def main():
args = configure_args()
exp_dir = args.exp_dir or "results/training/toy_evaluation"
params_filename = glob.glob(os.path.join(exp_dir, "params_*"))
params = ParameterServer(filename=params_filename[0])
behavior_params_filename = glob.glob(
os.path.join(exp_dir, "behavior_params*"))
if behavior_params_filename:
splits = 8
behavior_params = ParameterServer(filename=behavior_params_filename[0])
behavior_space = BehaviorSpace(behavior_params)
hypothesis_set, _ = behavior_space.create_hypothesis_set_fixed_split(
split=splits)
observer = BeliefObserver(params, hypothesis_set, splits=splits)
behavior = BehaviorDiscreteMacroActionsML(behavior_params)
else:
behavior = BehaviorDiscreteMacroActionsML(params)
observer = NearestAgentsObserver(params)
evaluator = GoalReached(params)
scenario_params = ParameterServer(
filename=
"configuration/database/scenario_sets/interaction_merging_light_dense_1D.json"
)
scenario_generator = ConfigurableScenarioGeneration(params=scenario_params,
num_scenarios=5)
scenario_file = glob.glob(os.path.join(exp_dir, "scenarios_list*"))
scenario_generator.load_scenario_list(scenario_file[0])
viewer = MPViewer(params=params,
x_range=[-35, 35],
y_range=[-35, 35],
follow_agent_id=True)
env = HyDiscreteHighway(behavior=behavior,
observer=observer,
evaluator=evaluator,
viewer=viewer,
scenario_generation=scenario_generator,
render=True)
env.reset()
actions = [0, 1, 2, 3, 4, 5, 6]
for action in actions:
print(action)
env.step(action)
agent = FQFAgent(env=env, test_env=env, params=params)
agent.load_models(os.path.join(exp_dir, "agent/checkpoints/final"))
def test_iqn_agent(self):
params = ParameterServer()
params["ML"]["BaseAgent"]["NumSteps"] = 2
params["ML"]["BaseAgent"]["MaxEpisodeSteps"] = 2
bp = DiscreteHighwayBlueprint(params,
number_of_senarios=10,
random_seed=0)
env = SingleAgentRuntime(blueprint=bp, render=False)
env._observer = NearestAgentsObserver(params)
env._action_wrapper = BehaviorDiscreteMacroActionsML(params)
iqn_agent = IQNAgent(agent_save_dir="./save_dir",
env=env,
params=params)
iqn_agent.train_episode()
iqn_agent.save(checkpoint_type="best")
iqn_agent.save(checkpoint_type="last")
loaded_agent = IQNAgent(agent_save_dir="./save_dir",
checkpoint_load="best")
loaded_agent2 = IQNAgent(agent_save_dir="./save_dir",
checkpoint_load="last")
loaded_agent_with_env = IQNAgent(env=env,
agent_save_dir="./save_dir",
checkpoint_load="last")
loaded_agent_with_env.train_episode()
self.assertEqual(loaded_agent.ml_behavior.action_space.n,
iqn_agent.ml_behavior.action_space.n)
return
def test_obs_traj(self):
map_filename = "external/bark_ml_project/bark_ml/environments/blueprints/highway/city_highway_straight.xodr"
params = ParameterServer()
behavior = BehaviorDiscreteMacroActionsML(params)
evaluator = GoalReached(params)
observer = NearestAgentsObserver(params)
viewer = MPViewer(params=params,
x_range=[-35, 35],
y_range=[-35, 35],
follow_agent_id=True)
env = HyDiscreteHighway(params=params,
map_filename=map_filename,
behavior=behavior,
evaluator=evaluator,
observer=observer,
viewer=viewer,
render=True)
env.reset()
actions = np.random.randint(0, 7, 100)
for action in actions:
concatenated_state, _, _, _ = env.step(action)
nn_ip_state = concatenated_state
ego_nn_input_state = deepcopy(
concatenated_state[0:observer._len_ego_state])
reverted_observed_state = observer.rev_observe_for_ego_vehicle(
nn_ip_state)
ext_reverted_observed_state = np.zeros(
(reverted_observed_state.shape[0] + 1))
ext_reverted_observed_state[1:] = reverted_observed_state
renormed_ego_state = observer._select_state_by_index(
