bo 2d finished - now testing

This commit is contained in:
Niko Feith 2023-08-28 15:58:36 +02:00
parent 33b8093a49
commit 597579bd98
4 changed files with 99 additions and 113 deletions

View File

@ -4,6 +4,7 @@ string metric
uint16 nr_weights
uint16 max_steps
uint16 nr_episodes
uint16 nr_dims
uint16 nr_runs
string acquisition_function
float32 metric_parameter

View File

@ -4,3 +4,4 @@ bool display_run
uint8 interactive_run
float64[] policy
float64[] weights
uint16 nr_dims

View File

@ -36,7 +36,7 @@ class ActiveBOTopic(Node):
self.active_bo_sub = self.create_subscription(ActiveBORequest,
'active_bo_request',
self.active_bo_callback,
self.bo_callback,
1, callback_group=bo_callback_group)
self.active_bo_pending = False
@ -63,7 +63,7 @@ class ActiveBOTopic(Node):
1, callback_group=rl_callback_group)
self.active_rl_sub = self.create_subscription(ActiveRLResponse,
'active_rl_response',
self.active_rl_callback,
self.rl_callback,
1, callback_group=rl_callback_group)
self.rl_pending = False
@ -105,6 +105,7 @@ class ActiveBOTopic(Node):
self.bo_metric = None
self.bo_fixed_seed = False
self.bo_nr_weights = None
self.bo_nr_dims = None
self.bo_steps = 0
self.bo_episodes = 0
self.bo_runs = 0
@ -120,7 +121,7 @@ class ActiveBOTopic(Node):
self.BO = None
self.overwrite = False
def active_bo_callback(self, msg):
def bo_callback(self, msg):
if not self.active_bo_pending:
# self.get_logger().info('Active Bayesian Optimization request pending!')
self.active_bo_pending = True
@ -128,6 +129,7 @@ class ActiveBOTopic(Node):
self.bo_metric = msg.metric
self.bo_fixed_seed = msg.fixed_seed
self.bo_nr_weights = msg.nr_weights
self.bo_nr_dims = msg.nr_dims
self.bo_steps = msg.max_steps
self.bo_episodes = msg.nr_episodes
self.bo_runs = msg.nr_runs
@ -135,19 +137,18 @@ class ActiveBOTopic(Node):
self.bo_metric_parameter = msg.metric_parameter
self.bo_metric_parameter_2 = msg.metric_parameter_2
self.save_result = msg.save_result
self.seed_array = np.zeros((1, self.bo_runs))
self.seed_array = np.zeros((self.bo_runs, 1))
self.overwrite = msg.overwrite
# initialize
self.reward = np.zeros((self.bo_episodes + self.nr_init - 1, self.bo_runs))
self.best_pol_reward = np.zeros((1, self.bo_runs))
self.best_policy = np.zeros((self.bo_steps, self.bo_runs))
self.best_weights = np.zeros((self.bo_nr_weights, self.bo_runs))
self.reward = np.zeros((self.bo_runs, self.bo_episodes + self.nr_init - 1))
self.best_pol_reward = np.zeros((self.bo_runs, 1))
self.best_policy = np.zeros((self.bo_runs, self.bo_steps, self.bo_nr_dims))
self.best_weights = np.zeros((self.bo_runs, self.bo_nr_weights, self.bo_nr_dims))
# set the seed
if self.bo_fixed_seed:
self.seed = int(np.random.randint(1, 2147483647, 1)[0])
# self.get_logger().info(str(self.seed))
else:
self.seed = None
@ -161,7 +162,7 @@ class ActiveBOTopic(Node):
self.rl_weights = None
self.rl_final_step = None
def active_rl_callback(self, msg):
def rl_callback(self, msg):
if self.rl_pending:
# self.get_logger().info('Active Reinforcement Learning response received!')
self.rl_weights = np.array(msg.weights, dtype=np.float64)
@ -201,7 +202,7 @@ class ActiveBOTopic(Node):
if self.BO is None:
self.BO = BayesianOptimization(self.bo_steps,
2,
self.bo_nr_dims,
self.bo_nr_weights,
acq=self.bo_acq_fcn)
@ -223,8 +224,9 @@ class ActiveBOTopic(Node):
rl_msg.seed = seed
rl_msg.display_run = False
rl_msg.interactive_run = 2
rl_msg.weights = self.BO.policy_model.random_policy().tolist()
rl_msg.policy = self.BO.policy_model.rollout().reshape(-1,).tolist()
rl_msg.weights = self.BO.policy_model.random_policy().flatten().tolist()
rl_msg.policy = self.BO.policy_model.rollout().flatten().tolist()
rl_msg.nr_dims = self.bo_nr_dims
self.active_rl_pub.publish(rl_msg)
@ -234,25 +236,15 @@ class ActiveBOTopic(Node):
bo_response = ActiveBOResponse()
best_policy_idx = np.argmax(self.best_pol_reward)
bo_response.best_policy = self.best_policy[:, best_policy_idx].tolist()
bo_response.best_weights = self.best_weights[:, best_policy_idx].tolist()
# self.get_logger().info(f'Best Policy: {self.best_pol_reward}')
self.get_logger().info(f'{best_policy_idx}, {int(self.seed_array[0, best_policy_idx])}')
bo_response.best_policy = self.best_policy[best_policy_idx, :, :].flatten().tolist()
bo_response.best_weights = self.best_weights[best_policy_idx, :, :].flatten().tolist()
bo_response.reward_mean = np.mean(self.reward, axis=1).tolist()
bo_response.reward_std = np.std(self.reward, axis=1).tolist()
