org.nd4j.linalg.api.ops.impl.transforms.custom.MultiHeadDotProductAttention Maven / Gradle / Ivy

Show more of this group Show more artifacts with this name
Show all versions of nd4j-api Show documentation
There is a newer version: 1.0.0-M2.1
/*******************************************************************************
 * Copyright (c) 2015-2019 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.transforms.custom;

import lombok.NoArgsConstructor;
import org.nd4j.autodiff.samediff.SDVariable;
import org.nd4j.autodiff.samediff.SameDiff;
import org.nd4j.base.Preconditions;
import org.nd4j.linalg.api.buffer.DataType;
import org.nd4j.linalg.api.ops.DynamicCustomOp;

import java.util.Arrays;
import java.util.Collections;
import java.util.List;


/**
 * (optionally scaled) multi head dot product attention
 *
 * See also "Attention is all you need" (https://arxiv.org/abs/1706.03762, pp. 4,5, "3.2.2 Multi-Head Attention")
 *
 * @author Paul Dubs
 */
@NoArgsConstructor
public class MultiHeadDotProductAttention extends DynamicCustomOp {
    private boolean withWeights;
    private boolean scaled;

    public MultiHeadDotProductAttention(SameDiff sameDiff, SDVariable queries, SDVariable keys, SDVariable values,
                                                           SDVariable Wq, SDVariable Wk, SDVariable Wv, SDVariable Wo,
                                                           SDVariable mask,
                                        boolean scaled, boolean withWeights) {
        super(null, sameDiff,
                mask == null ? new SDVariable[] {queries, keys, values, Wq, Wk, Wv, Wo}
                : new SDVariable[] {queries, keys, values, Wq, Wk, Wv, Wo, mask},
                false);
        this.scaled = scaled;
        this.withWeights = withWeights;
        addIArgument(scaled ? 1 : 0);
        addIArgument(withWeights ? 1 : 0);
    }

    @Override
    public String opName() {
        return "multi_head_dot_product_attention";
    }

    @Override
    public List doDiff(List gradient) {
        return sameDiff.f().multiHeadDotProductAttentionBp(arg(0), arg(1), arg(2), arg(3), arg(4), arg(5), arg(6), gradient.get(0), args().length > 7 ? arg(7) : null, scaled);
    }

    @Override
    public List calculateOutputDataTypes(List dataTypes){
        Preconditions.checkState(dataTypes != null && (dataTypes.size() == 7 || dataTypes.size() == 8), "Expected 7 or 8 input datatypes, got %s", dataTypes);
        DataType first = dataTypes.get(0);
        for( int i=0; i 0){
                Preconditions.checkState(first == dataTypes.get(i), "All datatypes must be same type, got input datatypes %s", dataTypes);
            }
        }
        if(withWeights){
            return Arrays.asList(first, first);
        }else{
            return Collections.singletonList(first);
        }
    }

    @Override
    public int getNumOutputs() {
        if(withWeights){
            return 2;
        }else{
            return 1;
        }
    }
}