observer._norm(ext_reverted_observed_state))
time.sleep(0.2)
np.testing.assert_array_almost_equal(ego_nn_input_state,
renormed_ego_state)
def test_collect_demonstrations(self):
params = ParameterServer()
bp = DiscreteHighwayBlueprint(params,
number_of_senarios=10,
random_seed=0)
env = SingleAgentRuntime(blueprint=bp, render=False)
env._observer = NearestAgentsObserver(params)
env._action_wrapper = BehaviorDiscreteMacroActionsML(params)
env._evaluator = TestEvaluator()
demo_behavior = bark_ml.library_wrappers.lib_fqf_iqn_qrdqn.\
tests.test_demo_behavior.TestDemoBehavior(params)
collector = DemonstrationCollector()
collection_result = collector.CollectDemonstrations(env, demo_behavior, 4, "./test_demo_collected", \
use_mp_runner=False, runner_init_params={"deepcopy" : False})
self.assertTrue(
os.path.exists("./test_demo_collected/collection_result"))
print(collection_result.get_data_frame().to_string())
experiences = collector.ProcessCollectionResult(
eval_criteria={"goal_r1": lambda x: x})
# expected length = 2 scenarios (only every second reaches goal) x 3 steps (4 executed, but first not counted)
self.assertEqual(len(experiences), 2 * 3)
collector.dump("./final_collections")
loaded_collector = DemonstrationCollector.load("./final_collections")
experiences_loaded = loaded_collector.GetDemonstrationExperiences()
print(experiences_loaded)
self.assertEqual(len(experiences_loaded), 2 * 3)
def __init__(self, params=None, number_of_senarios=25, random_seed=0):
ml_behavior = BehaviorDiscreteMacroActionsML(params)
MergingBlueprint.__init__(self,
params=params,
number_of_senarios=number_of_senarios,
random_seed=random_seed,
ml_behavior=ml_behavior,
viewer=True)
def main():
args = configure_args()
if is_local:
dir_prefix = ""
else:
dir_prefix="hy-iqn-lfd-full-exp.runfiles/hythe/"
logging.info(f"Executing job: {args.jobname}")
logging.info(f"Experiment server at: {os.getcwd()}")
params = ParameterServer(filename=os.path.join(dir_prefix, params_file),
log_if_default=True)
params = configure_params(params, seed=args.jobname)
experiment_id = params["Experiment"]["random_seed"]
params_filename = os.path.join(params["Experiment"]["dir"], "params_{}.json".format(experiment_id))
behavior = BehaviorDiscreteMacroActionsML(params)
evaluator = GoalReached(params)
observer = NearestAgentsObserver(params)
viewer = MPViewer(params=params,
x_range=[-35, 35],
y_range=[-35, 35],
follow_agent_id=True)
# extract params and save experiment parameters
params["ML"]["BaseAgent"]["SummaryPath"] = os.path.join(params["Experiment"]["dir"], "agent/summaries")
params["ML"]["BaseAgent"]["CheckpointPath"] = os.path.join(params["Experiment"]["dir"], "agent/checkpoints")
params.Save(filename=params_filename)
logging.info('-' * 60)
logging.info("Writing params to :{}".format(params_filename))
logging.info('-' * 60)
# database creation
dbs1 = DatabaseSerializer(test_scenarios=1, test_world_steps=2,
num_serialize_scenarios=num_demo_scenarios)
dbs1.process(os.path.join(dir_prefix, "configuration/database"),
filter_sets="**/**/interaction_merging_light_dense_1D.json")
local_release_filename = dbs1.release(version="lfd_offline")
db1 = BenchmarkDatabase(database_root=local_release_filename)
scenario_generator1, _, _ = db1.get_scenario_generator(0)
env = HyDiscreteHighway(params=params,
scenario_generation=scenario_generator1,
behavior=behavior,
evaluator=evaluator,
observer=observer,
viewer=viewer,
render=False)
scenario, _ = scenario_generator1.get_next_scenario()
world = scenario.GetWorldState()
observer.Reset(world)
assert env.action_space._n == 8, "Action Space is incorrect!"
