def main():
args = parser.parse_args()
env = GridWorld(display=args.render,
obstacles=[np.asarray([1, 2])],
step_wrapper=utils.step_wrapper,
reset_wrapper=utils.reset_wrapper,
stepReward=.01)
model = ActorCritic(env,
gamma=0.99,
log_interval=100,
max_episodes=10**4,
max_ep_length=30)
if args.policy_path is not None:
model.policy.load(args.policy_path)
if not args.play:
model.train()
if not args.dont_save:
model.policy.save('./saved-models/')
if args.play:
env.tickSpeed = 15
assert args.policy_path is not None, 'pass a policy to play from!'
model.generate_trajectory(args.num_trajs, './trajs/ac_gridworld/')
def main():
args = parser.parse_args()
mp.set_start_method('spawn')
env = GridWorld(display=False,
obstacles=[np.asarray([1, 2])],
reset_wrapper=reset_wrapper,
step_wrapper=step_wrapper)
model = ActorCritic(env,
gamma=0.99,
log_interval=100,
max_episodes=10**4,
max_ep_length=30)
if args.policy_path is not None:
model.policy.load(args.policy_path)
if not args.play:
model.train_mp(n_jobs=4)
if not args.dont_save:
model.policy.save('./saved-models/')
if args.play:
assert args.policy_path is not None, 'pass a policy to play from!'
model.generate_trajectory(args.num_trajs, './trajs/mp_gridworld/')
def main():
args = parser.parse_args()
if args.render:
from envs.gridworld import GridWorld
else:
from envs.gridworld_clockless import GridWorldClockless as GridWorld
env = GridWorld(display=args.render,
obstacles=[np.asarray([1, 2])],
goal_state=np.asarray([5, 5]),
step_wrapper=step_wrapper,
reset_wrapper=reset_wrapper,
seed=3)
loss_t = LBT(list_size=100, stop_threshold=1.5, log_interval=100)
model = ActorCritic(env,
gamma=0.99,
log_interval=200,
max_episodes=5000,
max_ep_length=20,
termination=loss_t)
if args.policy_path is not None:
model.policy.load(args.policy_path)
if args.reward_net is not None:
reward_net = RewardNet(env.reset().shape[0])
reward_net.to('cuda')
reward_net.load('./saved-models-rewards/0.pt')
reward_net.eval()
else:
reward_net = None
if not args.play:
model.train_mp(n_jobs=4, reward_net=reward_net, irl=args.irl)
if not args.dont_save:
model.policy.save('./saved-models/')
if args.play:
env.tickSpeed = 15
assert args.policy_path is not None, 'pass a policy to play from!'
model.generate_trajectory(args.num_trajs, './trajs/ac_gridworld/')
def main():
args = parser.parse_args()
experiment_logger = Logger('temp_save.txt')
experiment_logger.log_header('Arguments for the experiment :')
experiment_logger.log_info(vars(args))
mp.set_start_method('spawn')
if args.render:
from envs.gridworld import GridWorld
else:
from envs.gridworld_clockless import GridWorldClockless as GridWorld
agent_width = 10
step_size = 10
obs_width = 10
grid_size = 10
if args.feat_extractor == 'Onehot':
feat_ext = OneHot(grid_rows=10, grid_cols=10)
if args.feat_extractor == 'SocialNav':
feat_ext = SocialNav(fieldList=['agent_state', 'goal_state'])
if args.feat_extractor == 'FrontBackSideSimple':
feat_ext = FrontBackSideSimple(
thresh1=1,
thresh2=2,
thresh3=3,
thresh4=4,
step_size=step_size,
agent_width=agent_width,
obs_width=obs_width,
