def test_default_encoder_factory(observation_shape, action_size,
discrete_action):
factory = DefaultEncoderFactory()
# test state encoder
encoder = factory.create(observation_shape)
if len(observation_shape) == 3:
assert isinstance(encoder, PixelEncoder)
else:
assert isinstance(encoder, VectorEncoder)
# test state-action encoder
encoder = factory.create_with_action(observation_shape, action_size,
discrete_action)
if len(observation_shape) == 3:
assert isinstance(encoder, PixelEncoderWithAction)
else:
assert isinstance(encoder, VectorEncoderWithAction)
assert encoder._discrete_action == discrete_action
# test get_params
assert factory.get_type() == "default"
params = factory.get_params()
new_factory = DefaultEncoderFactory(**params)
assert new_factory.get_params() == params
DummyScaler,
DummyActionScaler,
)
@pytest.mark.parametrize("observation_shape", [(100, ), (4, 84, 84)])
@pytest.mark.parametrize("action_size", [2])
@pytest.mark.parametrize("actor_learning_rate", [1e-3])
@pytest.mark.parametrize("critic_learning_rate", [1e-3])
@pytest.mark.parametrize("temp_learning_rate", [1e-3])
@pytest.mark.parametrize("alpha_learning_rate", [1e-3])
@pytest.mark.parametrize("actor_optim_factory", [AdamFactory()])
@pytest.mark.parametrize("critic_optim_factory", [AdamFactory()])
@pytest.mark.parametrize("temp_optim_factory", [AdamFactory()])
@pytest.mark.parametrize("alpha_optim_factory", [AdamFactory()])
@pytest.mark.parametrize("encoder_factory", [DefaultEncoderFactory()])
@pytest.mark.parametrize("q_func_factory", ["mean", "qr", "iqn", "fqf"])
@pytest.mark.parametrize("gamma", [0.99])
@pytest.mark.parametrize("tau", [0.05])
@pytest.mark.parametrize("n_critics", [2])
@pytest.mark.parametrize("bootstrap", [False])
@pytest.mark.parametrize("share_encoder", [False, True])
@pytest.mark.parametrize("initial_temperature", [1.0])
@pytest.mark.parametrize("initial_alpha", [5.0])
@pytest.mark.parametrize("alpha_threshold", [10.0])
@pytest.mark.parametrize("n_action_samples", [10])
@pytest.mark.parametrize("soft_q_backup", [True])
@pytest.mark.parametrize("scaler", [None, DummyScaler()])
@pytest.mark.parametrize("action_scaler", [None, DummyActionScaler()])
@pytest.mark.parametrize("augmentation", [DrQPipeline()])
def test_cql_impl(
from d3rlpy.argument_utility import (
check_action_scaler,
check_encoder,
check_q_func,
check_scaler,
check_use_gpu,
)
from d3rlpy.gpu import Device
from d3rlpy.models.encoders import DefaultEncoderFactory
from d3rlpy.models.q_functions import MeanQFunctionFactory
from d3rlpy.preprocessing.action_scalers import MinMaxActionScaler
from d3rlpy.preprocessing.scalers import MinMaxScaler
@pytest.mark.parametrize("value", ["default", DefaultEncoderFactory()])
def test_check_encoder(value):
assert isinstance(check_encoder(value), DefaultEncoderFactory)
@pytest.mark.parametrize("value", ["mean", MeanQFunctionFactory()])
def test_check_q_func(value):
assert isinstance(check_q_func(value), MeanQFunctionFactory)
@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:
