def __init__(self):
env_input_size: int = 2
super().__init__(env_input_size)
self.post_fn_remote = self.post_milp
self.get_nn_fn = self.get_nn
self.plot_fn = self.plot
self.assign_lbl_fn = self.assign_label
self.additional_seen_fn = self.additional_seen
self.template_2d: np.ndarray = np.array([[1, 0], [0, 1]])
input_boundaries, input_template = self.get_template(0)
self.input_boundaries: List = input_boundaries
self.input_template: np.ndarray = input_template
# _, template = self.get_template(1)
template = Experiment.octagon(env_input_size)
self.plotting_time_interval = 60 * 5
self.analysis_template: np.ndarray = template
self.safe_angle = 30 * 2 * math.pi / 360
theta = [self.e(env_input_size, 0)]
neg_theta = [-self.e(env_input_size, 0)]
# battery = [self.e(env_input_size, 4)]
self.unsafe_zone: List[Tuple] = [(theta, np.array([-self.safe_angle])),
(neg_theta,
np.array([-self.safe_angle]))]
epsilon = 1e-4
self.angle_split: List[Tuple] = [
(theta, np.array([self.safe_angle - epsilon])),
(neg_theta, np.array([self.safe_angle - epsilon]))
]
self.use_rounding = False
self.rounding_value = 1024
self.time_horizon = 300
self.nn_path = "/home/edoardo/ray_results/tune_PPO_pendulum/PPO_MonitoredPendulum_035b5_00000_0_2021-05-11_11-59-52/checkpoint_3333/checkpoint-3333"
self.tau = 0.02
self.n_actions = 3
def generate_nn_polyhedral_guard(self, nn, chosen_action, output_flag):
gurobi_model = grb.Model()
gurobi_model.setParam('OutputFlag', output_flag)
gurobi_model.setParam('Threads', 2)
observation = gurobi_model.addMVar(shape=(2, ),
lb=float("-inf"),
ub=float("inf"),
name="observation")
Experiment.generate_nn_guard(gurobi_model,
observation,
nn,
action_ego=chosen_action)
observable_template = Experiment.octagon(2)
# self.env_input_size = 2
observable_result = optimise(observable_template, gurobi_model,
observation)
# self.env_input_size = 6
return observable_template, observable_result
def additional_seen(self):
# adds an element that captures all the states where battery is <=0
gurobi_model = grb.Model()
gurobi_model.setParam('OutputFlag', self.output_flag)
additional = [
self.safe_angle, self.safe_angle,
float("inf"),
float("inf"),
float("inf"),
float("inf"),
float("inf"),
float("inf")
]
input = generate_input_region(gurobi_model,
Experiment.octagon(self.env_input_size),
additional, self.env_input_size)
x_results = optimise(self.analysis_template, gurobi_model, input)
if x_results is None:
print("Model unsatisfiable")
return None
root = tuple(x_results)
return [(root, 0)]
def run_parameterised_experiment(config):
# Hyperparameters
trial_dir = tune.get_trial_dir()
problem, method, other_config = config["main_params"]
n_workers = config["n_workers"]
experiment = CartpoleExperiment()
experiment.nn_path = other_config[
"folder"] # nn_paths_cartpole[other_config["nn_path"]]
experiment.tau = other_config["tau"]
if other_config["template"] == 2: # octagon
experiment.analysis_template = Experiment.octagon(
experiment.env_input_size)
elif other_config["template"] == 0: # box
experiment.analysis_template = Experiment.box(
experiment.env_input_size)
else:
_, template = experiment.get_template(1)
experiment.analysis_template = template # standard
experiment.n_workers = n_workers
experiment.show_progressbar = False
experiment.show_progress_plot = False
# experiment.use_rounding = False
experiment.save_dir = trial_dir
experiment.update_progress_fn = update_progress
elapsed_seconds, safe, max_t = experiment.run_experiment()
safe_value = 0
if safe is None:
safe_value = 0
elif safe:
safe_value = 1
elif not safe:
safe_value = -1
tune.report(elapsed_seconds=elapsed_seconds,
safe=safe_value,
max_t=max_t,
done=True)
policy = trainer.get_policy()
