org.nd4j.autodiff.samediff

Class TrainingConfig

java.lang.Object

org.nd4j.autodiff.samediff.TrainingConfig

public class TrainingConfig
extends Object

TrainingConfig is a simple configuration class for defining settings for training a SameDiff instance.
It defines the following settings:

• The IUpdater to use (i.e., Adam, Nesterovs etc. The IUpdater instance is also how the learning rate (or learning rate schedule) is set.
• The L1 and L2 regularization coefficients (set to 0.0 by default)
• The DataSet feature and label mapping - which defines how the feature/label arrays from the DataSet/MultiDataSet should be associated with SameDiff variables (usually placeholders)

The TrainingConfig instance also stores the iteration count and the epoch count - these values are updated during training and are used for example in learning rate schedules.

Author:

Alex Black

Nested Class Summary

Nested Classes

Modifier and Type	Class and Description
static class	TrainingConfig.Builder

Constructor Summary

Constructors

Modifier	Constructor and Description
	TrainingConfig(IUpdater updater, List<Regularization> regularization, boolean minimize, List<String> dataSetFeatureMapping, List<String> dataSetLabelMapping, List<String> dataSetFeatureMaskMapping, List<String> dataSetLabelMaskMapping, List<String> lossVariables) Create a training configuration suitable for training both single input/output and multi input/output networks. See also the TrainingConfig.Builder for creating a TrainingConfig
protected	TrainingConfig(IUpdater updater, List<Regularization> regularization, boolean minimize, List<String> dataSetFeatureMapping, List<String> dataSetLabelMapping, List<String> dataSetFeatureMaskMapping, List<String> dataSetLabelMaskMapping, List<String> lossVariables, Map<String,List<IEvaluation>> trainEvaluations, Map<String,Integer> trainEvaluationLabels, Map<String,List<IEvaluation>> validationEvaluations, Map<String,Integer> validationEvaluationLabels)
	TrainingConfig(IUpdater updater, List<Regularization> regularization, String dataSetFeatureMapping, String dataSetLabelMapping) Create a training configuration suitable for training a single input, single output network. See also the TrainingConfig.Builder for creating a TrainingConfig

Method Summary

Modifier and Type	Method and Description
static TrainingConfig.Builder	builder()
static TrainingConfig	fromJson(@NonNull String json)
void	incrementEpochCount() Increment the epoch count by 1
void	incrementIterationCount() Increment the iteration count by 1
int	labelIdx(String s) Get the index of the label array that the specified variable is associated with
static void	removeInstances(List<?> list, Class<?> remove) Remove any instances of the specified type from the list.
static void	removeInstancesWithWarning(List<?> list, Class<?> remove, String warning)
String	toJson()

Methods inherited from class java.lang.Object

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

Constructor Detail

TrainingConfig

public TrainingConfig(IUpdater updater,
                      List<Regularization> regularization,
                      String dataSetFeatureMapping,
                      String dataSetLabelMapping)

Create a training configuration suitable for training a single input, single output network.
See also the TrainingConfig.Builder for creating a TrainingConfig

Parameters:

updater - The updater configuration to use

dataSetFeatureMapping - The name of the placeholder/variable that should be set using the feature INDArray from the DataSet (or the first/only feature from a MultiDataSet). For example, if the network input placeholder was called "input" then this should be set to "input"

dataSetLabelMapping - The name of the placeholder/variable that should be set using the label INDArray from the DataSet (or the first/only feature from a MultiDataSet). For example, if the network input placeholder was called "input" then this should be set to "input"

TrainingConfig

public TrainingConfig(IUpdater updater,
                      List<Regularization> regularization,
                      boolean minimize,
                      List<String> dataSetFeatureMapping,
                      List<String> dataSetLabelMapping,
                      List<String> dataSetFeatureMaskMapping,
                      List<String> dataSetLabelMaskMapping,
                      List<String> lossVariables)

Create a training configuration suitable for training both single input/output and multi input/output networks.
See also the TrainingConfig.Builder for creating a TrainingConfig

Parameters:

updater - The updater configuration to use

regularization - Regularization for all trainable parameters;\

minimize - Set to true if the loss function should be minimized (usually true). False to maximize

dataSetFeatureMapping - The name of the placeholders/variables that should be set using the feature INDArray(s) from the DataSet or MultiDataSet. For example, if the network had 2 inputs called "input1" and "input2" and the MultiDataSet features should be mapped with MultiDataSet.getFeatures(0)->"input1" and MultiDataSet.getFeatures(1)->"input2", then this should be set to List<>("input1", "input2").

dataSetLabelMapping - As per dataSetFeatureMapping, but for the DataSet/MultiDataSet labels

dataSetFeatureMaskMapping - May be null. If non-null, the variables that the MultiDataSet feature mask arrays should be associated with.

dataSetLabelMaskMapping - May be null. If non-null, the variables that the MultiDataSet label mask arrays should be associated with.

TrainingConfig

protected TrainingConfig(IUpdater updater,
                         List<Regularization> regularization,
                         boolean minimize,
                         List<String> dataSetFeatureMapping,
                         List<String> dataSetLabelMapping,
                         List<String> dataSetFeatureMaskMapping,
                         List<String> dataSetLabelMaskMapping,
                         List<String> lossVariables,
                         Map<String,List<IEvaluation>> trainEvaluations,
                         Map<String,Integer> trainEvaluationLabels,
                         Map<String,List<IEvaluation>> validationEvaluations,
                         Map<String,Integer> validationEvaluationLabels)

Method Detail

incrementIterationCount

public void incrementIterationCount()

Increment the iteration count by 1

incrementEpochCount

public void incrementEpochCount()

Increment the epoch count by 1

builder

public static TrainingConfig.Builder builder()

labelIdx

public int labelIdx(String s)

Get the index of the label array that the specified variable is associated with

Parameters:

s - Name of the variable

Returns:

The index of the label variable, or -1 if not found

removeInstances

public static void removeInstances(List<?> list,
                                   Class<?> remove)

Remove any instances of the specified type from the list. This includes any subtypes.

Parameters:

list - List. May be null

remove - Type of objects to remove

removeInstancesWithWarning

public static void removeInstancesWithWarning(List<?> list,
                                              Class<?> remove,
                                              String warning)

toJson

public String toJson()

fromJson

public static TrainingConfig fromJson(@NonNull
                                      @NonNull String json)