def test_discrete_cql_impl(
observation_shape,
action_size,
learning_rate,
optim_factory,
encoder_factory,
q_func_factory,
gamma,
n_critics,
target_reduction_type,
alpha,
scaler,
reward_scaler,
):
impl = DiscreteCQLImpl(
observation_shape=observation_shape,
action_size=action_size,
learning_rate=learning_rate,
optim_factory=optim_factory,
encoder_factory=encoder_factory,
q_func_factory=create_q_func_factory(q_func_factory),
gamma=gamma,
n_critics=n_critics,
target_reduction_type=target_reduction_type,
alpha=alpha,
use_gpu=None,
scaler=scaler,
reward_scaler=reward_scaler,
)
torch_impl_tester(
impl, discrete=True, deterministic_best_action=q_func_factory != "iqn"
)
def test_discrete_fqe_impl(
observation_shape,
action_size,
learning_rate,
optim_factory,
encoder_factory,
q_func_factory,
gamma,
n_critics,
scaler,
):
fqe = DiscreteFQEImpl(
observation_shape=observation_shape,
action_size=action_size,
learning_rate=learning_rate,
optim_factory=optim_factory,
encoder_factory=encoder_factory,
q_func_factory=create_q_func_factory(q_func_factory),
gamma=gamma,
n_critics=n_critics,
use_gpu=None,
scaler=scaler,
action_scaler=None,
)
torch_impl_tester(fqe, True)
def test_discrete_fqe_impl(
observation_shape,
action_size,
learning_rate,
optim_factory,
encoder_factory,
q_func_factory,
gamma,
n_critics,
bootstrap,
share_encoder,
scaler,
):
fqe = DiscreteFQEImpl(
observation_shape,
action_size,
learning_rate,
optim_factory,
encoder_factory,
create_q_func_factory(q_func_factory),
gamma,
n_critics,
bootstrap,
share_encoder,
use_gpu=False,
scaler=scaler,
action_scaler=None,
)
torch_impl_tester(fqe, True)
def test_double_dqn_impl(
observation_shape,
action_size,
learning_rate,
optim_factory,
encoder_factory,
q_func_factory,
gamma,
n_critics,
target_reduction_type,
scaler,
augmentation,
):
impl = DoubleDQNImpl(
observation_shape=observation_shape,
action_size=action_size,
learning_rate=learning_rate,
optim_factory=optim_factory,
encoder_factory=encoder_factory,
q_func_factory=create_q_func_factory(q_func_factory),
gamma=gamma,
n_critics=n_critics,
target_reduction_type=target_reduction_type,
use_gpu=None,
scaler=scaler,
augmentation=augmentation,
)
torch_impl_tester(impl,
discrete=True,
deterministic_best_action=q_func_factory != "iqn")
def test_discrete_bcq_impl(
observation_shape,
action_size,
learning_rate,
optim_factory,
encoder_factory,
q_func_factory,
gamma,
n_critics,
target_reduction_type,
action_flexibility,
beta,
scaler,
):
impl = DiscreteBCQImpl(
observation_shape=observation_shape,
action_size=action_size,
learning_rate=learning_rate,
optim_factory=optim_factory,
encoder_factory=encoder_factory,
q_func_factory=create_q_func_factory(q_func_factory),
gamma=gamma,
n_critics=n_critics,
target_reduction_type=target_reduction_type,
action_flexibility=action_flexibility,
beta=beta,
use_gpu=None,
scaler=scaler,
)
torch_impl_tester(impl,
discrete=True,
deterministic_best_action=q_func_factory != "iqn")
def test_discrete_cql_impl(
observation_shape,
action_size,
learning_rate,
optim_factory,
encoder_factory,
q_func_factory,
gamma,
n_critics,
bootstrap,
share_encoder,
scaler,
augmentation,
):
impl = DiscreteCQLImpl(
observation_shape,
action_size,
learning_rate,
optim_factory,
encoder_factory,
create_q_func_factory(q_func_factory),
gamma,
n_critics,
bootstrap,
share_encoder,
use_gpu=False,
scaler=scaler,