run(args, params, env, db=db1)
params.Save(params_filename)
logging.info('-' * 60)
logging.info("Writing params to :{}".format(params_filename))
logging.info('-' * 60)
def main():
map_filename = "external/bark_ml_project/bark_ml/environments/blueprints/highway/city_highway_straight.xodr"
params_filename = glob.glob(os.path.join(demo_root, "params_[!behavior]*"))
params = ParameterServer(filename=params_filename[0], log_if_default=True)
behavior = BehaviorDiscreteMacroActionsML(params)
evaluator = GoalReached(params)
if is_belief_observer:
splits = 2
bparams_filename = glob.glob(os.path.join(demo_root, "behavior_*"))
params_behavior = ParameterServer(filename=bparams_filename[0])
behavior_space = BehaviorSpace(params_behavior)
hypothesis_set, hypothesis_params = behavior_space.create_hypothesis_set_fixed_split(
split=splits)
observer = BeliefObserver(params, hypothesis_set, splits=splits)
else:
observer = NearestAgentsObserver(params)
viewer = MPViewer(params=params,
x_range=[-35, 35],
y_range=[-35, 35],
follow_agent_id=True)
env = HyDiscreteHighway(params=params,
map_filename=map_filename,
behavior=behavior,
evaluator=evaluator,
observer=observer,
viewer=viewer,
render=False)
env.reset()
ego_world_states = []
_, demonstrations = unpack_load_demonstrations(demo_root)
for demo in demonstrations:
ego_state = np.zeros((observer._len_ego_state + 1))
(nn_ip_state, action, reward, next_state, done, is_demo) = demo
ego_nn_input_state = deepcopy(nn_ip_state[0:observer._len_ego_state])
ego_state[1:] = ego_nn_input_state
reverted_observed_state = observer.rev_observe_for_ego_vehicle(
ego_state)
ego_world_states.append(
(reverted_observed_state[int(StateDefinition.X_POSITION)],
reverted_observed_state[int(StateDefinition.Y_POSITION)],
reverted_observed_state[int(StateDefinition.THETA_POSITION)],
reverted_observed_state[int(StateDefinition.VEL_POSITION)],
action, int(is_demo)))
df = pd.DataFrame(ego_world_states,
columns=[
'pos_x', 'pos_y', 'orientation', 'velocity',
'action', 'is_demo'
])
print(df.head(10))
df.to_pickle(os.path.join(demo_root, "demonstrations/demo_dataframe"))
return
def __init__(self,
params=None,
num_scenarios=25,
random_seed=0,
mode="dense"):
ml_behavior = BehaviorDiscreteMacroActionsML(params)
HighwayBlueprint.__init__(self,
params=params,
num_scenarios=num_scenarios,
random_seed=random_seed,
ml_behavior=ml_behavior,
viewer=True,
mode=mode)
def test_env(self):
map_filename = "external/bark_ml_project/bark_ml/environments/blueprints/highway/city_highway_straight.xodr"
params = ParameterServer()
behavior = BehaviorDiscreteMacroActionsML(params)
evaluator = GoalReached(params)
observer = NearestAgentsObserver(params)
viewer = MPViewer(params=params,
x_range=[-35, 35],
y_range=[-35, 35],
follow_agent_id=True)
env = HyDiscreteHighway(params=params,
map_filename=map_filename,
behavior=behavior,
evaluator=evaluator,
observer=observer,
viewer=viewer,
render=True)
env.reset()
actions = [5] * 100
print(actions)
for action in actions:
env.step(action)
time.sleep(0.2)
return
def main():
print("Experiment server at:", os.getcwd())
params = ParameterServer(
filename=
"hy-exp-run.runfiles/hythe/configuration/params/default_exp_runne_params.json"
)
params = configure_params(params)
num_scenarios = 5
random_seed = 0
behavior = BehaviorDiscreteMacroActionsML(params)
# evaluator = GoalReachedGuiding(params)
evaluator = GoalReached(params)
observer = NearestAgentsObserver(params)
viewer = MPViewer(params=params,
x_range=[-35, 35],
y_range=[-35, 35],
follow_agent_id=True)
experiment_id = params["Experiment"]["random_seed"]
params_filename = "./{}_default_exp_runner_params.json".format(
experiment_id)
print(params_filename)
params.Save(filename=params_filename)
# database creation
dbs = DatabaseSerializer(
test_scenarios=2, test_world_steps=2, num_serialize_scenarios=20
) # increase the number of serialize scenarios to 100
dbs.process("hy-exp-run.runfiles/hythe/configuration/database")