)
if args.feat_extractor == 'LocalGlobal':
feat_ext = LocalGlobal(
window_size=3,
grid_size=grid_size,
agent_width=agent_width,
obs_width=obs_width,
step_size=step_size,
)
experiment_logger.log_header('Parameters of the feature extractor :')
experiment_logger.log_info(feat_ext.__dict__)
'''
np.asarray([2,2]),np.asarray([7,4]),np.asarray([3,5]),
np.asarray([5,2]),np.asarray([8,3]),np.asarray([7,5]),
np.asarray([3,3]),np.asarray([3,7]),np.asarray([5,7])
'''
env = GridWorld(display=args.render,
is_onehot=False,
is_random=True,
rows=100,
agent_width=agent_width,
step_size=step_size,
obs_width=obs_width,
width=grid_size,
cols=100,
seed=7,
buffer_from_obs=0,
obstacles=3,
goal_state=np.asarray([5, 5]))
experiment_logger.log_header('Environment details :')
experiment_logger.log_info(env.__dict__)
model = ActorCritic(env,
feat_extractor=feat_ext,
gamma=0.99,
log_interval=100,
max_ep_length=40,
hidden_dims=args.policy_net_hidden_dims,
max_episodes=4000)
experiment_logger.log_header('Details of the RL method :')
experiment_logger.log_info(model.__dict__)
pdb.set_trace()
if args.policy_path is not None:
model.policy.load(args.policy_path)
if not args.play and not args.play_user:
if args.reward_path is None:
model.train_mp(n_jobs=4)
else:
from irlmethods.deep_maxent import RewardNet
state_size = featExtract.extract_features(env.reset()).shape[0]
reward_net = RewardNet(state_size)
reward_net.load(args.reward_path)
print(next(reward_net.parameters()).is_cuda)
model.train_mp(reward_net=reward_net, n_jobs=4)
if not args.dont_save:
model.policy.save('./saved-models/')
if args.play:
env.tickSpeed = 15
assert args.policy_path is not None, 'pass a policy to play from!'
model.generate_trajectory(args.num_trajs,
'./trajs/ac_fbs_simple4_static_map7/')
if args.play_user:
env.tickSpeed = 200
model.generate_trajectory_user(args.num_trajs,
'./trajs/ac_gridworld_user/')
def main():
#####for the logger
ts = time.time()
st = datetime.datetime.fromtimestamp(ts).strftime("%Y-%m-%d %H:%M:%S")
###################
args = parser.parse_args()
seed_all(args.seed)
if args.on_server:
matplotlib.use("Agg")
# pygame without monitor
os.environ["SDL_VIDEODRIVER"] = "dummy"
from matplotlib import pyplot as plt
mp.set_start_method("spawn")
from rlmethods.b_actor_critic import ActorCritic
from rlmethods.soft_ac import SoftActorCritic, QSoftActorCritic
from rlmethods.rlutils import ReplayBuffer
from envs.gridworld_drone import GridWorldDrone
from featureExtractor.drone_feature_extractor import (
DroneFeatureSAM1,
DroneFeatureOccup,
DroneFeatureRisk,
DroneFeatureRisk_v2,
VasquezF1,
VasquezF2,
VasquezF3,
Fahad,
GoalConditionedFahad,
)
from featureExtractor.gridworld_featureExtractor import (
FrontBackSide,
LocalGlobal,
OneHot,
SocialNav,
FrontBackSideSimple,
)
from featureExtractor.drone_feature_extractor import (
DroneFeatureRisk_speed,
DroneFeatureRisk_speedv2,
)
from featureExtractor.drone_feature_extractor import VasquezF1
save_folder = None
if not args.dont_save and not args.play:
if not args.save_folder:
print("Provide save folder.")
