org.nd4j.linalg.learning.AdamUpdater Maven / Gradle / Ivy
/*******************************************************************************
* Copyright (c) 2015-2018 Skymind, Inc.
*
* This program and the accompanying materials are made available under the
* terms of the Apache License, Version 2.0 which is available at
* https://www.apache.org/licenses/LICENSE-2.0.
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
* WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the
* License for the specific language governing permissions and limitations
* under the License.
*
* SPDX-License-Identifier: Apache-2.0
******************************************************************************/
package org.nd4j.linalg.learning;
import lombok.Data;
import org.apache.commons.math3.util.FastMath;
import org.nd4j.linalg.api.ndarray.INDArray;
import org.nd4j.linalg.api.shape.Shape;
import org.nd4j.linalg.indexing.NDArrayIndex;
import org.nd4j.linalg.learning.config.Adam;
import org.nd4j.linalg.ops.transforms.Transforms;
/**
* The Adam updater.
* http://arxiv.org/abs/1412.6980
*
* @author Adam Gibson
*/
@Data
public class AdamUpdater implements GradientUpdater {
private Adam config;
private INDArray m, v; // moving avg & sqrd gradients
private char gradientReshapeOrder;
public AdamUpdater(Adam config) {
this.config = config;
}
@Override
public void setStateViewArray(INDArray viewArray, long[] gradientShape, char gradientOrder, boolean initialize) {
if (!viewArray.isRowVector())
throw new IllegalArgumentException("Invalid input: expect row vector input");
if (initialize)
viewArray.assign(0);
long length = viewArray.length();
this.m = viewArray.get(NDArrayIndex.point(0), NDArrayIndex.interval(0, length / 2));
this.v = viewArray.get(NDArrayIndex.point(0), NDArrayIndex.interval(length / 2, length));
//Reshape to match the expected shape of the input gradient arrays
this.m = Shape.newShapeNoCopy(this.m, gradientShape, gradientOrder == 'f');
this.v = Shape.newShapeNoCopy(this.v, gradientShape, gradientOrder == 'f');
if (m == null || v == null)
throw new IllegalStateException("Could not correctly reshape gradient view arrays");
this.gradientReshapeOrder = gradientOrder;
}
/**
* Calculate the update based on the given gradient
*
* @param gradient the gradient to get the update for
* @param iteration
* @return the gradient
*/
@Override
public void applyUpdater(INDArray gradient, int iteration, int epoch) {
if (m == null || v == null)
throw new IllegalStateException("Updater has not been initialized with view state");
double beta1 = config.getBeta1();
double beta2 = config.getBeta2();
double learningRate = config.getLearningRate(iteration, epoch);
double epsilon = config.getEpsilon();
INDArray oneMinusBeta1Grad = gradient.mul(1.0 - beta1);
m.muli(beta1).addi(oneMinusBeta1Grad);
INDArray oneMinusBeta2GradSquared = gradient.mul(gradient).muli(1 - beta2);
v.muli(beta2).addi(oneMinusBeta2GradSquared);
double beta1t = FastMath.pow(beta1, iteration + 1);
double beta2t = FastMath.pow(beta2, iteration + 1);
double alphat = learningRate * FastMath.sqrt(1 - beta2t) / (1 - beta1t);
if (Double.isNaN(alphat) || alphat == 0.0)
alphat = epsilon;
INDArray sqrtV = Transforms.sqrt(v.dup(gradientReshapeOrder), false).addi(epsilon);
gradient.assign(m).muli(alphat).divi(sqrtV);
}
}