local_release_filename = dbs.release(version="test",
sub_dir="hy_bark_packaged_databases")
db = BenchmarkDatabase(database_root=local_release_filename)
# switch this to other generator to get other index
# scenario_generator, _, _ = db.get_scenario_generator(0)
scenario_generator, _, _ = db.get_scenario_generator(1)
#
# env = GymSingleAgentRuntime(ml_behavior = behavior,
# observer = observer,
# evaluator = evaluator,
# step_time=0.2,
# viewer=viewer,
# scenario_generator=scenario_generator,
# render=False)
env = HyDiscreteHighway(params=params,
scenario_generation=scenario_generator,
behavior=behavior,
evaluator=evaluator,
observer=observer,
viewer=viewer,
render=False)
run(params, env)
# video_renderer.drawWorld(env._world)
env._viewer.export_video("./test_video")
# from gym.envs.registration import register
# register(
# id='highway-v1',
# entry_point='bark_ml.environments.gym:DiscreteHighwayGym'
# )
# import gym
# env = gym.make("highway-v1")
# env.reset()
# actions = [5]*100
# print(actions)
# for action in actions:
# env.step(action)
# time.sleep(0.2)
return
def main():
args = configure_args()
if is_local:
dir_prefix = ""
else:
dir_prefix = "hy-x-iqn-beliefs.runfiles/hythe/"
print("Experiment server at :", os.getcwd())
params = ParameterServer(filename=os.path.join(
dir_prefix, "configuration/params/iqn_params.json"))
params = configure_params(params, seed=args.jobname)
experiment_id = params["Experiment"]["random_seed"]
params_filename = os.path.join(params["Experiment"]["dir"],
"params_{}.json".format(experiment_id))
params_behavior_filename = os.path.join(
params["Experiment"]["dir"],
"behavior_params_{}.json".format(experiment_id))
# check if exp exists and handle preemption
exp_exists = check_if_exp_exists(params)
if exp_exists:
print("Loading existing experiment from: {}".format(
args.jobname, (params["Experiment"]["dir"])))
if os.path.isfile(params_filename):
params = ParameterServer(filename=params_filename,
log_if_default=True)
if os.path.isfile(params_behavior_filename):
params_behavior = ParameterServer(
filename=params_behavior_filename, log_if_default=True)
else:
Path(params["Experiment"]["dir"]).mkdir(parents=True, exist_ok=True)
params["ML"]["BaseAgent"]["SummaryPath"] = os.path.join(
params["Experiment"]["dir"], "agent/summaries")
params["ML"]["BaseAgent"]["CheckpointPath"] = os.path.join(
params["Experiment"]["dir"], "agent/checkpoints")
params_behavior = ParameterServer(filename=os.path.join(
dir_prefix, "configuration/params/1D_desired_gap_no_prior.json"),
log_if_default=True)
params.Save(filename=params_filename)
params_behavior.Save(filename=params_behavior_filename)
# configure belief observer
splits = 2
behavior_space = configure_behavior_space(params_behavior)
hypothesis_set, hypothesis_params = behavior_space.create_hypothesis_set_fixed_split(
split=splits)
observer = BeliefObserver(params, hypothesis_set, splits=splits)
behavior = BehaviorDiscreteMacroActionsML(params)
evaluator = GoalReached(params)
viewer = MPViewer(params=params,
x_range=[-35, 35],
y_range=[-35, 35],
follow_agent_id=True)
# database creation
dbs = DatabaseSerializer(test_scenarios=2,
test_world_steps=2,
num_serialize_scenarios=num_scenarios)
dbs.process(os.path.join(dir_prefix, "configuration/database"),
filter_sets="**/**/interaction_merging_mid_dense_1D_new.json")
local_release_filename = dbs.release(version="test")
db = BenchmarkDatabase(database_root=local_release_filename)
scenario_generator, _, _ = db.get_scenario_generator(scenario_set_id=0)
env = HyDiscreteHighway(params=params,
scenario_generation=scenario_generator,
behavior=behavior,
evaluator=evaluator,
observer=observer,
viewer=viewer,
render=is_local)
print('Observation/state space size', env.observation_space)
run(params, env, exp_exists)
params.Save(filename=params_filename)
logging.info('-' * 60)
logging.info("Writing params to :{}".format(params_filename))