exit()
policy_net_dims = "-policy_net-"
for dim in args.policy_net_hidden_dims:
policy_net_dims += str(dim)
policy_net_dims += "-"
reward_net_dims = "-reward_net-"
for dim in args.reward_net_hidden_dims:
reward_net_dims += str(dim)
reward_net_dims += "-"
save_folder = (
"./results/"
+ args.save_folder
+ st
+ args.feat_extractor
+ "-seed-"
+ str(args.seed)
+ policy_net_dims
+ reward_net_dims
+ "-total-ep-"
+ str(args.total_episodes)
+ "-max-ep-len-"
+ str(args.max_ep_length)
)
experiment_logger = Logger(save_folder, "experiment_info.txt")
experiment_logger.log_header("Arguments for the experiment :")
repo = git.Repo(search_parent_directories=True)
experiment_logger.log_info({'From branch : ' : repo.active_branch.name})
experiment_logger.log_info({'Commit number : ' : repo.head.object.hexsha})
experiment_logger.log_info(vars(args))
window_size = 9
step_size = 2
agent_width = 10
obs_width = 10
grid_size = 10
feat_ext = None
# initialize the feature extractor to be used
if args.feat_extractor == "Onehot":
feat_ext = OneHot(grid_rows=10, grid_cols=10)
if args.feat_extractor == "SocialNav":
feat_ext = SocialNav(fieldList=["agent_state", "goal_state"])
if args.feat_extractor == "FrontBackSideSimple":
feat_ext = FrontBackSideSimple(
thresh1=1,
thresh2=2,
thresh3=3,
thresh4=4,
step_size=step_size,
agent_width=agent_width,
obs_width=obs_width,
)
if args.feat_extractor == "LocalGlobal":
feat_ext = LocalGlobal(
window_size=11,
grid_size=grid_size,
agent_width=agent_width,
obs_width=obs_width,
step_size=step_size,
)
if args.feat_extractor == "DroneFeatureSAM1":
feat_ext = DroneFeatureSAM1(
agent_width=agent_width,
obs_width=obs_width,
step_size=step_size,
grid_size=grid_size,
thresh1=15,
thresh2=30,
)
if args.feat_extractor == "DroneFeatureOccup":
feat_ext = DroneFeatureOccup(
agent_width=agent_width,
obs_width=obs_width,
step_size=step_size,
grid_size=grid_size,
window_size=window_size,
)
if args.feat_extractor == "DroneFeatureRisk":
feat_ext = DroneFeatureRisk(
agent_width=agent_width,
obs_width=obs_width,
step_size=step_size,
grid_size=grid_size,
show_agent_persp=False,
thresh1=15,
thresh2=30,
)
if args.feat_extractor == "DroneFeatureRisk_v2":
feat_ext = DroneFeatureRisk_v2(
agent_width=agent_width,
obs_width=obs_width,
step_size=step_size,
grid_size=grid_size,
show_agent_persp=False,
thresh1=15,
thresh2=30,
)
if args.feat_extractor == "DroneFeatureRisk_speed":
feat_ext = DroneFeatureRisk_speed(
agent_width=agent_width,
obs_width=obs_width,
step_size=step_size,
grid_size=grid_size,
show_agent_persp=False,
return_tensor=False,
thresh1=10,
thresh2=15,
)
if args.feat_extractor == "DroneFeatureRisk_speedv2":
feat_ext = DroneFeatureRisk_speedv2(
agent_width=agent_width,
obs_width=obs_width,
step_size=step_size,
grid_size=grid_size,
show_agent_persp=False,
return_tensor=False,
thresh1=18,
thresh2=30,
)
if args.feat_extractor == "VasquezF1":
feat_ext = VasquezF1(agent_width * 6, 0.5, 1.0)
if args.feat_extractor == "VasquezF2":
feat_ext = VasquezF1(agent_width * 6, 0.5, 1.0)
if args.feat_extractor == "VasquezF3":
feat_ext = VasquezF3(agent_width)
if args.feat_extractor == "Fahad":
feat_ext = Fahad(36, 60, 0.5, 1.0)
if args.feat_extractor == "GoalConditionedFahad":
feat_ext = GoalConditionedFahad(36, 60, 0.5, 1.0)
if feat_ext is None:
print("Please enter proper feature extractor!")