augmentation=augmentation,
)
torch_impl_tester(impl,
discrete=True,
deterministic_best_action=q_func_factory != "iqn")
def test_cql_impl(
observation_shape,
action_size,
actor_learning_rate,
critic_learning_rate,
temp_learning_rate,
alpha_learning_rate,
actor_optim_factory,
critic_optim_factory,
temp_optim_factory,
alpha_optim_factory,
encoder_factory,
q_func_factory,
gamma,
tau,
n_critics,
target_reduction_type,
initial_temperature,
initial_alpha,
alpha_threshold,
conservative_weight,
n_action_samples,
soft_q_backup,
scaler,
action_scaler,
reward_scaler,
):
impl = CQLImpl(
observation_shape=observation_shape,
action_size=action_size,
actor_learning_rate=actor_learning_rate,
critic_learning_rate=critic_learning_rate,
temp_learning_rate=temp_learning_rate,
alpha_learning_rate=alpha_learning_rate,
actor_optim_factory=actor_optim_factory,
critic_optim_factory=critic_optim_factory,
temp_optim_factory=temp_optim_factory,
alpha_optim_factory=alpha_optim_factory,
actor_encoder_factory=encoder_factory,
critic_encoder_factory=encoder_factory,
q_func_factory=create_q_func_factory(q_func_factory),
gamma=gamma,
tau=tau,
n_critics=n_critics,
target_reduction_type=target_reduction_type,
initial_temperature=initial_temperature,
initial_alpha=initial_alpha,
alpha_threshold=alpha_threshold,
conservative_weight=conservative_weight,
n_action_samples=n_action_samples,
soft_q_backup=soft_q_backup,
use_gpu=None,
scaler=scaler,
action_scaler=action_scaler,
reward_scaler=reward_scaler,
)
torch_impl_tester(
impl, discrete=False, deterministic_best_action=q_func_factory != "iqn"
)
def test_cql_impl(
observation_shape,
action_size,
actor_learning_rate,
critic_learning_rate,
temp_learning_rate,
alpha_learning_rate,
actor_optim_factory,
critic_optim_factory,
temp_optim_factory,
alpha_optim_factory,
encoder_factory,
q_func_factory,
gamma,
tau,
n_critics,
bootstrap,
share_encoder,
initial_temperature,
initial_alpha,
alpha_threshold,
n_action_samples,
soft_q_backup,
scaler,
action_scaler,
augmentation,
):
impl = CQLImpl(
observation_shape,
action_size,
actor_learning_rate,
critic_learning_rate,
temp_learning_rate,
alpha_learning_rate,
actor_optim_factory,
critic_optim_factory,
temp_optim_factory,
alpha_optim_factory,
encoder_factory,
encoder_factory,
create_q_func_factory(q_func_factory),
gamma,
tau,
n_critics,
bootstrap,
share_encoder,
initial_temperature,
initial_alpha,
alpha_threshold,
n_action_samples,
soft_q_backup,
use_gpu=False,
scaler=scaler,
action_scaler=action_scaler,
augmentation=augmentation,
)
torch_impl_tester(impl,
discrete=False,
deterministic_best_action=q_func_factory != "iqn")
def test_create_q_func_factory(name):
factory = create_q_func_factory(name)
if name == "mean":
assert isinstance(factory, MeanQFunctionFactory)
elif name == "qr":
assert isinstance(factory, QRQFunctionFactory)
elif name == "iqn":
assert isinstance(factory, IQNQFunctionFactory)
elif name == "fqf":
assert isinstance(factory, FQFQFunctionFactory)
def test_plas_with_perturbation_impl(
observation_shape,
action_size,
actor_learning_rate,
critic_learning_rate,
imitator_learning_rate,
actor_optim_factory,
critic_optim_factory,
imitator_optim_factory,
encoder_factory,
q_func_factory,
gamma,
tau,
n_critics,
bootstrap,
share_encoder,
target_reduction_type,
lam,
beta,
action_flexibility,
scaler,
action_scaler,
augmentation,
):
