def build_serving_module(
self,
seq_len: int,
synthetic_reward_network: ModelBase,
state_normalization_data: NormalizationData,
action_normalization_data: Optional[NormalizationData] = None,
discrete_action_names: Optional[List[str]] = None,
) -> torch.nn.Module:
"""
Returns a TorchScript predictor module
"""
state_preprocessor = Preprocessor(
state_normalization_data.dense_normalization_parameters)
if not discrete_action_names:
assert action_normalization_data is not None
action_preprocessor = Preprocessor(
action_normalization_data.dense_normalization_parameters)
return SyntheticRewardPredictorWrapper(
seq_len,
state_preprocessor,
action_preprocessor,
# pyre-fixme[29]: `Union[torch.Tensor, torch.nn.Module]` is not a
# function.
synthetic_reward_network.export_mlp().cpu().eval(),
)
else:
# TODO add Discrete Single Step Synthetic Reward Predictor
return torch.jit.script(torch.nn.Linear(1, 1))
def build_serving_module(
self,
synthetic_reward_network: ModelBase,
state_normalization_data: NormalizationData,
action_normalization_data: Optional[NormalizationData] = None,
discrete_action_names: Optional[List[str]] = None,
) -> torch.nn.Module:
"""
Returns a TorchScript predictor module
"""
state_preprocessor = Preprocessor(
state_normalization_data.dense_normalization_parameters
)
if not discrete_action_names:
assert action_normalization_data is not None
action_preprocessor = Preprocessor(
action_normalization_data.dense_normalization_parameters
)
synthetic_reward_with_preprocessor = ParametricDqnWithPreprocessor(
# pyre-fixme[29]: `Union[torch.Tensor, torch.nn.Module]` is not a
# function.
synthetic_reward_network.export_mlp().cpu().eval(),
state_preprocessor,
action_preprocessor,
)
return ParametricSingleStepSyntheticRewardPredictorWrapper(
synthetic_reward_with_preprocessor
)
else:
raise NotImplementedError(
"Discrete Single Step Synthetic Reward Predictor has not been implemented"
)