logging.info('-' * 60)
params_behavior.Save(filename=params_behavior_filename)
logging.info('-' * 60)
logging.info(
"Writing behavior params to :{}".format(params_behavior_filename))
logging.info('-' * 60)
return
def main():
args = configure_args()
if is_local:
dir_prefix = ""
else:
dir_prefix = "hy-iqnfd-beliefs-exp.runfiles/hythe/"
print("Executing job :", args.jobname)
print("Experiment server at :", os.getcwd())
params = ParameterServer(filename=os.path.join(
dir_prefix, "configuration/params/iqn_params_demo_full_local.json"),
log_if_default=True)
params = configure_params(params, seed=args.jobname)
experiment_id = params["Experiment"]["random_seed"]
params_filename = os.path.join(params["Experiment"]["dir"],
"params_{}.json".format(experiment_id))
behavior = BehaviorDiscreteMacroActionsML(params)
evaluator = GoalReached(params)
params_behavior_filename = os.path.join(
params["Experiment"]["dir"],
"behavior_params_{}.json".format(experiment_id))
params_behavior = ParameterServer(filename=os.path.join(
dir_prefix, "configuration/params/1D_desired_gap_no_prior.json"),
log_if_default=True)
params_behavior.Save(filename=params_behavior_filename)
# configure belief observer
splits = 2
behavior_space = BehaviorSpace(params_behavior)
hypothesis_set, hypothesis_params = behavior_space.create_hypothesis_set_fixed_split(
split=splits)
observer = BeliefObserver(params, hypothesis_set, splits=splits)
viewer = MPViewer(params=params,
x_range=[-35, 35],
y_range=[-35, 35],
follow_agent_id=True)
# extract params and save experiment parameters
params["ML"]["BaseAgent"]["SummaryPath"] = os.path.join(
params["Experiment"]["dir"], "agent/summaries")
params["ML"]["BaseAgent"]["CheckpointPath"] = os.path.join(
params["Experiment"]["dir"], "agent/checkpoints")
params.Save(filename=params_filename)
logging.info('-' * 60)
logging.info("Writing params to :{}".format(params_filename))
logging.info('-' * 60)
# database creation
dbs = DatabaseSerializer(test_scenarios=1,
test_world_steps=2,
num_serialize_scenarios=num_scenarios)
dbs.process(os.path.join(dir_prefix, "configuration/database"),
filter_sets="**/**/interaction_merging_light_dense_1D.json")
local_release_filename = dbs.release(version="test")
db = BenchmarkDatabase(database_root=local_release_filename)
scenario_generator, _, _ = db.get_scenario_generator(0)
env = HyDiscreteHighway(params=params,
scenario_generation=scenario_generator,
behavior=behavior,
evaluator=evaluator,
observer=observer,
viewer=viewer,
render=False)
scenario, _ = scenario_generator.get_next_scenario()
world = scenario.GetWorldState()
observer.Reset(world)
assert env.action_space._n == 8, "Action Space is incorrect!"
run(args, params, env, db=db)
params.Save(params_filename)
logging.info('-' * 60)
logging.info("Writing params to :{}".format(params_filename))
logging.info('-' * 60)
params_behavior.Save(filename=params_behavior_filename)
logging.info('-' * 60)
logging.info(
"Writing behavior params to :{}".format(params_behavior_filename))
logging.info('-' * 60)
def main():
map_filename = "external/bark_ml_project/bark_ml/environments/blueprints/highway/city_highway_straight.xodr"
params = ParameterServer(filename=os.path.join(
exp_root, "params_iqn_pre_exp_2sdi64.json"),
log_if_default=True)
params_behavior = ParameterServer(filename=os.path.join(
exp_root, "behavior_params_iqn_pre_exp_2sdi64.json"))
behavior = BehaviorDiscreteMacroActionsML(params)
evaluator = GoalReached(params)
if is_belief_observer:
splits = 2
behavior_space = BehaviorSpace(params_behavior)
hypothesis_set, hypothesis_params = behavior_space.create_hypothesis_set_fixed_split(
split=splits)
observer = BeliefObserver(params, hypothesis_set, splits=splits)
else:
observer = NearestAgentsObserver(params)
viewer = MPViewer(params=params,
x_range=[-35, 35],
y_range=[-35, 35],
follow_agent_id=True)
# database creation
dbs1 = DatabaseSerializer(test_scenarios=1,
