losses.rst

Losses & Metrics

asteroid.losses

asteroid.losses

Permutation invariant training (PIT) made easy

Asteroid supports regular Permutation Invariant Training (PIT), it's extension using Sinkhorn algorithm (SinkPIT) as well as Mixture Invariant Training (MixIT).

PIT

asteroid.losses.pit_wrapper

MixIT

asteroid.losses.mixit_wrapper

SinkPIT

asteroid.losses.sinkpit_wrapper

Available loss functions

PITLossWrapper supports three types of loss function. For "easy" losses, we implement the three types (pairwise point, single-source loss and multi-source loss). For others, we only implement the single-source loss which can be aggregated into both PIT and nonPIT training.

MSE

asteroid.losses.mse.PairwiseMSE

asteroid.losses.mse.SingleSrcMSE

asteroid.losses.mse.MultiSrcMSE

SDR

asteroid.losses.sdr.PairwiseNegSDR

asteroid.losses.sdr.SingleSrcNegSDR

asteroid.losses.sdr.MultiSrcNegSDR

PMSQE

asteroid.losses.pmsqe.SingleSrcPMSQE

STOI

asteroid.losses.stoi.NegSTOILoss

MultiScale Spectral Loss

asteroid.losses.multi_scale_spectral.SingleSrcMultiScaleSpectral

Deep clustering (Affinity) loss

asteroid.losses.cluster.deep_clustering_loss

Computing metrics

asteroid.metrics.get_metrics