def _build_configuration(self):
"""Builds a configuration using an SAC agent
"""
self._scenario_generator = \
UniformVehicleDistribution(num_scenarios=20,
random_seed=0,
params=self._params)
self._observer = CustomObserver(params=self._params)
self._behavior_model = DynamicModel(params=self._params)
self._evaluator = CustomEvaluator(params=self._params)
self._viewer = MPViewer(params=self._params,
x_range=[-30, 30],
y_range=[-60, 20],
follow_agent_id=True)
#self._viewer = VideoRenderer(renderer=viewer, world_step_time=0.2)
self._runtime = RuntimeRL(action_wrapper=self._behavior_model,
observer=self._observer,
evaluator=self._evaluator,
step_time=0.2,
viewer=self._viewer,
scenario_generator=self._scenario_generator)
tfa_env = tf_py_environment.TFPyEnvironment(TFAWrapper(self._runtime))
self._agent_0 = SACAgent(tfa_env, params=self._params)
self._agent_1 = SACAgent(tfa_env, params=self._params)
self._runner = SACRunner(tfa_env, [self._agent_0, self._agent_1],
params=self._params,
unwrapped_runtime=self._runtime)
def test_tfa_runtime():
params = ParameterServer(
filename="tests/data/deterministic_scenario_test.json")
scenario_generation = DeterministicScenarioGeneration(num_scenarios=3,
random_seed=0,
params=params)
state_observer = ClosestAgentsObserver(params=params)
action_wrapper = DynamicModel(params=params)
evaluator = GoalReached(params=params)
viewer = MPViewer(params=params,
x_range=[-30, 30],
y_range=[-20, 40],
follow_agent_id=True) # use_world_bounds=True
runtimerl = RuntimeRL(action_wrapper=action_wrapper,
observer=state_observer,
evaluator=evaluator,
step_time=0.05,
viewer=viewer,
scenario_generator=scenario_generation)
tfa_env = TFAWrapper(runtimerl)
_ = tfa_env.reset()
utils.validate_py_environment(tfa_env, episodes=5)
_ = tf_py_environment.TFPyEnvironment(tfa_env)
def test_runner():
params = ParameterServer(
filename="tests/data/deterministic_scenario_test.json")
base_dir = os.path.dirname(os.path.dirname(__file__))
params["BaseDir"] = base_dir
scenario_generation = DeterministicScenarioGeneration(num_scenarios=3,
random_seed=0,
params=params)
state_observer = ClosestAgentsObserver(params=params)
action_wrapper = DynamicModel(params=params)
evaluator = GoalReached(params=params)
viewer = MPViewer(params=params,
x_range=[-30, 30],
y_range=[-20, 40],
follow_agent_id=True)
runtimerl = RuntimeRL(action_wrapper=action_wrapper,
observer=state_observer,
evaluator=evaluator,
step_time=0.2,
viewer=viewer,
scenario_generator=scenario_generation,
render=False)
tfa_env = tf_py_environment.TFPyEnvironment(TFAWrapper(runtimerl))
sac_agent = SACAgent(tfa_env, params=params)
tfa_runner = TFARunner(tfa_env,
sac_agent,
params=params,
unwrapped_runtime=runtimerl)
tfa_runner.collect_initial_episodes()
# main functionalities
tfa_runner.train()
tfa_runner.visualize()
tfa_runner.evaluate()
def test_agent(self):
params = ParameterServer(
filename="tests/data/deterministic_scenario_test.json")
base_dir = os.path.dirname(os.path.dirname(__file__))
params["BaseDir"] = base_dir
scenario_generation = DeterministicScenarioGeneration(num_scenarios=2,
random_seed=0,
params=params)
state_observer = ClosestAgentsObserver(params=params)
action_wrapper = DynamicModel(params=params)
evaluator = GoalReached(params=params)
viewer = MPViewer(params=params,
x_range=[-30, 30],
y_range=[-20, 40],
follow_agent_id=True) # use_world_bounds=True
runtimerl = RuntimeRL(action_wrapper=action_wrapper,
observer=state_observer,
evaluator=evaluator,
step_time=0.05,
viewer=viewer,
scenario_generator=scenario_generation)
tfa_env = tf_py_environment.TFPyEnvironment(TFAWrapper(runtimerl))
sac_agent = SACAgent(tfa_env, params=params)
self.assertEqual(sac_agent._agent.name, "sac_agent")
sac_agent.reset()
# TODO(@hart): does not work because of read-only file-system
# sac_agent.save()
sac_agent.load()
def test_triple_int(self):
params = ParameterServer(
filename="tests/data/deterministic_scenario_drone_test.json")
scenario_generation = DeterministicScenarioGeneration(num_scenarios=3,
random_seed=0,
params=params)
state_observer = SimpleObserver(params=params)
action_wrapper = DynamicModel(model_name="TripleIntegratorModel",
params=params)
evaluator = GoalReached(params=params)
viewer = MPViewer(params=params,
x_range=[-30, 30],
y_range=[-40, 40],
use_world_bounds=True)
runtimerl = RuntimeRL(action_wrapper=action_wrapper,
observer=state_observer,
evaluator=evaluator,
step_time=0.2,
viewer=viewer,
scenario_generator=scenario_generation,
render=True)
for _ in range(0, 1):
runtimerl.reset()
for _ in range(0, 10): # run each scenario for 10 steps
action = action_wrapper.action_space.sample()
next_observed_state, reward, done, info = \
runtimerl.step(action)
if done:
print("State: {} \n Reward: {} \n Done {}, Info: {} \n \
=================================================" . \
