org.nd4j.linalg.api.ops.impl.scatter.ScatterUpdate 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.api.ops.impl.scatter;
import org.nd4j.autodiff.samediff.SDVariable;
import org.nd4j.autodiff.samediff.SameDiff;
import org.nd4j.imports.NoOpNameFoundException;
import org.nd4j.linalg.api.ops.DynamicCustomOp;
import java.util.ArrayList;
import java.util.List;
/**
* Scatter update op
*
* @author Alex Black
*/
public class ScatterUpdate extends DynamicCustomOp {
public ScatterUpdate(SameDiff sameDiff, SDVariable ref, SDVariable indices, SDVariable updates) {
super(null, sameDiff, new SDVariable[]{ref, indices, updates}, false);
}
public ScatterUpdate(){}
@Override
public String opName() {
return "scatter_upd";
}
@Override
public String onnxName() {
throw new NoOpNameFoundException("No onnx op opName found for " + opName());
}
@Override
public String tensorflowName() {
return "ScatterUpdate";
}
@Override
public List doDiff(List gradOut){
//3 args: ref, indices, updates
//For non-modified indices, input gradient (reference) is same as output gradient
//For modified indices, dL/dref = dL/dOut * dOut/dRef = dL/dOut * d(update)/dRef = 0
//And for updates, dL/du = dL/dOut * dOut/du = dL/dOut * d(update)/du = dL/dOut -> gather op
SDVariable indices = arg(1);
SDVariable updates = arg(2);
List ret = new ArrayList<>(3);
SDVariable zerosUpdate = f().zerosLike(updates);
SDVariable gradRef = f().scatterMul(gradOut.get(0), indices, zerosUpdate); //TODO optimize
ret.add(gradRef); //Reference array gradient
ret.add(f().zerosLike(arg(1))); //Indices
SDVariable gather = f().gather(gradOut.get(0), indices, 0); //Updates
ret.add(gather);
return ret;
}
}