test_world_steps=2,
num_serialize_scenarios=1)
dbs1.process(os.path.join("", "configuration/database"),
filter_sets="**/**/interaction_merging_light_dense_1D.json")
local_release_filename = dbs1.release(version="lfd_offline")
db1 = BenchmarkDatabase(database_root=local_release_filename)
scenario_generator1, _, _ = db1.get_scenario_generator(0)
env = HyDiscreteHighway(params=params,
scenario_generation=scenario_generator1,
map_filename=map_filename,
behavior=behavior,
evaluator=evaluator,
observer=observer,
viewer=viewer,
render=False)
scenario, _ = scenario_generator1.get_next_scenario()
world = scenario.GetWorldState()
observer.Reset(world)
env.reset()
agent = pick_agent(exp_root, env, params)
ego_world_states = []
memory = agent.memory
learned_states = memory["state"]
actions = memory['action']
is_demos = memory['is_demo']
for state, action, is_demo in zip(learned_states, actions, is_demos):
ego_state = np.zeros((observer._len_ego_state + 1))
ego_nn_input_state = deepcopy(state[0:observer._len_ego_state])
ego_state[1:] = ego_nn_input_state
reverted_observed_state = observer.rev_observe_for_ego_vehicle(
ego_state)
ego_world_states.append(
(reverted_observed_state[int(StateDefinition.X_POSITION)],
reverted_observed_state[int(StateDefinition.Y_POSITION)],
reverted_observed_state[int(StateDefinition.THETA_POSITION)],
reverted_observed_state[int(StateDefinition.VEL_POSITION)],
action[0], is_demo[0]))
df = pd.DataFrame(ego_world_states,
columns=[
'pos_x', 'pos_y', 'orientation', 'velocity',
'action', 'is_demo'
])
print(df.head(10))
if not os.path.exists(os.path.join(exp_root, "demonstrations")):
os.makedirs(os.path.join(exp_root, "demonstrations"))
df.to_pickle(os.path.join(exp_root, "demonstrations", "learned_dataframe"))
return
def main():
print("Experiment server at :", os.getcwd())
args = configure_args()
#load exp params
exp_dir = args.output_dir
params_filename = glob.glob(os.path.join(exp_dir, "params_[!behavior]*"))
params = ParameterServer(filename=params_filename[0])
params.load(fn=params_filename[0])
params["ML"]["BaseAgent"]["SummaryPath"] = os.path.join(exp_dir, "agent/summaries")
params["ML"]["BaseAgent"]["CheckpointPath"] = os.path.join(exp_dir, "agent/checkpoints")
splits = 8
behavior_params_filename = glob.glob(os.path.join(exp_dir, "behavior_params*"))
if behavior_params_filename:
params_behavior = ParameterServer(filename=behavior_params_filename[0])
else:
params_behavior = ParameterServer(filename="configuration/params/1D_desired_gap_no_prior.json")
behavior_space = configure_behavior_space(params_behavior)
hypothesis_set, hypothesis_params = behavior_space.create_hypothesis_set_fixed_split(split=splits)
observer = BeliefObserver(params, hypothesis_set, splits=splits)
behavior = BehaviorDiscreteMacroActionsML(params_behavior)
evaluator = GoalReached(params)
viewer = MPViewer(params=params,
x_range=[-35, 35],
y_range=[-35, 35],
follow_agent_id=True)
# database creation
dir_prefix = ""
dbs = DatabaseSerializer(test_scenarios=2, test_world_steps=2,
num_serialize_scenarios=10)
dbs.process(os.path.join(dir_prefix, "configuration/database"), filter_sets="**/**/interaction_merging_light_dense_1D.json")
local_release_filename = dbs.release(version="test")
db = BenchmarkDatabase(database_root=local_release_filename)
scenario_generator, _, _ = db.get_scenario_generator(0)
video_renderer = VideoRenderer(renderer=viewer, world_step_time=0.2)
env = HyDiscreteHighway(params=params,
scenario_generation=scenario_generator,
behavior=behavior,
evaluator=evaluator,
observer=observer,
viewer=video_renderer,
render=is_local)
# non-agent evaluation mode
num_steps = 100
num_samples = params_behavior["BehaviorSpace"]["Hypothesis"]["BehaviorHypothesisIDM"]["NumSamples"]
print("Steps, samples, splits", num_steps, num_samples, splits)
step = 1
env.reset()
threshold = observer.is_enabled_threshold
discretize = observer.is_discretize