# sequential_nn = convert_ray_simple_policy_to_sequential(policy).cpu()
sequential_nn = convert_ray_policy_to_sequential(policy).cpu()
# l0 = torch.nn.Linear(5, 3, bias=False)
# l0.weight = torch.nn.Parameter(torch.tensor([[0, 0, 1, 0, 0], [0, 0, 0, 1, 0], [0, 0, 0, 0, 1]], dtype=torch.float32))
layers = []
for l in sequential_nn:
layers.append(l)
nn = torch.nn.Sequential(*layers)
torch.save(nn, pickled_path)
# ray.shutdown()
return nn
if __name__ == '__main__':
os.environ["CUDA_VISIBLE_DEVICES"] = "0"
ray.init(log_to_driver=False, local_mode=False)
experiment = PendulumExperimentProbabilistic()
experiment.save_dir = "/home/edoardo/ray_results/tune_PPO_pendulum/octagon_psi05_h7_nocontain2"
experiment.plotting_time_interval = 60
experiment.show_progressbar = True
experiment.show_progress_plot = False
experiment.analysis_template = Experiment.octagon(2)
experiment.use_contained = False
experiment.use_split_with_seen = False
# experiment.max_probability_split = 0.5
experiment.time_horizon = 6
# experiment.load_graph = True
experiment.run_experiment()
def get_experiment_instance(config, trial_dir):
# Hyperparameters
problem, other_config = config["main_params"]
n_workers = config["n_workers"]
if problem == "bouncing_ball":
experiment = BouncingBallExperimentProbabilistic()
experiment.nn_path = os.path.join(utils.get_agents_dir(), nn_paths_bouncing_ball[other_config["nn_path"]])
assert os.path.exists(experiment.nn_path)
experiment.use_contained = other_config["use_contain"]
if other_config["initial_state"] == 0:
experiment.input_boundaries = [9, -5, 0, 0.1]
elif other_config["initial_state"] == 1:
experiment.input_boundaries = [9, -5, 1, 1]
else:
raise NotImplementedError()
if other_config["template"] == 1: # octagon
experiment.analysis_template = Experiment.octagon(experiment.env_input_size)
elif other_config["template"] == 0: # box
experiment.analysis_template = Experiment.box(experiment.env_input_size)
else:
raise NotImplementedError()
experiment.max_probability_split = other_config["phi"]
experiment.n_workers = n_workers
experiment.show_progressbar = False
experiment.show_progress_plot = False
experiment.save_dir = trial_dir
# experiment.update_progress_fn = update_progress
elif problem == "stopping_car":
experiment = StoppingCarExperimentProbabilistic()
experiment.nn_path = os.path.join(utils.get_agents_dir(), nn_paths_stopping_car[other_config["nn_path"]])
experiment.max_probability_split = other_config["phi"]
if other_config["template"] == 2: # octagon
experiment.analysis_template = Experiment.octagon(experiment.env_input_size)
elif other_config["template"] == 1:
delta_x = Experiment.e(experiment.env_input_size, 0)
v_ego = Experiment.e(experiment.env_input_size, 1)
template = np.array([delta_x, -delta_x, v_ego, -v_ego, 1 / 4.5 * delta_x + v_ego, 1 / 4.5 * delta_x - v_ego, -1 / 4.5 * delta_x + v_ego,
-1 / 4.5 * delta_x - v_ego])
experiment.analysis_template = template
elif other_config["template"] == 0: # box
experiment.analysis_template = Experiment.box(experiment.env_input_size)
else:
raise NotImplementedError()
experiment.n_workers = n_workers
experiment.show_progressbar = False
experiment.show_progress_plot = False
experiment.save_dir = trial_dir
# experiment.update_progress_fn = update_progress
else:
experiment = PendulumExperimentProbabilistic()
experiment.nn_path = os.path.join(utils.get_agents_dir(), nn_paths_cartpole[other_config["nn_path"]])
if other_config["template"] == 1: # octagon
experiment.analysis_template = Experiment.octagon(experiment.env_input_size)
elif other_config["template"] == 0: # box
experiment.analysis_template = Experiment.box(experiment.env_input_size)
else:
raise NotImplementedError()
experiment.max_probability_split = other_config["phi"]
experiment.n_workers = n_workers
