Others utils

Here, we document other utility functions that might come in handy for Scores development.

ak_lpe(index, k, query, *[, self_query])	Mean distance from \(\approx k\)-th nearest neighbors.
batched_grad(outputs, inputs, *[, retain_graph])	Compute gradients for batched inputs/outputs.
dirmult_surprise(counts, alpha)	Compute DCM surprise.
fgm_direction(grad, *[, p, check])	Return FGM attack direction.
get_aggregator(aggr_or_ord)	Get a \(1\)D aggregator function.
kldiv(inputs, expected)	KL div for (potentially batched) \(1\)D inputs.
knn_label_count(index, labels, n_classes, k, ...)	Count labels of neighbors for batched queries.
multinomial_test(counts, prior, mode)	Prepare test after having observed counts, with prior belief.
normalize_samples(tensor, *, ord[, ...])	Vector-normalizes all samples if ord is provided.
torch_quantile(tensor, q[, dim, keepdim, ...])	Improved torch.quantile() for one scalar quantile.