def run_sac_experiment(main,
mode,
include_folders=None,
log_dir=None,
exp_prefix="experiment",
exp_name=None,
**kwargs):
if exp_name is None:
exp_name = timestamp()
if log_dir is None:
log_dir = os.path.join(DEFAULT_LOG_DIR, "local",
exp_prefix.replace("_", "-"), exp_name)
if include_folders is None:
include_folders = list()
if mode == 'ec2':
include_folders.append('sac')
all_symlinks = list()
for folder in include_folders:
all_symlinks.append(_create_symlink(folder))
kwargs.update(added_project_directories=all_symlinks)
run_experiment_lite(stub_method_call=main,
mode=mode,
exp_prefix=exp_prefix,
exp_name=exp_name,
log_dir=log_dir,
**kwargs)
def parse_args():
parser = argparse.ArgumentParser()
parser.add_argument('--cost_type', type=float, default=10)
parser.add_argument('--domain',
type=str,
choices=AVAILABLE_DOMAINS,
default=None)
parser.add_argument('--task',
type=str,
choices=AVAILABLE_TASKS,
default='default')
parser.add_argument('--policy',
type=str,
choices=('gaussian', 'gmm', 'lsp'),
default='gaussian')
parser.add_argument('--envn', type=int, default=10)
parser.add_argument('--scale', type=float, default=1)
parser.add_argument('--exp_name', type=str, default=timestamp())
parser.add_argument('--mode', type=str, default='local')
args = parser.parse_args()
args.env = ENVS[args.envn]
log_dir = '/home/wisrl/Downloads/log_sac/Mujoco/' + args.env
log_dir += '_c%s'%args.cost_type
if args.task == 'delayed':
log_dir += '_delayed%s'%str(DELAY_CONST)
if 'cross' in args.task:
log_dir += '_cross'
log_dir = log_dir + '/SAC'
if not args.scale==1.0:
log_dir += '_s%s'%str(args.scale)
args.log_dir = log_dir
return args
def parse_args():
parser = argparse.ArgumentParser()
parser.add_argument('--domain',
type=str,
choices=AVAILABLE_DOMAINS,
default=None)
parser.add_argument('--task',
type=str,
choices=AVAILABLE_TASKS,
default='default')
parser.add_argument('--policy',
type=str,
choices=('gaussian', 'gmm', 'lsp'),
default='gaussian')
parser.add_argument('--env', type=str, default=DEFAULT_ENV)
parser.add_argument('--exp_name', type=str, default=timestamp())
parser.add_argument('--mode', type=str, default='local')
parser.add_argument('--log_dir', type=str, default=None)
## RC: Added to run time-complexity experiments.
parser.add_argument('--n_train_repeat', type=int, default=1)
parser.add_argument('--n_parallel', type=int, default=1)
parser.add_argument('--n_epochs', type=int, default=1000)
parser.add_argument('--gpu_fraction', type=float, default=1.0)
args = parser.parse_args()
return args
def parse_args():
parser = argparse.ArgumentParser()
parser.add_argument('--env',
type=str,
choices=AVAILABLE_ENVS,
default='swimmer')
parser.add_argument('--exp_name',type=str, default=timestamp())
parser.add_argument('--mode', type=str, default='local')
parser.add_argument('--log_dir', type=str, default=None)
args = parser.parse_args()
return args
def parse_args():
parser = argparse.ArgumentParser()
parser.add_argument('--domain',
type=str,
choices=AVAILABLE_DOMAINS,
default=None)
parser.add_argument('--task',
type=str,
choices=AVAILABLE_TASKS,
default='default')
parser.add_argument('--policy',
type=str,
choices=('gaussian', 'gmm', 'lsp'),
default='gaussian')
parser.add_argument('--env', type=str, default=DEFAULT_ENV)
parser.add_argument('--exp_name', type=str, default=timestamp())
parser.add_argument('--mode', type=str, default='local')
parser.add_argument('--tau', type=float, default=0.005)
parser.add_argument('--log_dir', type=str, default=None)
parser.add_argument('--lr', type=float, default=-1.0)
parser.add_argument('--l1regpi', type=float, default=0.0) #L1 reg policy
parser.add_argument('--l2regpi', type=float, default=0.0) #L2 reg policy
parser.add_argument(
'--l1regvf', type=float, default=0.0
) #L1 reg value (V only, the two Q networks are not regularized for simplicity)
parser.add_argument('--l2regvf', type=float, default=0.0) #L2 reg value
parser.add_argument('--wclippi', type=float,
default=0.0) #Weight clip policy
parser.add_argument('--wclipvf', type=float,
default=0.0) #Weight clip value
parser.add_argument('--dropoutpi', type=float,
default=1.0) #Dropout policy keep prob
parser.add_argument('--dropoutvf', type=float,
default=1.0) #Dropout value keep prob
parser.add_argument(
'--ent_coef', type=float, default=0.0
) #Entropy regularization coefficient, not the temperature of maximum entropy formulation (In this SAC implementation, this temperature term is fixed to be 1.
