ibl_bwb_eval.metrics#

Metric definitions the benchmark reports.

These are the metrics a submission is scored on, so they live with the scorer rather than with the training code, and training-time logging imports them from here too: a change to one moves every number already reported, which importing them by way of core would hide. core.nn.metrics keeps only diagnostics that are scored by nothing.

BPS

Bits Per Spike metric for Poisson spiking data.

PoissonD2Score

Minimal implementation of Cohen pseudo-R^2 score for Poisson NLL loss.

aggregate_metrics

Compute and flatten a dict of torchmetrics metrics to plain Python scalars.