format(next_observed_state, reward, done, info))
break
def _build_configuration(self):
"""Builds a configuration using an SAC agent
"""
self._scenario_generator = \
DeterministicDroneChallengeGeneration(num_scenarios=3,
random_seed=0,
params=self._params)
self._observer = CustomObserver(params=self._params)
self._behavior_model = DynamicModel(model_name="TripleIntegratorModel",
params=self._params)
self._evaluator = CustomEvaluator(params=self._params)
viewer = MPViewer(params=self._params,
x_range=[-20, 20],
y_range=[-20, 20],
follow_agent_id=True)
self._viewer = viewer
# self._viewer = VideoRenderer(renderer=viewer, world_step_time=0.2)
self._runtime = RuntimeRL(action_wrapper=self._behavior_model,
observer=self._observer,
evaluator=self._evaluator,
step_time=0.2,
viewer=self._viewer,
scenario_generator=self._scenario_generator)
# tfa_env = tf_py_environment.TFPyEnvironment(TFAWrapper(self._runtime))
tfa_env = tf_py_environment.TFPyEnvironment(
parallel_py_environment.ParallelPyEnvironment(
[lambda: TFAWrapper(self._runtime)] * self._params["ML"]["Agent"]["num_parallel_environments"]))
self._agent = SACAgent(tfa_env, params=self._params)
self._runner = SACRunner(tfa_env,
self._agent,
params=self._params,
unwrapped_runtime=self._runtime)
def test_runtime_rl(self):
params = ParameterServer(
filename="tests/data/deterministic_scenario_test.json")
scenario_generation = DeterministicScenarioGeneration(num_scenarios=2,
random_seed=0,
params=params)
state_observer = SimpleObserver(params=params)
action_wrapper = DynamicModel(params=params)
evaluator = GoalReached(params=params)
viewer = MPViewer(params=params,
x_range=[-30, 30],
y_range=[-40, 40],
use_world_bounds=True) #use_world_bounds=True) #
runtimerl = RuntimeRL(action_wrapper=action_wrapper,
observer=state_observer,
evaluator=evaluator,
step_time=0.2,
viewer=viewer,
scenario_generator=scenario_generation,
render=False)
start_time = time.time()
for _ in range(0, 100):
runtimerl.reset()
done = False
reward = 0.
for _ in range(0, 50): # run each scenario for 10 steps
action = action_wrapper.action_space.sample(
) / 100 # to go straight
next_observed_state, reward, done, info = \
runtimerl.step(action)
# observer
self.assertEqual(len(next_observed_state), 16)
np.testing.assert_array_equal(
next_observed_state[0:4],
state_observer._normalize(
runtimerl._world.agents[100].state)[1:5])
np.testing.assert_array_equal(
next_observed_state[4:8],
state_observer._normalize(
runtimerl._world.agents[101].state)[1:5])
if done:
print("State: {} \n Reward: {} \n Done {}, Info: {} \n \
=================================================" . \
format(next_observed_state, reward, done, info))
break
# must assert to equal as the agent reaches the goal in the
# specified number of steps
self.assertEqual(done, True)
# goal must have been reached which returns a reward of 1.
self.assertEqual(reward, 1.)
self.assertEqual(runtimerl._world.agents[100].id, 100)
self.assertEqual(runtimerl._world.agents[101].id, 101)
end_time = time.time()
print("100 runs took {}s.".format(str(end_time - start_time)))
def _build_configuration(self):
"""Builds a configuration using an SAC agent
"""
# TODO(@hart): increase scenario number
self._scenario_generator = \
ConfigurableScenarioGeneration(num_scenarios=100,
params=self._params)
# USE THIS FOR DETERMINISTIC SCENARIO GEN.
# self._scenario_generator = \
# DeterministicScenarioGeneration(num_scenarios=100,
# params=self._params)
# self._observer = NearestObserver(self._params)
self._observer = ClosestAgentsObserver(self._params)
self._behavior_model = DynamicModel(params=self._params)
self._evaluator = CustomEvaluator(params=self._params)
sim_step_time = 0.2
real_time_factor = 5
resolution = (1920, 1080)
dpi = 300
fig_env = plt.figure(figsize=(resolution[0] / dpi,
resolution[1] / dpi),
dpi=dpi)
gs = gridspec.GridSpec(1, 1, left=0.0, right=1, bottom=0, top=0.9)
axis = plt.subplot(gs[0])
viewer = MPViewer(params=self._params,
y_length=80,
enforce_y_length=True,
enforce_x_length=False,
follow_agent_id=True,
axis=axis)
# self._viewer = VideoRenderer(renderer=viewer, world_step_time=0.2)
self._viewer = viewer
self._runtime = RuntimeRL(action_wrapper=self._behavior_model,
observer=self._observer,
evaluator=self._evaluator,
step_time=0.2,
viewer=self._viewer,
scenario_generator=self._scenario_generator)
tfa_env = tf_py_environment.TFPyEnvironment(TFAWrapper(self._runtime))
eval_tf_env = tf_py_environment.TFPyEnvironment(
TFAWrapper(self._runtime))
# self._agent = PPOAgent(tfa_env, params=self._params)
# self._runner = PPORunner(tfa_env,
# eval_tf_env,
# self._agent,
# params=self._params,
# unwrapped_runtime=self._runtime)
self._agent = SACAgent(tfa_env, params=self._params)
self._runner = SACRunner(tfa_env,
self._agent,
params=self._params,
unwrapped_runtime=self._runtime)