exit()
# log feature extractor info
if not args.dont_save and not args.play:
experiment_logger.log_header("Parameters of the feature extractor :")
experiment_logger.log_info(feat_ext.__dict__)
# initialize the environment
if args.replace_subject:
replace_subject = True
else:
replace_subject = False
env = GridWorldDrone(
display=args.render,
is_onehot=False,
seed=args.seed,
obstacles=None,
show_trail=False,
is_random=True,
annotation_file=args.annotation_file,
subject=args.subject,
tick_speed=60,
obs_width=10,
step_size=step_size,
agent_width=agent_width,
replace_subject=replace_subject,
segment_size=args.segment_size,
external_control=True,
step_reward=0.001,
show_comparison=True,
consider_heading=True,
show_orientation=True,
# rows=200, cols=200, width=grid_size)
rows=576,
cols=720,
width=grid_size,
)
# env = gym.make('Acrobot-v1')
# log environment info
if not args.dont_save and not args.play:
experiment_logger.log_header("Environment details :")
experiment_logger.log_info(env.__dict__)
# initialize RL
if args.rl_method == "ActorCritic":
model = ActorCritic(
env,
feat_extractor=feat_ext,
gamma=1,
log_interval=100,
max_episode_length=args.max_ep_length,
hidden_dims=args.policy_net_hidden_dims,
save_folder=save_folder,
lr=args.lr,
entropy_coeff=args.entropy_coeff,
max_episodes=args.total_episodes,
)
if args.rl_method == "SAC":
replay_buffer = ReplayBuffer(args.replay_buffer_size)
model = SoftActorCritic(
env,
replay_buffer,
feat_ext,
buffer_sample_size=args.replay_buffer_sample_size,
entropy_tuning=True,
play_interval=args.play_interval,
entropy_target=args.entropy_target,
gamma=args.gamma,
learning_rate=args.lr,
)
if args.rl_method == "discrete_QSAC":
replay_buffer = ReplayBuffer(args.replay_buffer_size)
model = QSoftActorCritic(
env,
replay_buffer,
feat_ext,
buffer_sample_size=args.replay_buffer_sample_size,
entropy_tuning=True,
play_interval=args.play_interval,
entropy_target=args.entropy_target,
gamma=args.gamma,
learning_rate=args.lr,
)
# log RL info
if not args.dont_save and not args.play:
experiment_logger.log_header("Details of the RL method :")
experiment_logger.log_info(model.__dict__)
if args.policy_path is not None:
from debugtools import numericalSort
policy_file_list = []
reward_across_models = []
# print(args.policy_path)
if os.path.isfile(args.policy_path):
policy_file_list.append(args.policy_path)
if os.path.isdir(args.policy_path):
policy_names = glob.glob(os.path.join(args.policy_path, "*.pt"))
policy_file_list = sorted(policy_names, key=numericalSort)
xaxis = np.arange(len(policy_file_list))
if not args.play and not args.play_user:
# no playing of any kind, so training
if args.reward_path is None:
if args.policy_path:
model.policy.load(args.policy_path)
if args.rl_method == "SAC" or args.rl_method == "discrete_QSAC":
model.train(args.total_episodes, args.max_ep_length)
else:
model.train()
else:
from irlmethods.deep_maxent import RewardNet
state_size = feat_ext.extract_features(env.reset()).shape[0]
reward_net = RewardNet(state_size, args.reward_net_hidden_dims)
reward_net.load(args.reward_path)
print(next(reward_net.parameters()).is_cuda)
model.train(reward_net=reward_net)
if not args.dont_save:
model.policy.save(save_folder + "/policy-models/")
if args.play:
# env.tickSpeed = 15
from debugtools import compile_results
xaxis = []
counter = 1
plt.figure(0)
avg_reward_list = []
frac_good_run_list = []
print(policy_file_list)
for policy_file in policy_file_list:
print("Playing for policy :", policy_file)
model.policy.load(policy_file)
policy_folder = policy_file.strip().split("/")[0:-2]
save_folder = ""
for p in policy_folder:
save_folder = save_folder + p + "/"
print("The final save folder ", save_folder)
# env.tickSpeed = 10
assert args.policy_path is not None, "pass a policy to play from!"