parser.add_argument('--batchnormpi', type=bool,
default=False) #Batchnorm policy (T/F)
parser.add_argument('--batchnormvf', type=bool,
default=False) #Batchnorm value (T/F)
#parser.add_argument('--gaussianreg', type=float, default=1e-3) #This term appears in the original code release and regularizes the mu and logsigma of policy output. However, this regularization term is not applied when we use gaussian policy. Since all of our experiements adopt gaussian policy, this term does not affect our result in anyway.
parser.add_argument('--reward_scale', type=float, default=-1.0)
parser.add_argument('--num_hidden', type=int, default=256)
parser.add_argument('--seed', type=int, default=1)
parser.add_argument('--policypath', type=str,
default='') #Policy network save path
parser.add_argument('--valuepath', type=str,
default='') #Value network save path
args = parser.parse_args()
return args
def main():
args = parse_args()
variant = {'policy_type': args.policy_type}
run_experiment_lite(
run,
exp_prefix='multigoal',
exp_name=timestamp(),
variant=variant,
snapshot_mode='last',
n_parallel=1,
seed=1,
mode='local',
)
def parse_args():
parser = argparse.ArgumentParser()
parser.add_argument('--env',
type=str,
choices=AVAILABLE_ENVS,
default='swimmer')
parser.add_argument('--exp_name', type=str, default=timestamp())
parser.add_argument('--mode', type=str, default='local')
parser.add_argument('--log_dir', type=str, default='./logs/unity')
parser.add_argument('--idx', type=int, default=0)
parser.add_argument('--no_graphics', type=bool, default=False)
args = parser.parse_args()
return args
def parse_args():
parser = argparse.ArgumentParser()
parser.add_argument('--env',
type=str,
choices=AVAILABLE_ENVS,
default=DEFAULT_ENV)
parser.add_argument('--exp_name', type=str, default=timestamp())
parser.add_argument('--mode', type=str, default='local')
parser.add_argument('--log_dir', type=str, default=None)
parser.add_argument('--low_level_policy_path',
'-p',
type=str,
default=None)
args = parser.parse_args()
return args
def launch_experiments(args):
num_experiments = 1
print('Launching {} experiments.'.format(num_experiments))
for i in range(num_experiments):
print("Experiment: {}/{}".format(i + 1, num_experiments))
run_sac_experiment(
run_experiment,
mode='local',
n_parallel=1,
terminate_machine=True,
log_dir='/root/code/log/{0}/{1}'.format(args.domain, timestamp()),
snapshot_mode='gap',
snapshot_gap=100,
sync_s3_pkl=True,
)
def parse_args():
parser = argparse.ArgumentParser()
parser.add_argument('--domain',
type=str,
choices=AVAILABLE_DOMAINS,
default='ant-cross-maze')
parser.add_argument('--policy',
type=str,
choices=('gaussian', 'gaussian_ptr'),
default='gaussian_ptr')
parser.add_argument('--env', type=str, default=DEFAULT_ENV)
parser.add_argument('--exp_name', type=str, default=timestamp())
parser.add_argument('--mode', type=str, default='local')
parser.add_argument('--log_dir', type=str, default=None)
args = parser.parse_args()
return args
def parse_args():
parser = argparse.ArgumentParser()
parser.add_argument('--env',
type=str,
choices=AVAILABLE_ENVS,
default='swimmer')
parser.add_argument('--exp_name', type=str, default=timestamp())
parser.add_argument('--mode', type=str, default='local')
parser.add_argument('--log_dir', type=str, default=None)
parser.add_argument('--seed', type=int, default=1)
parser.add_argument('--num_skills', type=int, default=None)
parser.add_argument('--eval_freq', type=int, default=None)
parser.add_argument('--xdir', default=None)
