optax.scale_by_rprop

Contents

optax.scale_by_rprop#

optax.scale_by_rprop(learning_rate: jax.typing.ArrayLike, eta_minus: jax.typing.ArrayLike = 0.5, eta_plus: jax.typing.ArrayLike = 1.2, min_step_size: jax.typing.ArrayLike = 1e-06, max_step_size: jax.typing.ArrayLike = 50.0) → optax.GradientTransformation[source]#

Scale with the Rprop optimizer.

See optax.rprop() for more details.

Parameters:

  • learning_rate – The initial step size.

  • eta_minus – Multiplicative factor for decreasing step size. This is applied when the gradient changes sign from one step to the next.

  • eta_plus – Multiplicative factor for increasing step size. This is applied when the gradient has the same sign from one step to the next.

  • min_step_size – Minimum allowed step size. Smaller steps will be clipped to this value.

  • max_step_size – Maximum allowed step size. Larger steps will be clipped to this value.

Returns:

The corresponding optax.GradientTransformation.

Contents