Quantities

Cockpit offers a large set of so called quantities that can be efficiently tracked during the training process.

Loss(track_schedule[, verbose])

Loss Quantity class tracking the mini-batch training loss during training.

Parameters(track_schedule[, verbose])

Parameter Quantitiy class tracking the current parameters in each iteration.

Distance(track_schedule[, verbose])

Distance Quantity class tracking distance of the parameters from their init.

UpdateSize(track_schedule[, verbose])

Quantity class for tracking parameter update sizes.

GradNorm(track_schedule[, verbose])

Quantitiy Class for tracking the norm of the mean gradient.

Time(track_schedule[, verbose])

Time Quantity Class tracking the time during training.

Alpha(track_schedule[, verbose])

Alpha Quantity class for the normalized step length α.

CABS(track_schedule[, verbose])

CABS Quantity class for the suggested batch size using the CABS criterion.

EarlyStopping(track_schedule[, verbose, epsilon])

Quantity class for the evidence-based early-stopping criterion.

GradHist1d(track_schedule[, verbose, bins, ...])

Quantity class for one-dimensional histograms of indivdual gradient elements.

GradHist2d(track_schedule[, verbose, bins, ...])

Quantity class for two-dimensional histograms over gradient and parameters.

NormTest(track_schedule[, verbose])

Quantitiy Class for the norm test.

InnerTest(track_schedule[, verbose])

Quantitiy Class for tracking the result of the inner product test.

OrthoTest(track_schedule[, verbose])

Quantity Class for the orthogonality test.

HessMaxEV(track_schedule[, verbose, use_power])

Quantity Class tracking the Hessian's largest eigenvalue.

HessTrace(track_schedule[, verbose, curvature])

Quantitiy Class tracking the trace of the Hessian during training.

TICDiag(track_schedule[, verbose, ...])

Quantity class for tracking the TIC using diagonal curvature approximation.

TICTrace(track_schedule[, verbose, ...])

Quantity class for the TIC using the trace of curvature and gradient covariance.

MeanGSNR(track_schedule[, verbose, epsilon])

Quantitiy Class for the mean gradient signal-to-noise ratio (GSNR).