def create_tasks(number_of_runs, log_dir, env_id):
ts = list()
alg = 'iwpg'
defaults = utils.get_defaults_kwargs(alg=alg, env_id=env_id)
task_number = 0
gen = hyper_parameter_generator(number_of_runs=number_of_runs)
hms_time = time.strftime("%Y-%m-%d__%H-%M-%S")
try:
while True:
generated_params = next(gen)
task_number += 1
kwargs = defaults.copy()
experiment_path = os.path.join('experiments_a', env_id)
logger_kwargs = setup_logger_kwargs(base_dir=log_dir,
exp_name=experiment_path,
seed=task_number,
hms_time=hms_time,
use_tensor_board=False,
verbose=False)
kwargs.update(logger_kwargs=logger_kwargs,
seed=task_number,
env_id=env_id,
alg=alg,
**generated_params)
target_fn = run_training
t = Task(id=task_number, target_function=target_fn, kwargs=kwargs)
ts.append(t)
except StopIteration:
print(f'Created {task_number} tasks.')
return ts
def create_tasks(number_of_runs, log_dir, env_id):
ts = list()
alg = 'iwpg'
task_number = 0
gen = hyper_parameter_generator(number_of_runs)
try:
while True:
generated_params = next(gen)
task_number += 1
kwargs = utils.get_defaults_kwargs(alg=alg, env_id=env_id)
experiment_path = os.path.join('experiments_c', env_id)
logger_kwargs = setup_logger_kwargs(base_dir=log_dir,
exp_name=experiment_path,
seed=task_number,
use_tensor_board=False,
verbose=False)
kwargs.update(logger_kwargs=logger_kwargs,
seed=task_number,
env_id=env_id,
alg=alg,
**generated_params)
# deactivate reward scaling for manipulation tasks
env_type, _ = utils.get_env_type(env_id=env_id)
if env_type == 'gym_manipulator_envs' or env_type == 'bullet':
kwargs['use_reward_scaling'] = False
target_fn = run_iwpg_training
t = Task(id=task_number, target_function=target_fn, kwargs=kwargs)
ts.append(t)
except StopIteration:
print(f'Created {task_number} tasks.')
return ts
def __init__(self,
alg: str,
env_id: str,
log_dir: str,
seed: int,
unparsed_args: list = ()
) -> None:
""" Class Constructor """
self.alg = alg
self.env_id = env_id
self.log_dir = log_dir
self.seed = seed
self.multi_thread = False
self.num_runs = 1
self.training = False
self.compiled = False
self.trained = False
self.default_kwargs = utils.get_defaults_kwargs(alg=alg,
env_id=env_id)
self.kwargs = self.default_kwargs.copy()
# update algorithm kwargs with unparsed arguments from command line
keys = [k[2:] for k in unparsed_args[0::2]] # remove -- from argument
values = [eval(v) for v in unparsed_args[1::2]]
unparsed_dict = {k: v for k, v in zip(keys, values)}
self.kwargs.update(**unparsed_dict)
self.logger_kwargs = None # defined by compile (a specific seed might be passed)
self.exp_name = os.path.join(self.env_id, self.alg)
# assigned by class methods
self.model = None
self.env = None
self.scheduler = None
def learn(env_id, **kwargs) -> tuple:
defaults = utils.get_defaults_kwargs(alg='npg', env_id=env_id)
defaults.update(**kwargs)
alg = NaturalPolicyGradientAlgorithm(env_id=env_id, **kwargs)
ac, env = alg.learn()
return ac, env
def create_tasks(number_of_runs, log_dir, env_id):
ts = list()
task_number = 1000 # === seed number
alg = 'iwpg'
if env_id == 'all':
env_ids = [
'HalfCheetahBulletEnv-v0',
'AntBulletEnv-v0',
'HopperBulletEnv-v0',
'Walker2DBulletEnv-v0',
'HumanoidBulletEnv-v0',
'ReacherBulletEnv-v0',
'PusherBulletEnv-v0',
'KukaBulletEnv-v0',
]
else:
env_ids = [
env_id,
]
for env_id in env_ids:
defaults = utils.get_defaults_kwargs(alg=alg, env_id=env_id)
gen = hyper_parameter_generator(number_of_runs)
try:
while True:
generated_params = next(gen)
task_number += 1
kwargs = defaults.copy()
experiment_path = os.path.join('experiments_d', env_id)
logger_kwargs = setup_logger_kwargs(base_dir=log_dir,
exp_name=experiment_path,
seed=task_number,
use_tensor_board=False,
verbose=False)
kwargs.update(logger_kwargs=logger_kwargs,
seed=task_number,
env_id=env_id,
alg=alg,
**generated_params)
# deactivate reward scaling for manipulation tasks
env_type, _ = utils.get_env_type(env_id=env_id)
if env_type == 'gym_manipulator_envs' or env_type == 'bullet':
kwargs['use_reward_scaling'] = False
target_fn = run_training
t = Task(id=task_number,
target_function=target_fn,
kwargs=kwargs)
ts.append(t)
except StopIteration:
pass
print(f'Created {task_number} tasks.')
return ts
def check_alg(alg_name, env_id):
"""" Run one epoch update with algorithm."""
print(f'Run {alg_name}.')
defaults = U.get_defaults_kwargs(alg=alg_name, env_id=env_id)
defaults['epochs'] = 1
defaults['num_mini_batches'] = 4
defaults['steps_per_epoch'] = 1000
defaults['verbose'] = False
learn_fn = U.get_learn_function(alg_name)
defaults['logger_kwargs'] = setup_logger_kwargs(exp_name='unittest',
seed=None,
base_dir='/var/tmp/',
datestamp=True,
use_tensor_board=True,
verbose=False)
return learn_fn(env_id, **defaults)