Skip to content

Optimization ​

livn.optimization is livn's differentiable ML layer containing metrics, losses, and fitting machinary used to estimate model parameters from data.

TIP

Naturally, this section assumes a differentiable environment requiring a JAX backend.

Installation ​

livn.optimization is an extra of livn:

sh
pip install livn[optimization]

It depends on livn[diffrax], so this also installs JAX, diffrax, equinox, and optax, plus grain for the data pipeline. The module works with whatever LIVN_BACKEND is set; only building a differentiable Env needs LIVN_BACKEND=diffrax.

Hard and soft ​

Metrics come in two flavours:

  • Hard metrics bin spike times onto a grid before comparing them. They are the numbers to report but the binning has no gradient, so they cannot be optimized against.
  • Soft metrics are differentiable approximations over a continuous rate or raster. They are usable as fitting terms, and they are not expected to reproduce their hard counterpart's value.

Released under the MIT License.