org.nd4j.autodiff.util.SameDiffUtils Maven / Gradle / Ivy
/*
* ******************************************************************************
* *
* *
* * 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.
* *
* * See the NOTICE file distributed with this work for additional
* * information regarding copyright ownership.
* * 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.autodiff.util;
import java.util.*;
import lombok.AccessLevel;
import lombok.NoArgsConstructor;
import org.nd4j.autodiff.functions.DifferentialFunction;
import org.nd4j.autodiff.samediff.SDVariable;
import org.nd4j.autodiff.samediff.SameDiff;
import org.nd4j.common.base.Preconditions;
import org.nd4j.linalg.api.ndarray.INDArray;
import org.nd4j.linalg.api.ops.impl.layers.ExternalErrorsFunction;
import org.nd4j.linalg.api.ops.impl.shape.ReductionShape;
import org.nd4j.linalg.api.shape.Shape;
import org.nd4j.linalg.dataset.api.iterator.MultiDataSetIterator;
import org.nd4j.linalg.exception.ND4JException;
import org.nd4j.linalg.factory.Nd4j;
@NoArgsConstructor(access = AccessLevel.PRIVATE)
public class SameDiffUtils {
/**
* Stack batch outputs, like an output from {@link org.nd4j.autodiff.samediff.SameDiff#output(MultiDataSetIterator, String...)}
*/
public static Map stackOutputs(List> outputs){
Map> outs = new HashMap<>();
for(Map batch : outputs){
for(String k : batch.keySet()){
if(!outs.containsKey(k))
outs.put(k, new ArrayList());
outs.get(k).add(batch.get(k));
}
}
Map ret = new HashMap<>();
for(String k : outs.keySet()){
try {
ret.put(k, Nd4j.concat(0, outs.get(k).toArray(new INDArray[0])));
} catch(Exception e){
throw new ND4JException("Error concatenating batch outputs", e);
}
}
return ret;
}
/**
* Get a list of batch outputs for a single variable from a list of batch outputs for all variables
*/
public static List getSingleOutput(List> outputs, String output){
List batches = new ArrayList<>();
for(Map batch : outputs)
batches.add(batch.get(output));
return batches;
}
public static ExternalErrorsFunction externalErrors(SameDiff sameDiff, Map externalGradients, SDVariable... inputs) {
Preconditions.checkArgument(inputs != null && inputs.length > 0, "Require at least one SDVariable to" +
" be specified when using external errors: got %s", inputs);
ExternalErrorsFunction fn = new ExternalErrorsFunction(sameDiff, Arrays.asList(inputs), externalGradients);
fn.outputVariable();
return fn;
}
public static ExternalErrorsFunction externalErrors(SameDiff sameDiff, SDVariable[] inputs) {
return externalErrors(sameDiff, null, inputs);
}
/**
* Add 1s as required to the array make an array possible to be broadcast with the original (pre-reduce) array.
*
* Example: if doing [a,b,c].sum(1), result is [a,c]. To 'undo' this in a way that can be auto-broadcast,
* we want to expand as required - i.e., [a,c] -> [a,1,c] which can be auto-broadcast with the original [a,b,c].
* This is typically only used with reduction operations backprop.
*
* @param origRank Rank of the original array, before the reduction was executed
* @param reduceDims Dimensions that the original array was reduced from
* @param toExpand Array to add 1s to the shape to (such that it can be
* @return Reshaped array.
*/
public static SDVariable reductionBroadcastableWithOrigShape(int origRank, int[] reduceDims, SDVariable toExpand) {
if (Shape.isWholeArray(origRank, reduceDims)) {
//Output is [1,1] which is already broadcastable
return toExpand;
} else if (origRank == 2 && reduceDims.length == 1) {
//In this case: [a,b] -> [1,b] or [a,b] -> [a,1]
//both are already broadcastable
return toExpand;
} else {
//Example: [a,b,c].sum(1) -> [a,c]... want [a,1,c]
for (int d : reduceDims) {
toExpand = toExpand.getSameDiff().expandDims(toExpand, d);
}
return toExpand;
}
}
public static SDVariable reductionBroadcastableWithOrigShape(SDVariable origInput, SDVariable axis, SDVariable toExpand) {
SDVariable shape = origInput.shape();
SDVariable reduceShape = reductionShape(shape, axis, true);
SDVariable reshaped = toExpand.reshape(reduceShape);
return reshaped;
}
public static SDVariable reductionShape(SDVariable shape, SDVariable axis, boolean keepDim){
return new ReductionShape(shape.getSameDiff(), shape, axis, keepDim).outputVariable();
}
public static void validateDifferentialFunctionSameDiff(SameDiff sameDiff, SDVariable function, DifferentialFunction op) {
Preconditions.checkState(function != null, "Passed in function was null.");
Preconditions.checkState(function.getSameDiff() == sameDiff);
Preconditions.checkState(function.getSameDiff() == sameDiff,
"Function applications must be contained " +
"in same sameDiff. The left %s must match this function %s", function, op);
}
}