Package ai.djl.training.tracker
package ai.djl.training.tracker
Contains classes for having a gradually changing hyper-parameter.
It contains a main interface Tracker and various options that
extend it.
-
ClassDescription
CosineTrackeris an implementation ofTrackerwhich is updated by taking sections of a cosine curve to smoothly reduce learning rate until a specified step and base learning rate.The Builder to construct anCosineTrackerobject.CyclicalTrackeris an implementation ofTrackerwhich is a policy of learning rate adjustment that increases the learning rate off a base value in a cyclical nature, as detailed in the paper Cyclical Learning Rates for Training Neural Networks.The Builder to construct anCyclicalTrackerobject.CyclicalTrackerprovides three predefined cyclical modes and can be selected by this enum.ScaleFunctionis an interface to implement a custom scale function.FactorTrackeris an implementation ofTrackerwhich is updated by a multiplicative factor.The Builder to construct anFactorTrackerobject.FixedPerVarTrackeris an implementation ofTrackerwhich returns a fixed value.The Builder to construct anFixedPerVarTrackerobject.FactorTrackeris an implementation ofTrackerwhich is updated by a constant factor.The Builder to construct anLinearTrackerobject.MultiFactorTrackeris an implementation ofTrackerwhich returns piecewise constant values for fixed numbers of steps.The Builder to construct anMultiFactorTrackerobject.ATrackerrepresents a collection of hyperparameters orTrackers that changes gradually through the training process.Polynomial decayTracker.Builder for PolynomialDecayTracker.ATrackerrepresents a hyperparameter that changes gradually through the training process.AWarmUpTrackerapplies a simple warm-up before executing a mainTracker.The Builder to construct aWarmUpTracker.An enum that enumerates the types of warm-up modes for aWarmUpTracker.