org.apache.ignite.ml.nn

Class UpdatesStrategy<M,U extends Serializable>

java.lang.Object

org.apache.ignite.ml.nn.UpdatesStrategy<M,U>

Type Parameters:

M - Type of model to be optimized.

U - Type of update.

public class UpdatesStrategy<M,U extends Serializable>
extends Object

Class encapsulating update strategies for group trainers based on updates.

Constructor Summary

Constructors

Constructor and Description
UpdatesStrategy(ParameterUpdateCalculator<M,U> updatesCalculator, IgniteFunction<List<U>,U> locStepUpdatesReducer, IgniteFunction<List<U>,U> allUpdatesReducer) Construct instance of this class with given parameters.

Method Summary

Modifier and Type	Method and Description
IgniteFunction<List<U>,U>	allUpdatesReducer() Get function used to reduce updates from different trainings (for example, averaging of gradients of all parallel trainings).
ParameterUpdateCalculator<M,U>	getUpdatesCalculator() Get parameter update calculator (see ParameterUpdateCalculator).
IgniteFunction<List<U>,U>	locStepUpdatesReducer() Get function used to reduce updates in one training (for example, sum all sequential gradient updates to get one gradient update).

Methods inherited from class java.lang.Object

clone, equals, finalize, getClass, hashCode, notify, notifyAll, toString, wait, wait, wait

Constructor Detail

UpdatesStrategy

public UpdatesStrategy(ParameterUpdateCalculator<M,U> updatesCalculator,
                       IgniteFunction<List<U>,U> locStepUpdatesReducer,
                       IgniteFunction<List<U>,U> allUpdatesReducer)

Construct instance of this class with given parameters.

Parameters:

updatesCalculator - Parameter update calculator.

locStepUpdatesReducer - Function used to reduce updates in one training (for example, sum all sequential gradient updates to get one gradient update).

allUpdatesReducer - Function used to reduce updates from different trainings (for example, averaging of gradients of all parallel trainings).

Method Detail

getUpdatesCalculator

public ParameterUpdateCalculator<M,U> getUpdatesCalculator()

Get parameter update calculator (see ParameterUpdateCalculator).

Returns:

Parameter update calculator.

locStepUpdatesReducer

public IgniteFunction<List<U>,U> locStepUpdatesReducer()

Get function used to reduce updates in one training (for example, sum all sequential gradient updates to get one gradient update).

Returns:

Function used to reduce updates in one training (for example, sum all sequential gradient updates to get on gradient update).

allUpdatesReducer

public IgniteFunction<List<U>,U> allUpdatesReducer()

Get function used to reduce updates from different trainings (for example, averaging of gradients of all parallel trainings).

Returns:

Function used to reduce updates from different trainings.