if args.exp_trajectory_path is not None:
from irlmethods.irlUtils import calculate_expert_svf
expert_svf = calculate_expert_svf(
args.exp_trajectory_path,
max_time_steps=args.max_ep_length,
feature_extractor=feat_ext,
gamma=1,
)
# reward_across_models.append(model.generate_trajectory(args.num_trajs, args.render))
if args.exp_trajectory_path is None:
if args.dont_save:
rewards, state_info, sub_info = model.generate_trajectory(
args.num_trajs, args.render
)
else:
rewards, state_info, sub_info = model.generate_trajectory(
args.num_trajs,
args.render,
store_raw=args.store_raw_states,
path=save_folder + "/agent_generated_trajectories/",
)
else:
if args.dont_save:
rewards, state_info, sub_info = model.generate_trajectory(
args.num_trajs, args.render, expert_svf=expert_svf
)
else:
rewards, state_info, sub_info = model.generate_trajectory(
args.num_trajs,
args.render,
path=save_folder + "/agent_generated_trajectories/",
expert_svf=expert_svf,
)
avg_reward, good_run_frac = compile_results(
rewards, state_info, sub_info
)
avg_reward_list.append(avg_reward)
frac_good_run_list.append(good_run_frac)
plt.plot(avg_reward_list, c="r")
plt.plot(frac_good_run_list, c="g")
plt.draw()
plt.show()
if args.play_user:
env.tickSpeed = 200
model.generate_trajectory_user(
args.num_trajs, args.render, path="./user_generated_trajectories/"
)
def main():
args = parser.parse_args()
utils.seed_all(args.seed)
if args.on_server:
# matplotlib without monitor
matplotlib.use('Agg')
# pygame without monitor
os.environ['SDL_VIDEODRIVER'] = 'dummy'
from matplotlib import pyplot as plt
save_folder = None
if not args.dont_save:
save_folder = './results/'+ args.save_folder
experiment_logger = Logger(save_folder,'experiment_info.txt')
experiment_logger.log_header('Arguments for the experiment :')
experiment_logger.log_info(vars(args))
mp.set_start_method('spawn')
if args.render:
from envs.gridworld import GridWorld
else:
from envs.gridworld_clockless import GridWorldClockless as GridWorld
if args.feat_extractor=='MCFeatures':
feat_ext = MCFeatures(args.state_discretization[0], args.state_discretization[1])
elif args.feat_extractor=='MCFeaturesOnehot':
feat_ext = MCFeaturesOnehot(args.state_discretization[0], args.state_discretization[1])
else:
print('Enter proper feature extractor value.')
exit()
if not args.dont_save:
experiment_logger.log_header('Parameters of the feature extractor :')
experiment_logger.log_info(feat_ext.__dict__)
'''
np.asarray([2,2]),np.asarray([7,4]),np.asarray([3,5]),
np.asarray([5,2]),np.asarray([8,3]),np.asarray([7,5]),
np.asarray([3,3]),np.asarray([3,7]),np.asarray([5,7])
env = GridWorld(display=args.render, is_onehot= False,is_random=True,
rows=100, agent_width=agent_width,step_size=step_size,
obs_width=obs_width,width=grid_size,
cols=100,
seed=7,
buffer_from_obs=0,
obstacles=3,
goal_state=np.asarray([5,5]))
'''
env = gym.make('MountainCar-v0')
env = env.unwrapped
if not args.dont_save:
experiment_logger.log_header('Environment details :')
experiment_logger.log_info(env.__dict__)
model = ActorCritic(env, feat_extractor=feat_ext, gamma=0.99, plot_loss=False,
log_interval=10, max_ep_length=300, hidden_dims=args.policy_net_hidden_dims,
max_episodes=30, save_folder=save_folder)
if not args.dont_save:
experiment_logger.log_header('Details of the RL method :')
experiment_logger.log_info(model.__dict__)
#pdb.set_trace()
if args.policy_path is not None:
policy_file_list = []
reward_across_models = []
if os.path.isfile(args.policy_path):
policy_file_list.append(args.policy_path)
if os.path.isdir(args.policy_path):
policy_names = glob.glob(os.path.join(args.policy_path, '*.pt'))
policy_file_list = sorted(policy_names, key=numericalSort)
xaxis = np.arange(len(policy_file_list))
if not args.play and not args.play_user:
if args.reward_path is None:
model.train_mp(n_jobs=4)
else:
from irlmethods.deep_maxent import RewardNet
state_size = feat_ext.state_rep_size
reward_net = RewardNet(state_size, args.policy_net_hidden_dims)
reward_net.load(args.reward_path)
print(next(reward_net.parameters()).is_cuda)
model.train_mp(reward_net = reward_net,n_jobs = 4)
if not args.dont_save:
model.policy.save(save_folder+'/policy/')
if args.play:
xaxis = []
counter = 1
print(policy_file_list)
for policy_file in policy_file_list:
model.policy.load(policy_file)
env.tickSpeed = 15
assert args.policy_path is not None, 'pass a policy to play from!'
