def main(args):
dataset, env = get_pybullet(args.dataset)
d3rlpy.seed(args.seed)
train_episodes, test_episodes = train_test_split(dataset, test_size=0.2)
device = None if args.gpu is None else Device(args.gpu)
encoder_factory = VectorEncoderFactory(hidden_units=[256, 256, 256, 256])
awac = AWAC(actor_encoder_factory=encoder_factory,
critic_encoder_factory=encoder_factory,
q_func_factory=args.q_func,
use_gpu=device)
awac.fit(train_episodes,
eval_episodes=test_episodes,
n_epochs=1000,
scorers={
'environment': evaluate_on_environment(env),
'td_error': td_error_scorer,
'discounted_advantage': discounted_sum_of_advantage_scorer,
'value_scale': average_value_estimation_scorer,
'value_std': value_estimation_std_scorer,
'action_diff': continuous_action_diff_scorer
})
def test_device(mock):
device = Device()
copy_device = copy.deepcopy(device)
assert device.get_id() == 0
assert copy_device.get_id() == 0
with parallel():
inc_device = copy.deepcopy(device)
assert device.get_id() == 1
assert inc_device.get_id() == 1
# check circulation
inc2_device = copy.deepcopy(device)
assert device.get_id() == 0
assert inc2_device.get_id() == 0
def main(args):
dataset, env = get_pybullet(args.dataset)
d3rlpy.seed(args.seed)
train_episodes, test_episodes = train_test_split(dataset, test_size=0.2)
device = None if args.gpu is None else Device(args.gpu)
dynamics = ProbabilisticEnsembleDynamics(use_gpu=device)
dynamics.fit(train_episodes,
eval_episodes=test_episodes,
n_steps=100000,
scorers={
"obs_error": dynamics_observation_prediction_error_scorer,
"reward_error": dynamics_reward_prediction_error_scorer,
})
combo = COMBO(q_func_factory=args.q_func,
dynamics=dynamics,
use_gpu=device)
combo.fit(train_episodes,
eval_episodes=test_episodes,
n_steps=1000000,
scorers={
'environment': evaluate_on_environment(env),
'td_error': td_error_scorer,
'discounted_advantage': discounted_sum_of_advantage_scorer,
'value_scale': average_value_estimation_scorer,
'value_std': value_estimation_std_scorer,
'action_diff': continuous_action_diff_scorer
})
def main(args):
dataset, env = get_atari(args.dataset)
d3rlpy.seed(args.seed)
train_episodes, test_episodes = train_test_split(dataset, test_size=0.2)
device = None if args.gpu is None else Device(args.gpu)
bc = DiscreteBC(n_epochs=100,
scaler='pixel',
use_batch_norm=False,
use_gpu=device)
bc.fit(train_episodes,
eval_episodes=test_episodes,
scorers={'environment': evaluate_on_environment(env, epsilon=0.05)})
def main(args):
dataset, env = get_pybullet(args.dataset)
d3rlpy.seed(args.seed)
train_episodes, test_episodes = train_test_split(dataset, test_size=0.2)
device = None if args.gpu is None else Device(args.gpu)
bc = BC(n_epochs=100, use_gpu=device)
bc.fit(train_episodes,
eval_episodes=test_episodes,
scorers={
'environment': evaluate_on_environment(env),
'action_diff': continuous_action_diff_scorer
})
def main(args):
dataset, env = get_pybullet(args.dataset)
d3rlpy.seed(args.seed)
train_episodes, test_episodes = train_test_split(dataset, test_size=0.2)
device = None if args.gpu is None else Device(args.gpu)
awr = AWR(n_epochs=100, use_gpu=device)
awr.fit(train_episodes,
eval_episodes=test_episodes,
scorers={
'environment': evaluate_on_environment(env),
'td_error': td_error_scorer,
'value_scale': average_value_estimation_scorer,
'action_diff': continuous_action_diff_scorer
})
def main(args):
dataset, env = get_pybullet(args.dataset)
d3rlpy.seed(args.seed)
train_episodes, test_episodes = train_test_split(dataset, test_size=0.2)
device = None if args.gpu is None else Device(args.gpu)
sac = SAC(n_epochs=100, q_func_type=args.q_func_type, use_gpu=device)
sac.fit(train_episodes,
eval_episodes=test_episodes,
scorers={
'environment': evaluate_on_environment(env),
'td_error': td_error_scorer,
'discounted_advantage': discounted_sum_of_advantage_scorer,
'value_scale': average_value_estimation_scorer,
'value_std': value_estimation_std_scorer,
'action_diff': continuous_action_diff_scorer
})
def main(args):
dataset, env = get_atari(args.dataset)
d3rlpy.seed(args.seed)
train_episodes, test_episodes = train_test_split(dataset, test_size=0.2)
device = None if args.gpu is None else Device(args.gpu)
dqn = DQN(n_epochs=100,
q_func_type=args.q_func_type,
scaler='pixel',
use_batch_norm=False,
use_gpu=device)
dqn.fit(train_episodes,
eval_episodes=test_episodes,
scorers={
'environment': evaluate_on_environment(env, epsilon=0.05),
'td_error': td_error_scorer,
'discounted_advantage': discounted_sum_of_advantage_scorer,
'value_scale': average_value_estimation_scorer
})
@pytest.mark.parametrize('value', ['min_max', MinMaxScaler(), None])
def test_check_scaler(value):
scaler = check_scaler(value)
if value is None:
assert scaler is None
else:
assert isinstance(scaler, MinMaxScaler)
@pytest.mark.parametrize('value', [['random_shift'], [RandomShift()], None])
def test_check_augmentation(value):
pipeline = check_augmentation(value)
assert isinstance(pipeline, AugmentationPipeline)
if value is None:
assert len(pipeline.augmentations) == 0
else:
assert isinstance(pipeline.augmentations[0], RandomShift)
@pytest.mark.parametrize('value', [False, True, 0, Device(0)])
def test_check_use_gpu(value):
device = check_use_gpu(value)
if type(value) == bool and value:
assert device.get_id() == 0
elif type(value) == bool and not value:
assert device is None
elif type(value) == int:
assert device.get_id() == 0
elif isinstance(value, Device):
assert device.get_id() == 0
def to_gpu(self, device=Device()):
self.device = 'cuda:%d' % device.get_id()
to_cuda(self, self.device)