beliefs_df = pd.DataFrame(columns=["Step", "Action", "Agent", "Beliefs", "HyNum"])
beliefs_orig_df = pd.DataFrame(columns=["Step", "Action", "Agent", "Beliefs", "HyNum"])
while step <= num_steps:
action = 5 #np.random.randint(0, behavior.action_space.n)
next_state, reward, done, info = env.step(action)
for agent, beliefs in observer.beliefs.items():
beliefs = np.asarray(beliefs)
oring = deepcopy(beliefs)
for i, belief in enumerate(oring):
beliefs_orig_df = beliefs_orig_df.append({"Step": step, "Action": action, "Agent": agent, "Beliefs": belief, "HyNum": i}, ignore_index=True)
if discretize:
beliefs = observer.discretize_beliefs(beliefs)
if threshold:
beliefs = observer.threshold_beliefs(beliefs)
for i, belief in enumerate(beliefs):
beliefs_df = beliefs_df.append({"Step": step, "Action": action, "Agent": agent, "Beliefs": belief, "HyNum": i}, ignore_index=True)
step += 1
suffix = "switch"
if threshold:
suffix += "_threshold"
if discretize:
suffix += "_discretize"
beliefs_data_filename = "beliefs_{}_{}_{}".format(splits, num_samples, num_steps)
beliefs_data_filename += suffix
print(beliefs_data_filename)
beliefs_df.to_pickle(os.path.join(str(Path.home()), "master_thesis/code/hythe-src/beliefs_data/", beliefs_data_filename))
beliefs_data_filename = "orig_beliefs_{}_{}_{}".format(splits, num_samples, num_steps)
beliefs_data_filename += suffix
print(beliefs_data_filename)
beliefs_orig_df.to_pickle(os.path.join(str(Path.home()), "master_thesis/code/hythe-src/beliefs_data/", beliefs_data_filename))
video_filename = os.path.join(str(Path.home()), "master_thesis/code/hythe-src/beliefs_data/", "video_{}".format(num_samples))
print(video_filename)
video_filename += suffix
print(video_filename)
video_renderer.export_video(filename=video_filename)
return
def main():
args = configure_args()
if is_local:
dir_prefix = ""
else:
dir_prefix = "hy-iqn-exp.runfiles/hythe/"
print("Executing job :", args.jobname)
print("Experiment server at :", os.getcwd())
params = ParameterServer(filename=os.path.join(
dir_prefix, "configuration/params/iqn_params.json"),
log_if_default=True)
params = configure_params(params, seed=args.jobname)
experiment_id = params["Experiment"]["random_seed"]
params_filename = os.path.join(params["Experiment"]["dir"],
"params_{}.json".format(experiment_id))
# check if exp exists and handle preemption
exp_exists = check_if_exp_exists(params)
if exp_exists:
print("Loading existing experiment from: {}".format(
args.jobname, (params["Experiment"]["dir"])))
if os.path.isfile(params_filename):
params = ParameterServer(filename=params_filename,
log_if_default=True)
else:
Path(params["Experiment"]["dir"]).mkdir(parents=True, exist_ok=True)
behavior = BehaviorDiscreteMacroActionsML(params)
evaluator = GoalReached(params)
observer = NearestAgentsObserver(params)
viewer = MPViewer(params=params,
x_range=[-35, 35],
y_range=[-35, 35],
follow_agent_id=True)
# extract params and save experiment parameters
params["ML"]["BaseAgent"]["SummaryPath"] = os.path.join(
params["Experiment"]["dir"], "agent/summaries")
params["ML"]["BaseAgent"]["CheckpointPath"] = os.path.join(
params["Experiment"]["dir"], "agent/checkpoints")
params.Save(filename=params_filename)
logging.info('-' * 60)
logging.info("Writing params to :{}".format(params_filename))
logging.info('-' * 60)
# database creation
dbs = DatabaseSerializer(test_scenarios=2,
test_world_steps=2,
num_serialize_scenarios=num_scenarios)
dbs.process(os.path.join(dir_prefix, "configuration/database"),
filter_sets="**/**/interaction_merging_light_dense_1D.json")
local_release_filename = dbs.release(version="test")
db = BenchmarkDatabase(database_root=local_release_filename)
scenario_generator, _, _ = db.get_scenario_generator(0)
env = HyDiscreteHighway(params=params,
scenario_generation=scenario_generator,
behavior=behavior,
evaluator=evaluator,
observer=observer,
viewer=viewer,
render=False)
assert env.action_space._n == 8, "Action Space is incorrect!"