reward_across_models.append(model.generate_trajectory(args.num_trajs, args.render))
#plotting the 2d list
xaxis.append(counter)
counter += 1
reward_across_models_np = np.array(reward_across_models)
mean_rewards = np.mean(reward_across_models_np, axis=1)
std_rewards = np.std(reward_across_models_np, axis=1)
plt.plot(xaxis,mean_rewards,color = 'r',label='IRL trained agent')
plt.fill_between(xaxis , mean_rewards-std_rewards ,
mean_rewards+std_rewards, alpha = 0.5, facecolor = 'r')
plt.draw()
plt.pause(0.001)
'''
print('RAM usage :')
display_memory_usage(process.memory_info().rss)
print('GPU usage :')
display_memory_usage(torch.cuda.memory_allocated())
torch.cuda.empty_cache()
display_memory_usage(torch.cuda.memory_allocated())
'''
#plt.show()
plt.show()
if args.play_user:
env.tickSpeed = 200
model.generate_trajectory_user(args.num_trajs, './trajs/ac_gridworld_user/')
def main():
args = parser.parse_args()
mp.set_start_method('spawn')
from envs.gridworld_drone import GridWorldDrone
agent_width = 10
step_size = 2
obs_width = 10
grid_size = 10
if args.feat_extractor == 'Onehot':
feat_ext = OneHot(grid_rows=10, grid_cols=10)
if args.feat_extractor == 'SocialNav':
feat_ext = SocialNav(fieldList=['agent_state', 'goal_state'])
if args.feat_extractor == 'FrontBackSideSimple':
feat_ext = FrontBackSideSimple(
thresh1=1,
thresh2=2,
thresh3=3,
thresh4=4,
step_size=step_size,
agent_width=agent_width,
obs_width=obs_width,
fieldList=['agent_state', 'goal_state', 'obstacles'])
if args.feat_extractor == 'LocalGlobal':
feat_ext = LocalGlobal(
window_size=3,
grid_size=grid_size,
agent_width=agent_width,
obs_width=obs_width,
step_size=step_size,
fieldList=['agent_state', 'goal_state', 'obstacles'])
#featExtract = OneHot(grid_rows=10,grid_cols=10)
#featExtract = FrontBackSideSimple(thresh1 = 1,fieldList = ['agent_state','goal_state','obstacles'])
#featExtract = SocialNav(fieldList = ['agent_state','goal_state'])
'''
np.asarray([2,2]),np.asarray([7,4]),np.asarray([3,5]),
np.asarray([5,2]),np.asarray([8,3]),np.asarray([7,5]),
np.asarray([3,3]),np.asarray([3,7]),np.asarray([5,7])
env = GridWorld(display=args.render, is_onehot= False,is_random=True,
rows=10, agent_width=agent_width,step_size=step_size,
obs_width=obs_width,width=grid_size,
cols=10,
seed = 7,
obstacles = '../envs/map3.jpg',
goal_state = np.asarray([5,5]))
'''
env = GridWorldDrone(display=args.render,
is_onehot=False,
seed=999,
obstacles=None,
show_trail=False,
is_random=False,
annotation_file=args.annotation_file,
subject=None,
tick_speed=90,
obs_width=10,
step_size=step_size,
agent_width=agent_width,
show_comparison=True,
rows=576,
cols=720,
width=grid_size)
model = ActorCritic(env,
feat_extractor=featExtract,
gamma=0.99,
log_interval=50,
max_ep_length=500,
max_episodes=2000)
if args.policy_path is not None:
model.policy.load(args.policy_path)
if not args.play and not args.play_user:
if args.reward_path is None:
model.train_mp(n_jobs=4)
else:
from irlmethods.deep_maxent import RewardNet
state_size = featExtract.extract_features(env.reset()).shape[0]
reward_net = RewardNet(state_size)
reward_net.load(args.reward_path)
print(next(reward_net.parameters()).is_cuda)
model.train_mp(reward_net=reward_net, n_jobs=4)
if not args.dont_save:
model.policy.save('./saved-models/')
if args.play:
#env.tickSpeed = 15
assert args.policy_path is not None, 'pass a policy to play from!'
model.generate_trajectory(
args.num_trajs, './trajs/ac_loc_glob_rectified_win_3_static_map3/')
if args.play_user:
env.tickSpeed = 200
model.generate_trajectory_user(args.num_trajs,
'./trajs/ac_gridworld_user/')