impl = PLASWithPerturbationImpl(
observation_shape,
action_size,
actor_learning_rate,
critic_learning_rate,
imitator_learning_rate,
actor_optim_factory,
critic_optim_factory,
imitator_optim_factory,
encoder_factory,
encoder_factory,
encoder_factory,
create_q_func_factory(q_func_factory),
gamma,
tau,
n_critics,
bootstrap,
share_encoder,
target_reduction_type,
lam,
beta,
action_flexibility,
use_gpu=False,
scaler=scaler,
action_scaler=action_scaler,
augmentation=augmentation,
)
torch_impl_tester(
impl, discrete=False, deterministic_best_action=q_func_factory != "iqn"
)
def test_plas_with_perturbation_impl(
observation_shape,
action_size,
actor_learning_rate,
critic_learning_rate,
imitator_learning_rate,
actor_optim_factory,
critic_optim_factory,
imitator_optim_factory,
encoder_factory,
q_func_factory,
gamma,
tau,
n_critics,
target_reduction_type,
lam,
beta,
action_flexibility,
scaler,
action_scaler,
reward_scaler,
):
impl = PLASWithPerturbationImpl(
observation_shape=observation_shape,
action_size=action_size,
actor_learning_rate=actor_learning_rate,
critic_learning_rate=critic_learning_rate,
imitator_learning_rate=imitator_learning_rate,
actor_optim_factory=actor_optim_factory,
critic_optim_factory=critic_optim_factory,
imitator_optim_factory=imitator_optim_factory,
actor_encoder_factory=encoder_factory,
critic_encoder_factory=encoder_factory,
imitator_encoder_factory=encoder_factory,
q_func_factory=create_q_func_factory(q_func_factory),
gamma=gamma,
tau=tau,
n_critics=n_critics,
target_reduction_type=target_reduction_type,
lam=lam,
beta=beta,
action_flexibility=action_flexibility,
use_gpu=None,
scaler=scaler,
action_scaler=action_scaler,
reward_scaler=reward_scaler,
)
torch_impl_tester(impl,
discrete=False,
deterministic_best_action=q_func_factory != "iqn")
def test_awac_impl(
observation_shape,
action_size,
actor_learning_rate,
critic_learning_rate,
actor_optim_factory,
critic_optim_factory,
encoder_factory,
q_func_factory,
gamma,
tau,
lam,
n_action_samples,
max_weight,
n_critics,
bootstrap,
share_encoder,
target_reduction_type,
scaler,
action_scaler,
augmentation,
):
impl = AWACImpl(
observation_shape,
action_size,
actor_learning_rate,
critic_learning_rate,
actor_optim_factory,
critic_optim_factory,
encoder_factory,
encoder_factory,
create_q_func_factory(q_func_factory),
gamma,
tau,
lam,
n_action_samples,
max_weight,
n_critics,
bootstrap,
share_encoder,
target_reduction_type,
use_gpu=False,
scaler=scaler,
action_scaler=action_scaler,
augmentation=augmentation,
)
torch_impl_tester(impl,
discrete=False,
deterministic_best_action=q_func_factory != "iqn")
def test_td3_impl(
observation_shape,
action_size,
actor_learning_rate,
critic_learning_rate,
actor_optim_factory,
critic_optim_factory,
encoder_factory,
q_func_factory,
gamma,
tau,
n_critics,
bootstrap,
share_encoder,
target_reduction_type,
target_smoothing_sigma,
target_smoothing_clip,
scaler,
action_scaler,
augmentation,
):
impl = TD3Impl(
observation_shape,
action_size,
actor_learning_rate,
critic_learning_rate,
actor_optim_factory,
critic_optim_factory,
encoder_factory,
encoder_factory,
create_q_func_factory(q_func_factory),
gamma,
tau,
n_critics,
bootstrap,
share_encoder,
target_reduction_type,
target_smoothing_sigma,
target_smoothing_clip,
use_gpu=False,
scaler=scaler,
action_scaler=action_scaler,
augmentation=augmentation,
)
torch_impl_tester(impl,