run(params, env, exp_exists)
params.Save(params_filename)
logging.info('-' * 60)
logging.info("Writing params to :{}".format(params_filename))
logging.info('-' * 60)
def __init__(self, env, test_env, params, bark_behavior=None):
BehaviorModel.__init__(self, params)
self._params = params
self.env = env
self.test_env = test_env
self._bark_behavior_model = bark_behavior or BehaviorDiscreteMacroActionsML(
params)
self.device = torch.device("cuda" if self._params["ML"]["BaseAgent"][
"Cuda", "", True] and torch.cuda.is_available() else "cpu")
self.online_net = None
self.target_net = None
self.steps = 0
self.learning_steps = 0
self.episodes = 0
self.best_eval_score = -np.inf
self.num_actions = self.env.action_space.n
self.num_steps = self._params["ML"]["BaseAgent"]["NumSteps", "",
5000000]
self.batch_size = self._params["ML"]["BaseAgent"]["BatchSize", "", 32]
self.double_q_learning = self._params["ML"]["BaseAgent"][
"Double_q_learning", "", False]
self.dueling_net = self._params["ML"]["BaseAgent"]["DuelingNet", "",
False]
self.noisy_net = self._params["ML"]["BaseAgent"]["NoisyNet", "", False]
self.use_per = self._params["ML"]["BaseAgent"]["Use_per", "", False]
self.reward_log_interval = self._params["ML"]["BaseAgent"][
"RewardLogInterval", "", 5]
self.summary_log_interval = self._params["ML"]["BaseAgent"][
"SummaryLogInterval", "", 100]
self.eval_interval = self._params["ML"]["BaseAgent"]["EvalInterval",
"", 25000]
self.num_eval_steps = self._params["ML"]["BaseAgent"]["NumEvalSteps",
"", 12500]
self.gamma_n = \
self._params["ML"]["BaseAgent"]["Gamma", "", 0.99] ** \
self._params["ML"]["BaseAgent"]["Multi_step", "", 1]
self.start_steps = self._params["ML"]["BaseAgent"]["StartSteps", "",
5000]
self.epsilon_train = LinearAnneaer(
1.0, self._params["ML"]["BaseAgent"]["EpsilonTrain", "", 0.01],
self._params["ML"]["BaseAgent"]["EpsilonDecaySteps", "", 25000])
self.epsilon_eval = self._params["ML"]["BaseAgent"]["EpsilonEval", "",
0.001]
self.update_interval = \
self._params["ML"]["BaseAgent"]["Update_interval", "", 4]
self.target_update_interval = self._params["ML"]["BaseAgent"][
"TargetUpdateInterval", "", 5000]
self.max_episode_steps = \
self._params["ML"]["BaseAgent"]["MaxEpisodeSteps", "", 10000]
self.grad_cliping = self._params["ML"]["BaseAgent"]["GradCliping", "",
5.0]
self.summary_dir = \
self._params["ML"]["BaseAgent"]["SummaryPath", "", ""]
self.model_dir = \
self._params["ML"]["BaseAgent"]["CheckpointPath", "", ""]
if not os.path.exists(self.model_dir) and self.model_dir:
os.makedirs(self.model_dir)
if not os.path.exists(self.summary_dir) and self.summary_dir:
os.makedirs(self.summary_dir)
self.writer = SummaryWriter(log_dir=self.summary_dir)
self.train_return = RunningMeanStats(self.summary_log_interval)
# NOTE: by default we do not want the action to be set externally
# as this enables the agents to be plug and played in BARK.
self._set_action_externally = False
# Replay memory which is memory-efficient to store stacked frames.
if self.use_per:
beta_steps = (self.num_steps - self.start_steps) / \
self.update_interval
self.memory = LazyPrioritizedMultiStepMemory(
self._params["ML"]["BaseAgent"]["MemorySize", "", 10**6],
self.env.observation_space.shape,
self.device,
self._params["ML"]["BaseAgent"]["Gamma", "", 0.99],
self._params["ML"]["BaseAgent"]["Multi_step", "", 1],
beta_steps=beta_steps)
else:
self.memory = LazyMultiStepMemory(
self._params["ML"]["BaseAgent"]["MemorySize", "", 10**6],
self.env.observation_space.shape, self.device,
self._params["ML"]["BaseAgent"]["Gamma", "", 0.99],
self._params["ML"]["BaseAgent"]["Multi_step", "", 1])
def test_discrete_behavior(self):
params = ParameterServer()
discrete_behavior = BehaviorDiscreteMacroActionsML(params)
# sets 0-th motion primitive active
discrete_behavior.ActionToBehavior(0)
print(discrete_behavior.action_space)
dbs.process("configuration/database", filter_sets="**/**/interaction_merging_light_dense_1D.json")
local_release_filename = dbs.release(version="resume")
db = BenchmarkDatabase(database_root=local_release_filename)
scenario_generator, _, _ = db.get_scenario_generator(0)
# load belief observer specifics
if is_belief_observer:
splits = 2
behavior_params_filename = glob.glob(os.path.join(exp_dir, "behavior_params*"))[0]
params_behavior = ParameterServer(filename=behavior_params_filename, log_if_default=True)
behavior_space = BehaviorSpace(params_behavior)
hypothesis_set, hypothesis_params = behavior_space.create_hypothesis_set_fixed_split(split=splits)
observer = BeliefObserver(params, hypothesis_set, splits=splits)
behavior = BehaviorDiscreteMacroActionsML(params_behavior)
# if not, load default observer
else:
behavior = BehaviorDiscreteMacroActionsML(params)
observer = NearestAgentsObserver(params)
evaluator = GoalReached(params)
viewer = MPViewer(
params=params,
center= [960, 1000.8],
enforce_x_length=True,
x_length = 100.0,
use_world_bounds=False)
# load env