RewardClipping

class torchrl.envs.transforms.RewardClipping(clamp_min: Optional[float] = None, clamp_max: Optional[float] = None, in_keys: Optional[Sequence[NestedKey]] = None, out_keys: Optional[Sequence[NestedKey]] = None)

Clips the reward between clamp_min and clamp_max.

Parameters:

• clip_min (scalar) – minimum value of the resulting reward.

• clip_max (scalar) – maximum value of the resulting reward.

transform_reward_spec(reward_spec: TensorSpec) → TensorSpec

Transforms the reward spec such that the resulting spec matches transform mapping.

Parameters:

reward_spec (TensorSpec) – spec before the transform

Returns:

expected spec after the transform