discrete=False,
deterministic_best_action=q_func_factory != "iqn")
def test_sac_impl(
observation_shape,
action_size,
actor_learning_rate,
critic_learning_rate,
temp_learning_rate,
actor_optim_factory,
critic_optim_factory,
temp_optim_factory,
encoder_factory,
q_func_factory,
gamma,
tau,
n_critics,
bootstrap,
share_encoder,
initial_temperature,
scaler,
action_scaler,
augmentation,
):
impl = SACImpl(
observation_shape,
action_size,
actor_learning_rate,
critic_learning_rate,
temp_learning_rate,
actor_optim_factory,
critic_optim_factory,
temp_optim_factory,
encoder_factory,
encoder_factory,
create_q_func_factory(q_func_factory),
gamma,
tau,
n_critics,
bootstrap,
share_encoder,
initial_temperature,
use_gpu=False,
scaler=scaler,
action_scaler=action_scaler,
augmentation=augmentation,
)
torch_impl_tester(
impl, discrete=False, deterministic_best_action=q_func_factory != "iqn"
)
def test_crr_impl(
observation_shape,
action_size,
actor_learning_rate,
critic_learning_rate,
actor_optim_factory,
critic_optim_factory,
encoder_factory,
q_func_factory,
gamma,
beta,
n_action_samples,
advantage_type,
weight_type,
max_weight,
n_critics,
target_reduction_type,
scaler,
action_scaler,
augmentation,
):
impl = CRRImpl(
observation_shape=observation_shape,
action_size=action_size,
actor_learning_rate=actor_learning_rate,
critic_learning_rate=critic_learning_rate,
actor_optim_factory=actor_optim_factory,
critic_optim_factory=critic_optim_factory,
actor_encoder_factory=encoder_factory,
critic_encoder_factory=encoder_factory,
q_func_factory=create_q_func_factory(q_func_factory),
gamma=gamma,
beta=beta,
n_action_samples=n_action_samples,
advantage_type=advantage_type,
weight_type=weight_type,
max_weight=max_weight,
n_critics=n_critics,
target_reduction_type=target_reduction_type,
use_gpu=None,
scaler=scaler,
action_scaler=action_scaler,
augmentation=augmentation,
)
torch_impl_tester(impl, discrete=False, deterministic_best_action=False)
def test_td3_plus_bc_impl(
observation_shape,
action_size,
actor_learning_rate,
critic_learning_rate,
actor_optim_factory,
critic_optim_factory,
encoder_factory,
q_func_factory,
gamma,
tau,
n_critics,
target_reduction_type,
target_smoothing_sigma,
target_smoothing_clip,
alpha,
scaler,
action_scaler,
reward_scaler,
):
impl = TD3PlusBCImpl(
observation_shape=observation_shape,
action_size=action_size,
actor_learning_rate=actor_learning_rate,
critic_learning_rate=critic_learning_rate,
actor_optim_factory=actor_optim_factory,
critic_optim_factory=critic_optim_factory,
actor_encoder_factory=encoder_factory,
critic_encoder_factory=encoder_factory,
q_func_factory=create_q_func_factory(q_func_factory),
gamma=gamma,
tau=tau,
n_critics=n_critics,
target_reduction_type=target_reduction_type,
target_smoothing_sigma=target_smoothing_sigma,
target_smoothing_clip=target_smoothing_clip,
alpha=alpha,
use_gpu=None,
scaler=scaler,
action_scaler=action_scaler,
reward_scaler=reward_scaler,
)
torch_impl_tester(impl,
discrete=False,
deterministic_best_action=q_func_factory != "iqn")
def test_awac_impl(
observation_shape,
action_size,
actor_learning_rate,
critic_learning_rate,
actor_optim_factory,
critic_optim_factory,
encoder_factory,
q_func_factory,
gamma,
tau,
lam,
n_action_samples,
max_weight,
n_critics,
target_reduction_type,
scaler,
action_scaler,
reward_scaler,
):
impl = AWACImpl(