experiment.show_progressbar = False
experiment.show_progress_plot = False
# experiment.use_rounding = False
experiment.save_dir = trial_dir
# experiment.update_progress_fn = update_progress
return experiment
def run_parameterised_experiment(config):
# Hyperparameters
trial_dir = tune.get_trial_dir()
problem, method, other_config = config["main_params"]
n_workers = config["n_workers"]
if problem == "bouncing_ball":
experiment = BouncingBallExperiment()
experiment.nn_path = os.path.join(utils.get_save_dir(), nn_paths_bouncing_ball[other_config["nn_path"]])
experiment.tau = other_config["tau"]
if other_config["template"] == 2: # octagon
experiment.analysis_template = Experiment.octagon(experiment.env_input_size)
elif other_config["template"] == 0: # box
experiment.analysis_template = Experiment.box(experiment.env_input_size)
else:
raise NotImplementedError()
experiment.n_workers = n_workers
experiment.show_progressbar = False
experiment.show_progress_plot = False
experiment.save_dir = trial_dir
experiment.update_progress_fn = update_progress
elapsed_seconds, safe, max_t = experiment.run_experiment()
elif problem == "stopping_car":
if method == "ora":
experiment = ORAStoppingCarExperiment()
experiment.nn_path = os.path.join(utils.get_save_dir(), nn_paths_stopping_car[other_config["nn_path"]])
experiment.input_epsilon = other_config["epsilon_input"]
if other_config["template"] == 2: # octagon
experiment.analysis_template = Experiment.octagon(experiment.env_input_size)
elif other_config["template"] == 0: # box
experiment.analysis_template = Experiment.box(experiment.env_input_size)
else:
_, template = experiment.get_template(1)
experiment.analysis_template = template # standard
experiment.n_workers = n_workers
experiment.show_progressbar = False
experiment.show_progress_plot = False
experiment.save_dir = trial_dir
experiment.update_progress_fn = update_progress
elapsed_seconds, safe, max_t = experiment.run_experiment()
else:
experiment = StoppingCarExperiment()
experiment.nn_path = os.path.join(utils.get_save_dir(), nn_paths_stopping_car[other_config["nn_path"]])
experiment.input_epsilon = other_config["epsilon_input"]
if other_config["template"] == 2: # octagon
experiment.analysis_template = Experiment.octagon(experiment.env_input_size)
elif other_config["template"] == 0: # box
experiment.analysis_template = Experiment.box(experiment.env_input_size)
else:
_, template = experiment.get_template(1)
experiment.analysis_template = template # standard
experiment.n_workers = n_workers
experiment.show_progressbar = False
experiment.show_progress_plot = False
experiment.save_dir = trial_dir
experiment.update_progress_fn = update_progress
elapsed_seconds, safe, max_t = experiment.run_experiment()
else:
experiment = CartpoleExperiment()
experiment.nn_path = os.path.join(utils.get_save_dir(), nn_paths_cartpole[other_config["nn_path"]])
experiment.tau = other_config["tau"]
if other_config["template"] == 2: # octagon
experiment.analysis_template = Experiment.octagon(experiment.env_input_size)
elif other_config["template"] == 0: # box
experiment.analysis_template = Experiment.box(experiment.env_input_size)
else:
_, template = experiment.get_template(1)
experiment.analysis_template = template # standard
experiment.n_workers = n_workers
experiment.show_progressbar = False
experiment.show_progress_plot = False
# experiment.use_rounding = False
experiment.save_dir = trial_dir
experiment.update_progress_fn = update_progress
elapsed_seconds, safe, max_t = experiment.run_experiment()
safe_value = 0
if safe is None:
safe_value = 0
elif safe:
safe_value = 1
elif not safe:
safe_value = -1
tune.report(elapsed_seconds=elapsed_seconds, safe=safe_value, max_t=max_t, done=True)