False], dtype=np.bool),
m_agent_id)
return (decision_step, terminal_step)
def _make_batched_step(
self, name: str, done: bool,
reward: float) -> Tuple[DecisionSteps, TerminalSteps]:
recurrent_obs_val = (self.goal[name]
if self.step_count[name] <= self.num_show_steps
else 0)
m_vector_obs = self._make_obs(recurrent_obs_val)
m_reward = np.array([reward], dtype=np.float32)
m_agent_id = np.array([self.agent_id[name]], dtype=np.int32)
action_mask = self._generate_mask()
decision_step = DecisionSteps(m_vector_obs, m_reward, m_agent_id,
action_mask)
terminal_step = TerminalSteps.empty(self.behavior_spec)
if done:
self.final_rewards[name].append(self.rewards[name])
self._reset_agent(name)
recurrent_obs_val = (self.goal[name] if
self.step_count[name] <= self.num_show_steps
else 0)
new_vector_obs = self._make_obs(recurrent_obs_val)
(
new_reward,
new_done,
new_agent_id,
new_action_mask,
) = self._construct_reset_step(name)
decision_step = DecisionSteps(new_vector_obs, new_reward,
new_agent_id, new_action_mask)
terminal_step = TerminalSteps(m_vector_obs, m_reward,
np.array([False], dtype=np.bool),
m_agent_id)
return (decision_step, terminal_step)
def test_empty_terminal_steps():
specs = BehaviorSpec(observation_shapes=[(3, 2), (5, )],
action_spec=ActionSpec.create_continuous(3))
ts = TerminalSteps.empty(specs)
assert len(ts.obs) == 2
assert ts.obs[0].shape == (0, 3, 2)
assert ts.obs[1].shape == (0, 5)
def test_empty_terminal_steps():
specs = BehaviorSpec(
sensor_specs=create_sensor_specs_with_shapes([(3, 2), (5, )]),
action_spec=ActionSpec.create_continuous(3),
)
ts = TerminalSteps.empty(specs)
assert len(ts.obs) == 2
assert ts.obs[0].shape == (0, 3, 2)
assert ts.obs[1].shape == (0, 5)
def test_empty_terminal_steps():
specs = BehaviorSpec(
observation_shapes=[(3, 2), (5, )],
action_type=ActionType.CONTINUOUS,
action_shape=3,
)
ts = TerminalSteps.empty(specs)
assert len(ts.obs) == 2
assert ts.obs[0].shape == (0, 3, 2)
assert ts.obs[1].shape == (0, 5)
def create_mock_steps(
num_agents: int = 1,
num_vector_observations: int = 0,
num_vis_observations: int = 0,
action_shape: List[int] = None,
discrete: bool = False,
done: bool = False,
) -> Tuple[DecisionSteps, TerminalSteps]:
"""
Creates a mock Tuple[DecisionSteps, TerminalSteps] with observations.
Imitates constant vector/visual observations, rewards, dones, and agents.
:int num_agents: Number of "agents" to imitate.
:int num_vector_observations: Number of "observations" in your observation space
:int num_vis_observations: Number of "observations" in your observation space
:int num_vector_acts: Number of actions in your action space
:bool discrete: Whether or not action space is discrete
:bool done: Whether all the agents in the batch are done
"""
if action_shape is None:
action_shape = [2]
obs_list = []
for _ in range(num_vis_observations):
obs_list.append(np.ones((num_agents, 84, 84, 3), dtype=np.float32))
if num_vector_observations > 1:
obs_list.append(
np.array(num_agents * [num_vector_observations * [1]],
dtype=np.float32))
action_mask = None
if discrete:
action_mask = [
np.array(num_agents * [action_size * [False]])
for action_size in action_shape
]
reward = np.array(num_agents * [1.0], dtype=np.float32)
interrupted = np.array(num_agents * [False], dtype=np.bool)
agent_id = np.arange(num_agents, dtype=np.int32)
behavior_spec = BehaviorSpec(
[(84, 84, 3)] * num_vis_observations +
[(num_vector_observations, 0, 0)],
ActionType.DISCRETE if discrete else ActionType.CONTINUOUS,
action_shape if discrete else action_shape[0],
)
if done:
return (
DecisionSteps.empty(behavior_spec),
TerminalSteps(obs_list, reward, interrupted, agent_id),
)
else:
return (
DecisionSteps(obs_list, reward, agent_id, action_mask),
TerminalSteps.empty(behavior_spec),
)
def create_mock_steps(
num_agents: int,
observation_shapes: List[Tuple],
action_shape: Union[int, Tuple[int]] = None,
discrete: bool = False,
done: bool = False,
) -> Tuple[DecisionSteps, TerminalSteps]:
"""
Creates a mock Tuple[DecisionSteps, TerminalSteps] with observations.
Imitates constant vector/visual observations, rewards, dones, and agents.
:int num_agents: Number of "agents" to imitate.