observation_shape=observation_shape,
action_size=action_size,
actor_learning_rate=actor_learning_rate,
critic_learning_rate=critic_learning_rate,
actor_optim_factory=actor_optim_factory,
critic_optim_factory=critic_optim_factory,
actor_encoder_factory=encoder_factory,
critic_encoder_factory=encoder_factory,
q_func_factory=create_q_func_factory(q_func_factory),
gamma=gamma,
tau=tau,
lam=lam,
n_action_samples=n_action_samples,
max_weight=max_weight,
n_critics=n_critics,
target_reduction_type=target_reduction_type,
use_gpu=None,
scaler=scaler,
action_scaler=action_scaler,
reward_scaler=reward_scaler,
)
torch_impl_tester(
impl, discrete=False, deterministic_best_action=q_func_factory != "iqn"
)
def test_sac_impl(
observation_shape,
action_size,
actor_learning_rate,
critic_learning_rate,
temp_learning_rate,
actor_optim_factory,
critic_optim_factory,
temp_optim_factory,
encoder_factory,
q_func_factory,
gamma,
tau,
n_critics,
target_reduction_type,
initial_temperature,
scaler,
action_scaler,
augmentation,
):
impl = SACImpl(
observation_shape=observation_shape,
action_size=action_size,
actor_learning_rate=actor_learning_rate,
critic_learning_rate=critic_learning_rate,
temp_learning_rate=temp_learning_rate,
actor_optim_factory=actor_optim_factory,
critic_optim_factory=critic_optim_factory,
temp_optim_factory=temp_optim_factory,
actor_encoder_factory=encoder_factory,
critic_encoder_factory=encoder_factory,
q_func_factory=create_q_func_factory(q_func_factory),
gamma=gamma,
tau=tau,
n_critics=n_critics,
target_reduction_type=target_reduction_type,
initial_temperature=initial_temperature,
use_gpu=None,
scaler=scaler,
action_scaler=action_scaler,
augmentation=augmentation,
)
torch_impl_tester(impl,
discrete=False,
deterministic_best_action=q_func_factory != "iqn")
def test_discrete_sac_impl(
observation_shape,
action_size,
actor_learning_rate,
critic_learning_rate,
temp_learning_rate,
actor_optim_factory,
critic_optim_factory,
temp_optim_factory,
encoder_factory,
q_func_factory,
gamma,
n_critics,
initial_temperature,
scaler,
reward_scaler,
):
impl = DiscreteSACImpl(
observation_shape=observation_shape,
action_size=action_size,
actor_learning_rate=actor_learning_rate,
critic_learning_rate=critic_learning_rate,
temp_learning_rate=temp_learning_rate,
actor_optim_factory=actor_optim_factory,
critic_optim_factory=critic_optim_factory,
temp_optim_factory=temp_optim_factory,
actor_encoder_factory=encoder_factory,
critic_encoder_factory=encoder_factory,
q_func_factory=create_q_func_factory(q_func_factory),
gamma=gamma,
n_critics=n_critics,
initial_temperature=initial_temperature,
use_gpu=None,
scaler=scaler,
reward_scaler=reward_scaler,
)
torch_impl_tester(
impl, discrete=True, deterministic_best_action=q_func_factory != "iqn"
)
def test_discrete_bcq_impl(
observation_shape,
action_size,
learning_rate,
optim_factory,
encoder_factory,
q_func_factory,
gamma,
n_critics,
bootstrap,
share_encoder,
target_reduction_type,
action_flexibility,
beta,
scaler,
augmentation,
):
impl = DiscreteBCQImpl(
observation_shape,
action_size,
learning_rate,
optim_factory,
encoder_factory,
create_q_func_factory(q_func_factory),
gamma,
n_critics,
bootstrap,
share_encoder,
target_reduction_type,
action_flexibility,
beta,
use_gpu=False,
scaler=scaler,
augmentation=augmentation,
)
torch_impl_tester(impl,
discrete=True,