if self.save_result:
if self.bo_env == "Mountain Car":
env = 'mc'
elif self.bo_env == "Cartpole":
env = 'cp'
elif self.bo_env == "Acrobot":
env = 'ab'
elif self.bo_env == "Pendulum":
env = 'pd'
if self.bo_env == "Reacher":
env = 're'
else:
raise NotImplementedError
@ -283,7 +275,6 @@ class ActiveBOTopic(Node):
if self.bo_fixed_seed:
seed = int(self.seed_array[0, best_policy_idx])
# self.get_logger().info(f'Used seed{seed}')
else:
seed = int(np.random.randint(1, 2147483647, 1)[0])
@ -291,8 +282,9 @@ class ActiveBOTopic(Node):
active_rl_request.seed = seed
active_rl_request.display_run = True
active_rl_request.interactive_run = 1
active_rl_request.policy = self.best_policy[:, best_policy_idx].tolist()
active_rl_request.weights = self.best_weights[:, best_policy_idx].tolist()
active_rl_request.policy = self.best_policy[best_policy_idx, :, :].flatten().tolist()
active_rl_request.weights = self.best_weights[best_policy_idx, :, :].flatten().tolist()
active_rl_request.nr_dims = self.bo_nr_dims
self.active_rl_pub.publish(active_rl_request)
@ -316,7 +308,7 @@ class ActiveBOTopic(Node):
user_query = ImprovementQuery(self.bo_metric_parameter,
self.bo_metric_parameter_2,
self.last_query,
self.reward[:self.current_episode, self.current_run])
self.reward[self.current_run, :self.current_episode])
elif self.bo_metric == "max acquisition":
user_query = MaxAcqQuery(self.bo_metric_parameter,
@ -347,30 +339,17 @@ class ActiveBOTopic(Node):
active_rl_request.seed = seed
active_rl_request.display_run = True
active_rl_request.interactive_run = 0
active_rl_request.policy = old_policy.tolist()
active_rl_request.weights = old_weights.tolist()
active_rl_request.policy = old_policy.flatten().tolist()
active_rl_request.weights = old_weights.flatten().tolist()
active_rl_request.nr_dims = self.bo_nr_dims
# self.get_logger().info('Calling: Active RL')
self.active_rl_pub.publish(active_rl_request)
self.rl_pending = True
else:
# if self.bo_acq_fcn == "Preference Expected Improvement":
# self.get_logger().info(f"{self.BO.acq_fun.proposal_mean}")
# self.get_logger().info(f"X: {self.BO.X}")
x_next = self.BO.next_observation()
# self.get_logger().info(f'x_next: {x_next}')
# self.get_logger().info(f'overwrite: {self.weight_preference}')
# self.get_logger().info(f'rl_weights: {self.rl_weights}')
if self.overwrite:
if self.weight_preference is not None and self.rl_weights is not None:
x_next[self.weight_preference] = self.rl_weights[self.weight_preference]
# self.get_logger().info(f'x_next: {x_next}')
# self.get_logger().info(f'overwrite: {self.weight_preference}')
# self.get_logger().info(f'rl_weights: {self.rl_weights}')
# self.get_logger().info('Next Observation BO!')
self.BO.policy_model.weights = np.around(x_next, decimals=8)
self.BO.policy_model.set_weights(np.around(x_next, decimals=8))
if self.bo_fixed_seed:
seed = self.seed
else:
@ -381,23 +360,23 @@ class ActiveBOTopic(Node):
rl_msg.seed = seed
rl_msg.display_run = False
rl_msg.interactive_run = 2
rl_msg.policy = self.BO.policy_model.rollout().reshape(-1,).tolist()
rl_msg.weights = x_next.tolist()
rl_msg.policy = self.BO.policy_model.rollout().flatten().tolist()
rl_msg.weights = x_next.flatten().tolist()
rl_msg.nr_dims = self.bo_nr_dims
self.rl_pending = True
self.active_rl_pub.publish(rl_msg)
self.reward[self.current_episode, self.current_run] = np.max(self.BO.Y)
self.reward[self.current_run, self.current_episode] = np.max(self.BO.Y)
self.get_logger().info(f'Current Episode: {self.current_episode},'
f' best reward: {self.reward[self.current_episode, self.current_run]}')
f' best reward: {self.reward[self.current_run, self.current_episode]}')
self.current_episode += 1
else:
self.best_policy[:, self.current_run], \
self.best_pol_reward[:, self.current_run], \
self.best_weights[:, self.current_run], idx = self.BO.get_best_result()
self.best_policy[self.current_run, :, :], \
self.best_pol_reward[self.current_run, :], \
self.best_weights[self.current_run, :, :], idx = self.BO.get_best_result()
if self.current_run < self.bo_runs - 1:
self.BO = None
@ -405,7 +384,8 @@ class ActiveBOTopic(Node):
self.current_episode = 0
self.last_query = 0
if self.bo_fixed_seed:
self.seed_array[0, self.current_run] = self.seed
self.seed_array[self.current_run, 0] = self.seed
else:
self.seed = int(np.random.randint(1, 2147483647, 1)[0])
# self.get_logger().info(f'{self.seed}')
self.current_run += 1
@ -443,4 +423,3 @@ def main(args=None):
if __name__ == '__main__':
main()

