org.nd4j.linalg.activations.IActivation Maven / Gradle / Ivy
package org.nd4j.linalg.activations;
import org.nd4j.linalg.api.ndarray.INDArray;
import org.nd4j.linalg.primitives.Pair;
import org.nd4j.serde.json.LegacyIActivationDeserializerHelper;
import org.nd4j.shade.jackson.annotation.JsonAutoDetect;
import org.nd4j.shade.jackson.annotation.JsonTypeInfo;
import java.io.Serializable;
/**
* Interface for implementing custom activation functions
*/
@JsonTypeInfo(use = JsonTypeInfo.Id.CLASS, include = JsonTypeInfo.As.PROPERTY, property = "@class",
defaultImpl = LegacyIActivationDeserializerHelper.class)
@JsonAutoDetect(fieldVisibility = JsonAutoDetect.Visibility.ANY, getterVisibility = JsonAutoDetect.Visibility.NONE,
setterVisibility = JsonAutoDetect.Visibility.NONE)
public interface IActivation extends Serializable {
/**
* Carry out activation function on the input array (usually known as 'preOut' or 'z')
* Implementations must overwrite "in", transform in place and return "in"
* Can support separate behaviour during test
* @param in
* @param training
* @return transformed activation
*/
INDArray getActivation(INDArray in, boolean training);
/**
* Backpropagate the errors through the activation function, given input z and epsilon dL/da.
* Returns 2 INDArrays:
* (a) The gradient dL/dz, calculated from dL/da, and
* (b) The parameter gradients dL/dw, where w is the weights in the activation function. For activation functions
* with no gradients, this will be null.
*
* @param in Input, before applying the activation function (z, or 'preOut')
* @param epsilon Gradient to be backpropagated: dL/da, where L is the loss function
* @return dL/dz and dL/dw, for weights w (null if activatino function has no weights)
*/
Pair backprop(INDArray in, INDArray epsilon);
int numParams(int inputSize);
void setParametersViewArray(INDArray viewArray, boolean initialize);
INDArray getParametersViewArray();
void setGradientViewArray(INDArray viewArray);
INDArray getGradientViewArray();
}