deterministic_best_action=q_func_factory != "iqn")
def test_bcq_impl(
observation_shape,
action_size,
actor_learning_rate,
critic_learning_rate,
imitator_learning_rate,
actor_optim_factory,
critic_optim_factory,
imitator_optim_factory,
encoder_factory,
q_func_factory,
gamma,
tau,
n_critics,
lam,
n_action_samples,
action_flexibility,
latent_size,
beta,
scaler,
action_scaler,
):
impl = BCQImpl(
observation_shape=observation_shape,
action_size=action_size,
actor_learning_rate=actor_learning_rate,
critic_learning_rate=critic_learning_rate,
imitator_learning_rate=imitator_learning_rate,
actor_optim_factory=actor_optim_factory,
critic_optim_factory=critic_optim_factory,
imitator_optim_factory=imitator_optim_factory,
actor_encoder_factory=encoder_factory,
critic_encoder_factory=encoder_factory,
imitator_encoder_factory=encoder_factory,
q_func_factory=create_q_func_factory(q_func_factory),
gamma=gamma,
tau=tau,
n_critics=n_critics,
lam=lam,
n_action_samples=n_action_samples,
action_flexibility=action_flexibility,
latent_size=latent_size,
beta=beta,
use_gpu=None,
scaler=scaler,
action_scaler=action_scaler,
)
impl.build()
# test internal methods
x = torch.rand(32, *observation_shape)
repeated_x = impl._repeat_observation(x)
assert repeated_x.shape == (32, n_action_samples) + observation_shape
action = impl._sample_repeated_action(repeated_x)
assert action.shape == (32, n_action_samples, action_size)
value = impl._predict_value(repeated_x, action)
assert value.shape == (n_critics, 32 * n_action_samples, 1)
best_action = impl._predict_best_action(x)
assert best_action.shape == (32, action_size)
torch_impl_tester(impl, discrete=False, deterministic_best_action=False)
def test_bear_impl(
observation_shape,
action_size,
actor_learning_rate,
critic_learning_rate,
imitator_learning_rate,
temp_learning_rate,
alpha_learning_rate,
actor_optim_factory,
critic_optim_factory,
imitator_optim_factory,
temp_optim_factory,
alpha_optim_factory,
encoder_factory,
q_func_factory,
gamma,
tau,
n_critics,
initial_temperature,
initial_alpha,
alpha_threshold,
lam,
n_action_samples,
mmd_kernel,
mmd_sigma,
scaler,
action_scaler,
augmentation,
):
impl = BEARImpl(
observation_shape=observation_shape,
action_size=action_size,
actor_learning_rate=actor_learning_rate,
critic_learning_rate=critic_learning_rate,
imitator_learning_rate=imitator_learning_rate,
temp_learning_rate=temp_learning_rate,
alpha_learning_rate=alpha_learning_rate,
actor_optim_factory=actor_optim_factory,
critic_optim_factory=critic_optim_factory,
imitator_optim_factory=imitator_optim_factory,
temp_optim_factory=temp_optim_factory,
alpha_optim_factory=alpha_optim_factory,
actor_encoder_factory=encoder_factory,
critic_encoder_factory=encoder_factory,
imitator_encoder_factory=encoder_factory,
q_func_factory=create_q_func_factory(q_func_factory),
gamma=gamma,
tau=tau,
n_critics=n_critics,
initial_temperature=initial_temperature,
initial_alpha=initial_alpha,
alpha_threshold=alpha_threshold,
lam=lam,
n_action_samples=n_action_samples,
mmd_kernel=mmd_kernel,
mmd_sigma=mmd_sigma,
use_gpu=None,
scaler=scaler,
action_scaler=action_scaler,
augmentation=augmentation,
)
impl.build()
x = torch.rand(32, *observation_shape)
target = impl.compute_target(x)
if q_func_factory == "mean":
assert target.shape == (32, 1)
else:
n_quantiles = impl._q_func.q_funcs[0]._n_quantiles
assert target.shape == (32, n_quantiles)