View File

@ -10,11 +10,7 @@ from rclpy.node import Node
from rclpy.callback_groups import ReentrantCallbackGroup
from active_bo_ros.ReinforcementLearning.ContinuousMountainCar import Continuous_MountainCarEnv
from active_bo_ros.ReinforcementLearning.CartPole import CartPoleEnv
from active_bo_ros.ReinforcementLearning.Pendulum import PendulumEnv
from active_bo_ros.ReinforcementLearning.Acrobot import AcrobotEnv
from dm_control import suite
import numpy as np
import time
@ -34,7 +30,7 @@ class ActiveRL(Node):
1, callback_group=rl_callback_group)
self.active_rl_sub = self.create_subscription(ActiveRLRequest,
'active_rl_request',
self.active_rl_callback,
self.rl_callback,
1, callback_group=rl_callback_group)
self.rl_env = None
@ -56,17 +52,20 @@ class ActiveRL(Node):
callback_group=topic_callback_group)
self.eval_sub = self.create_subscription(ActiveRLEvalResponse,
'active_rl_eval_response',
self.active_rl_eval_callback,
self.rl_eval_callback,
1,
callback_group=topic_callback_group)
self.eval_response_received = False
self.eval_policy = None
self.eval_weights = None
self.overwrite_weight = None
self.weight_preference = None
# RL Environments
self.env = None
self.rl_spec = None
self.rl_dims = None
self.pol_dims = None
# State Machine Variables
self.best_pol_shown = False
@ -83,61 +82,65 @@ class ActiveRL(Node):
def reset_rl_request(self):
self.rl_env = None
self.rl_spec = None
self.rl_seed = None
self.rl_policy = None
self.rl_weights = None
self.interactive_run = 0
self.display_run = False
self.rl_dims = None
self.pol_dims = None
def active_rl_callback(self, msg):
def rl_callback(self, msg):
self.rl_env = msg.env
self.rl_seed = msg.seed
self.display_run = msg.display_run
self.rl_policy = np.array(msg.policy, dtype=np.float64)
self.rl_dims = msg.nr_dims
self.rl_weights = msg.weights
pol = msg.policy
self.pol_dims = (len(pol)/self.rl_dims, self.rl_dims)
self.rl_policy = np.array(pol, dtype=np.float64).reshape(self.pol_dims)
self.interactive_run = msg.interactive_run
if self.rl_env == "Mountain Car":
self.env = Continuous_MountainCarEnv(render_mode="rgb_array")
elif self.rl_env == "Cartpole":
self.env = CartPoleEnv(render_mode="rgb_array")
elif self.rl_env == "Acrobot":
self.env = AcrobotEnv(render_mode="rgb_array")
elif self.rl_env == "Pendulum":
self.env = PendulumEnv(render_mode="rgb_array")
if self.rl_env == "Reacher":
random_state = np.random.RandomState(seed=self.rl_seed)
self.env = suite.load('reacher', 'hard', task_kwargs={'random': random_state})
self.rl_spec = self.env.action_spec()
self.env.reset()
else:
raise NotImplementedError
# self.get_logger().info('Active RL: Called!')
self.env.reset(seed=self.rl_seed)
self.rl_pending = True
self.policy_sent = False
self.rl_step = 0
def reset_eval_request(self):
self.eval_policy = None
self.eval_weights = None
def active_rl_eval_callback(self, msg):
self.eval_policy = np.array(msg.policy, dtype=np.float64)
def rl_eval_callback(self, msg):
self.eval_policy = np.array(msg.policy, dtype=np.float64).reshape(self.pol_dims)
self.eval_weights = msg.weights
self.overwrite_weight = msg.overwrite_weight
self.weight_preference = msg.weight_preference
self.get_logger().info('Active RL Eval: Responded!')
self.env.reset(seed=self.rl_seed)
self.env.reset()
self.eval_response_received = True
def next_image(self, policy, display_run):
action = policy[self.rl_step]
def step(self, policy, display_run):
done = False
action = policy[self.rl_step, :]
action_clipped = action.clip(min=-1.0, max=1.0)
output = self.env.step(action_clipped.astype(np.float64))
