reference/generated/torchrl.modules.functional.signed_hyperbolic

Run in Google Colab

Colab

Download Notebook

Notebook

View on GitHub

GitHub

signed_hyperbolic

class torchrl.modules.functional.signed_hyperbolic(value: Tensor, epsilon: float = 0.001)

Apply the signed hyperbolic value transform.

This is the scale-compressing transform introduced by Pohlen et al. and used by algorithms in the MuZero and Muesli families:

sign(value) * (sqrt(abs(value) + 1) - 1) + epsilon * value.

Parameters:

• value (torch.Tensor) – Input tensor.

• epsilon (float, optional) – Positive linear correction that keeps the inverse Lipschitz continuous. Defaults to 1e-3.

Returns:

A tensor with the same shape, dtype, and device as value.

Examples

>>> import torch
>>> from torchrl.modules import functional as F
>>> value = torch.tensor([-100.0, 0.0, 100.0])
>>> F.signed_hyperbolic(value)
tensor([-9.1499,  0.0000,  9.1499])

See also

signed_parabolic() for the inverse operation and SignedHyperbolicValueTransform for the module form.

Note

See Observe and Look Further: Achieving Consistent Performance on Atari (Pohlen et al., 2018).