torch_impl_tester(impl,
discrete=False,
deterministic_best_action=q_func_factory != "iqn")
def test_bear_impl(
observation_shape,
action_size,
actor_learning_rate,
critic_learning_rate,
imitator_learning_rate,
temp_learning_rate,
alpha_learning_rate,
actor_optim_factory,
critic_optim_factory,
imitator_optim_factory,
temp_optim_factory,
alpha_optim_factory,
encoder_factory,
q_func_factory,
gamma,
tau,
n_critics,
initial_temperature,
initial_alpha,
alpha_threshold,
lam,
n_action_samples,
n_target_samples,
n_mmd_action_samples,
mmd_kernel,
mmd_sigma,
vae_kl_weight,
scaler,
action_scaler,
):
impl = BEARImpl(
observation_shape=observation_shape,
action_size=action_size,
actor_learning_rate=actor_learning_rate,
critic_learning_rate=critic_learning_rate,
imitator_learning_rate=imitator_learning_rate,
temp_learning_rate=temp_learning_rate,
alpha_learning_rate=alpha_learning_rate,
actor_optim_factory=actor_optim_factory,
critic_optim_factory=critic_optim_factory,
imitator_optim_factory=imitator_optim_factory,
temp_optim_factory=temp_optim_factory,
alpha_optim_factory=alpha_optim_factory,
actor_encoder_factory=encoder_factory,
critic_encoder_factory=encoder_factory,
imitator_encoder_factory=encoder_factory,
q_func_factory=create_q_func_factory(q_func_factory),
gamma=gamma,
tau=tau,
n_critics=n_critics,
initial_temperature=initial_temperature,
initial_alpha=initial_alpha,
alpha_threshold=alpha_threshold,
lam=lam,
n_action_samples=n_action_samples,
n_target_samples=n_target_samples,
n_mmd_action_samples=n_mmd_action_samples,
mmd_kernel=mmd_kernel,
mmd_sigma=mmd_sigma,
vae_kl_weight=vae_kl_weight,
use_gpu=None,
scaler=scaler,
action_scaler=action_scaler,
)
impl.build()
torch_impl_tester(impl, discrete=False, deterministic_best_action=False)
def test_bcq_impl(
observation_shape,
action_size,
actor_learning_rate,
critic_learning_rate,
imitator_learning_rate,
actor_optim_factory,
critic_optim_factory,
imitator_optim_factory,
encoder_factory,
q_func_factory,
gamma,
tau,
n_critics,
bootstrap,
share_encoder,
lam,
n_action_samples,
action_flexibility,
latent_size,
beta,
scaler,
action_scaler,
augmentation,
):
impl = BCQImpl(
observation_shape,
action_size,
actor_learning_rate,
critic_learning_rate,
imitator_learning_rate,
actor_optim_factory,
critic_optim_factory,
imitator_optim_factory,
encoder_factory,
encoder_factory,
encoder_factory,
create_q_func_factory(q_func_factory),
gamma,
tau,
n_critics,
bootstrap,
share_encoder,
lam,
n_action_samples,
action_flexibility,
latent_size,
beta,
use_gpu=False,
scaler=scaler,
action_scaler=action_scaler,
augmentation=augmentation,
)
impl.build()
# test internal methods
x = torch.rand(32, *observation_shape)
repeated_x = impl._repeat_observation(x)
assert repeated_x.shape == (32, n_action_samples) + observation_shape
action = impl._sample_repeated_action(repeated_x)
assert action.shape == (32, n_action_samples, action_size)
value = impl._predict_value(repeated_x, action)
assert value.shape == (n_critics, 32 * n_action_samples, 1)
target = impl.compute_target(x)
if q_func_factory == "mean":
assert target.shape == (32, 1)
else:
assert target.shape == (32, impl._q_func.q_funcs[0]._n_quantiles)
best_action = impl._predict_best_action(x)
assert best_action.shape == (32, action_size)
torch_impl_tester(impl, discrete=False, deterministic_best_action=False)