self.rl_reward += output[1]
done = output[2]
self.rl_step += 1
if output.reward != 0.0:
self.rl_reward += output.reward * 10
done = True
else:
self.rl_step -= 1.0
if display_run:
rgb_array = self.env.render()
rgb_array = self.env.physics.render(camera_id=0, height=400, width=600)
rgb_shape = rgb_array.shape
red = rgb_array[:, :, 0].flatten().tolist()
@ -160,25 +163,27 @@ class ActiveRL(Node):
return done
def complete_run(self, policy):
def runner(self, policy):
env_reward = 0.0
step_count = 0
done = False
self.env.reset(seed=self.rl_seed)
self.env.reset()
for i in range(len(policy)):
action = policy[i]
for i in range(policy.shape[0]):
action = policy[i, :]
action_clipped = action.clip(min=-1.0, max=1.0)
output = self.env.step(action_clipped.astype(np.float64))
env_reward += output[1]
done = output[2]
step_count += 1
if output.reward != 0.0:
self.rl_reward += output.reward * 10
done = True
else:
self.rl_step -= 1.0
if done:
break
self.env.reset(seed=self.rl_seed)
return env_reward, step_count
def mainloop_callback(self):
@ -188,10 +193,10 @@ class ActiveRL(Node):
if not self.policy_sent:
self.rl_step = 0
self.rl_reward = 0.0
self.env.reset(seed=self.rl_seed)
self.env.reset()
eval_request = ActiveRLEvalRequest()
eval_request.policy = self.rl_policy.tolist()
eval_request.policy = self.rl_policy.flatten().tolist()
eval_request.weights = self.rl_weights
self.eval_pub.publish(eval_request)
@ -200,7 +205,7 @@ class ActiveRL(Node):
self.policy_sent = True
done = self.next_image(self.rl_policy, self.display_run)
done = self.step(self.rl_policy, self.display_run)
if done:
self.best_pol_shown = True
@ -212,18 +217,18 @@ class ActiveRL(Node):
pass
if self.eval_response_received:
done = self.next_image(self.eval_policy, self.display_run)
done = self.step(self.eval_policy, self.display_run)
if done:
rl_response = ActiveRLResponse()
rl_response.weights = self.eval_weights
rl_response.reward = self.rl_reward
rl_response.final_step = self.rl_step
rl_response.overwrite_weight = self.overwrite_weight
rl_response.weight_preference = self.weight_preference
self.active_rl_pub.publish(rl_response)
self.env.reset(seed=self.rl_seed)
self.env.reset()
# reset flags and attributes
self.reset_eval_request()
@ -240,10 +245,10 @@ class ActiveRL(Node):
if not self.policy_sent:
self.rl_step = 0
self.rl_reward = 0.0
self.env.reset(seed=self.rl_seed)
self.env.reset()
eval_request = ActiveRLEvalRequest()
eval_request.policy = self.rl_policy.tolist()
eval_request.policy = self.rl_policy.flatten().tolist()
eval_request.weights = self.rl_weights
self.eval_pub.publish(eval_request)
@ -252,7 +257,7 @@ class ActiveRL(Node):
self.policy_sent = True
done = self.next_image(self.rl_policy, self.display_run)
done = self.step(self.rl_policy, self.display_run)
if done:
self.rl_step = 0
@ -260,18 +265,18 @@ class ActiveRL(Node):
self.rl_pending = False
elif self.interactive_run == 2:
env_reward, step_count = self.complete_run(self.rl_policy)
env_reward, step_count = self.runner(self.rl_policy)
rl_response = ActiveRLResponse()
rl_response.weights = self.rl_weights
rl_response.reward = env_reward
rl_response.final_step = step_count
if self.overwrite_weight is None:
overwrite_weight = [False] * len(self.rl_weights)
if self.weight_preference is None:
weight_preference = [False] * len(self.rl_weights)
else:
overwrite_weight = self.overwrite_weight
weight_preference = self.weight_preference
rl_response.overwrite_weight = overwrite_weight
rl_response.weight_preference = weight_preference
self.active_rl_pub.publish(rl_response)