org.deeplearning4j.nn.weights

Interface IWeightInit

All Superinterfaces:

Serializable

All Known Implementing Classes:

WeightInitConstant, WeightInitDistribution, WeightInitEmbedding, WeightInitIdentity, WeightInitLecunUniform, WeightInitNormal, WeightInitRelu, WeightInitReluUniform, WeightInitSigmoidUniform, WeightInitUniform, WeightInitVarScalingNormalFanAvg, WeightInitVarScalingNormalFanIn, WeightInitVarScalingNormalFanOut, WeightInitVarScalingUniformFanAvg, WeightInitVarScalingUniformFanIn, WeightInitVarScalingUniformFanOut, WeightInitXavier, WeightInitXavierLegacy, WeightInitXavierUniform

public interface IWeightInit
extends Serializable

Interface for weight initialization.

Author:

Christian Skarby

Field Summary

Fields

Modifier and Type	Field and Description
static char	DEFAULT_WEIGHT_INIT_ORDER

Method Summary

Modifier and Type	Method and Description
INDArray	init(double fanIn, double fanOut, long[] shape, char order, INDArray paramView) Initialize parameters in the given view.

Field Detail

DEFAULT_WEIGHT_INIT_ORDER

static final char DEFAULT_WEIGHT_INIT_ORDER

See Also:

Constant Field Values

Method Detail

init

INDArray init(double fanIn,
              double fanOut,
              long[] shape,
              char order,
              INDArray paramView)

Initialize parameters in the given view. Double values are used for fanIn and fanOut as some layers (convolution with stride) results in a non-integer number which may be truncated to zero in certain configurations

Parameters:

fanIn - Number of input parameters

fanOut - Number of output parameters

shape - Desired shape of array (users shall assume paramView has this shape after method has finished)

order - Order of array, e.g. Fortran ('f') or C ('c')

paramView - View of parameters to initialize (and reshape)