def __init__(self,
env_spec: EnvSpec,
batch_size: int,
reward_multiplier: float,
lr: float = 3e-3,
logger: StepLogger = None,
device: str = 'cuda' if to.cuda.is_available() else 'cpu'):
"""
Constructor
:param env_spec: environment specification
:param batch_size: batch size for each update step
:param reward_multiplier: factor for the predicted probability
:param lr: learning rate
:param logger: logger for every step of the algorithm, if `None` the default logger will be created
"""
self.device = device
self.batch_size = batch_size
self.reward_multiplier = reward_multiplier
self.lr = lr
spec = EnvSpec(obs_space=BoxSpace.cat(
[env_spec.obs_space, env_spec.obs_space, env_spec.act_space]),
act_space=BoxSpace(bound_lo=[0], bound_up=[1]))
self.discriminator = FNNPolicy(spec=spec,
hidden_nonlin=to.tanh,
hidden_sizes=[62],
output_nonlin=to.sigmoid)
self.loss_fcn = nn.BCELoss()
self.optimizer = to.optim.Adam(self.discriminator.parameters(),
lr=lr,
eps=1e-5)
self.logger = logger
def test_cat_box_space(bs_list):
bs_cat = BoxSpace.cat(bs_list)
assert isinstance(bs_cat, BoxSpace)
assert bs_cat.flat_dim == sum([bs.flat_dim for bs in bs_list])
def obs_space(self):
outer_space = self.wrapped_env.obs_space
augmented_space = BoxSpace(0.5*self._nominal, 1.5*self._nominal, [self._nominal.shape[0]], self._params)
return BoxSpace.cat((outer_space, augmented_space))