parser.add_argument('--xname', default=None)
args = parser.parse_args()
return args
def parse_args():
parser = argparse.ArgumentParser()
parser.add_argument('--domain',
type=str,
default='Baxter')
parser.add_argument('--task',
type=str,
default='default')
parser.add_argument('--policy',
type=str,
choices=('gaussian', 'gmm', 'lsp'),
default='gaussian')
parser.add_argument('--env', type=str, default='BaxterReachNG')
parser.add_argument('--exp_name', type=str, default=timestamp())
parser.add_argument('--mode', type=str, default='local')
parser.add_argument('--log_dir', type=str, default=None)
args = parser.parse_args()
return args
def parse_args():
parser = argparse.ArgumentParser()
parser.add_argument('--domain',
type=str,
choices=AVAILABLE_DOMAINS,
default=None)
parser.add_argument('--task',
type=str,
choices=AVAILABLE_TASKS,
default='default')
parser.add_argument('--policy',
type=str,
choices=('lsp', 'gmm'),
default='lsp')
parser.add_argument('--env', type=str, default=DEFAULT_ENV)
parser.add_argument('--exp_name', type=str, default=timestamp())
parser.add_argument('--mode', type=str, default='local')
parser.add_argument('--log_dir', type=str, default=None)
args = parser.parse_args()
return args
def launch_experiments():
args = arg()
num_experiments = 1
print('Launching {} experiments.'.format(num_experiments))
for i in range(num_experiments):
print("Experiment: {}/{}".format(i + 1, num_experiments))
experiment_prefix = args.domain + '/' + args.exp_name
experiment_name = '{prefix}-{exp_name}-{i:02}'.format(
prefix=args.domain, exp_name=args.exp_name, i=0)
run_sac_experiment(
run_experiment,
mode='local',
exp_prefix=experiment_prefix,
exp_name=experiment_name,
n_parallel=1,
terminate_machine=True,
log_dir='/root/code/log/{0}/{1}'.format(args.domain, timestamp()),
snapshot_mode='gap',
snapshot_gap=100,
sync_s3_pkl=True,
)
env=env,
policy=policy,
initial_exploration_policy=initial_exploration_policy,
pool=pool,
qf1=qf1,
qf2=qf2,
vf=vf,
lr=3e-4,
scale_reward=20,
discount=0.99,
tau=0.005,
reparameterize=True,
target_update_interval=1,
action_prior='uniform',
save_full_state=False,
)
algorithm._sess.run(tf.global_variables_initializer())
algorithm.train()
if __name__ == "__main__":
run_sac_experiment(
run_experiment,
mode='local',
log_dir='/root/code/log/prim/reach/{0}'.format(timestamp()),
snapshot_mode='gap',
snapshot_gap=100,
)
env=env,
policy=policy,
initial_exploration_policy=initial_exploration_policy,
pool=pool,
qf1=qf1,
qf2=qf2,
vf=vf,
lr=3e-4,
scale_reward=20,
discount=0.99,
tau=0.005,
reparameterize=True,
target_update_interval=1,
action_prior='uniform',
save_full_state=False,
)
algorithm._sess.run(tf.global_variables_initializer())
algorithm.train()
if __name__ == "__main__":
run_sac_experiment(
run_experiment,
mode='local',
log_dir='/root/code/log/prim/pick/{0}'.format(timestamp()),
snapshot_mode='gap',
snapshot_gap=100,
)
# plotter=plotter,
lr=1e-3,
discount=0.99,
#tau=1e-4,
target_update_interval=20,
# reparameterize=False,
save_full_state=False
)
algo.train()
if __name__ == "__main__":
exp_prefix = 'DQN-VoltVar34-exp2'
exp_name = timestamp()
log_dir = os.path.join(
DEFAULT_LOG_DIR,
"local",
exp_prefix.replace("_", "-"),
exp_name)
os.makedirs(log_dir,exist_ok=True)
logger.set_snapshot_dir(log_dir)
logger.set_snapshot_mode('last')
tabular_log_file = os.path.join(log_dir,'progress.csv')
text_log_file = os.path.join(log_dir,'debug.log')
logger.add_text_output(text_log_file)
logger.add_tabular_output(tabular_log_file)
run_experiment()
# run_sac_experiment(