:List observation_shapes: A List of the observation spaces in your steps
:int num_vector_acts: Number of actions in your action space
:bool discrete: Whether or not action space is discrete
:bool done: Whether all the agents in the batch are done
"""
if action_shape is None:
action_shape = 2
obs_list = []
for _shape in observation_shapes:
obs_list.append(np.ones((num_agents, ) + _shape, dtype=np.float32))
action_mask = None
if discrete and isinstance(action_shape, Iterable):
action_mask = [
np.array(num_agents * [action_size * [False]])
for action_size in action_shape # type: ignore
] # type: ignore
reward = np.array(num_agents * [1.0], dtype=np.float32)
interrupted = np.array(num_agents * [False], dtype=np.bool)
agent_id = np.arange(num_agents, dtype=np.int32)
behavior_spec = BehaviorSpec(
observation_shapes,
ActionType.DISCRETE if discrete else ActionType.CONTINUOUS,
action_shape,
)
if done:
return (
DecisionSteps.empty(behavior_spec),
TerminalSteps(obs_list, reward, interrupted, agent_id),
)
else:
return (
DecisionSteps(obs_list, reward, agent_id, action_mask),
TerminalSteps.empty(behavior_spec),
)
def _update_state(self, output: UnityRLOutputProto) -> None:
"""
Collects experience information from all external brains in environment at current step.
"""
for brain_name in self._env_specs.keys():
if brain_name in output.agentInfos:
agent_info_list = output.agentInfos[brain_name].value
self._env_state[brain_name] = steps_from_proto(
agent_info_list, self._env_specs[brain_name]
)
else:
self._env_state[brain_name] = (
DecisionSteps.empty(self._env_specs[brain_name]),
TerminalSteps.empty(self._env_specs[brain_name]),
)
self._side_channel_manager.process_side_channel_message(output.side_channel)
def create_mock_steps(
num_agents: int,
observation_specs: List[ObservationSpec],
action_spec: ActionSpec,
done: bool = False,
grouped: bool = False,
) -> Tuple[DecisionSteps, TerminalSteps]:
"""
Creates a mock Tuple[DecisionSteps, TerminalSteps] with observations.
Imitates constant vector/visual observations, rewards, dones, and agents.
:int num_agents: Number of "agents" to imitate.
:List observation_specs: A List of the observation specs in your steps
:int action_spec: ActionSpec for the agent
:bool done: Whether all the agents in the batch are done
"""
obs_list = []
for obs_spec in observation_specs:
obs_list.append(np.ones((num_agents,) + obs_spec.shape, dtype=np.float32))
action_mask = None
if action_spec.is_discrete():
action_mask = [
np.array(num_agents * [action_size * [False]])
for action_size in action_spec.discrete_branches # type: ignore
] # type: ignore
reward = np.array(num_agents * [1.0], dtype=np.float32)
interrupted = np.array(num_agents * [False], dtype=np.bool)
agent_id = np.arange(num_agents, dtype=np.int32)
_gid = 1 if grouped else 0
group_id = np.array(num_agents * [_gid], dtype=np.int32)
group_reward = np.array(num_agents * [0.0], dtype=np.float32)
behavior_spec = BehaviorSpec(observation_specs, action_spec)
if done:
return (
DecisionSteps.empty(behavior_spec),
TerminalSteps(
obs_list, reward, interrupted, agent_id, group_id, group_reward
),
)
else:
return (
DecisionSteps(
obs_list, reward, agent_id, action_mask, group_id, group_reward
),
TerminalSteps.empty(behavior_spec),
)
def get_steps(self, behavior_name):
# This gets the individual DecisionSteps and TerminalSteps
# from the envs and merges them into a batch to be sent
# to the AgentProcessor.
dec_vec_obs = []
dec_reward = []
dec_group_reward = []
dec_agent_id = []
dec_group_id = []
ter_vec_obs = []
ter_reward = []
ter_group_reward = []
ter_agent_id = []
ter_group_id = []
interrupted = []
action_mask = None
terminal_step = TerminalSteps.empty(self.behavior_spec)
decision_step = None
for i in range(self.num_agents):
name_and_num = behavior_name + str(i)
env = self.envs[name_and_num]
_dec, _term = env.step_result[behavior_name]
if not self.dones[name_and_num]:
dec_agent_id.append(i)
dec_group_id.append(1)
if len(dec_vec_obs) > 0:
for j, obs in enumerate(_dec.obs):
dec_vec_obs[j] = np.concatenate((dec_vec_obs[j], obs),
axis=0)
else:
for obs in _dec.obs:
dec_vec_obs.append(obs)
dec_reward.append(_dec.reward[0])
dec_group_reward.append(_dec.group_reward[0])
if _dec.action_mask is not None:
if action_mask is None:
action_mask = []
if len(action_mask) > 0:
action_mask[0] = np.concatenate(
(action_mask[0], _dec.action_mask[0]), axis=0)
else:
action_mask.append(_dec.action_mask[0])
if len(_term.reward) > 0 and name_and_num in self.just_died:
ter_agent_id.append(i)
ter_group_id.append(1)
if len(ter_vec_obs) > 0:
for j, obs in enumerate(_term.obs):
ter_vec_obs[j] = np.concatenate((ter_vec_obs[j], obs),
axis=0)
else:
for obs in _term.obs:
ter_vec_obs.append(obs)
ter_reward.append(_term.reward[0])
ter_group_reward.append(_term.group_reward[0])
interrupted.append(False)
self.just_died.remove(name_and_num)
decision_step = DecisionSteps(
dec_vec_obs,
dec_reward,
dec_agent_id,
action_mask,
dec_group_id,
dec_group_reward,
)
terminal_step = TerminalSteps(
ter_vec_obs,
ter_reward,
interrupted,
ter_agent_id,
ter_group_id,
ter_group_reward,
)
return (decision_step, terminal_step)