deeplearning4j/deeplearning4j

/*
 *  ******************************************************************************
 *  *
 *  *
 *  * 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.deeplearning4j.nn.layers.samediff;

import lombok.val;
import org.deeplearning4j.nn.api.Layer;
import org.deeplearning4j.nn.api.MaskState;
import org.deeplearning4j.nn.api.TrainingConfig;
import org.deeplearning4j.nn.conf.layers.samediff.SDVertexParams;
import org.deeplearning4j.nn.conf.layers.samediff.SameDiffVertex;
import org.deeplearning4j.nn.gradient.DefaultGradient;
import org.deeplearning4j.nn.gradient.Gradient;
import org.deeplearning4j.nn.graph.ComputationGraph;
import org.deeplearning4j.nn.graph.vertex.BaseGraphVertex;
import org.deeplearning4j.nn.params.SameDiffParamInitializer;
import org.deeplearning4j.nn.workspace.ArrayType;
import org.deeplearning4j.nn.workspace.LayerWorkspaceMgr;
import org.nd4j.autodiff.samediff.SDVariable;
import org.nd4j.autodiff.samediff.SameDiff;
import org.nd4j.autodiff.samediff.array.SingleThreadArrayHolder;
import org.nd4j.autodiff.samediff.internal.InferenceSession;
import org.nd4j.autodiff.samediff.internal.SessionMemMgr;
import org.nd4j.autodiff.util.SameDiffUtils;
import org.nd4j.common.base.Preconditions;
import org.nd4j.linalg.api.buffer.DataType;
import org.nd4j.linalg.api.memory.MemoryWorkspace;
import org.nd4j.linalg.api.memory.conf.WorkspaceConfiguration;
import org.nd4j.linalg.api.ndarray.INDArray;
import org.nd4j.linalg.api.ops.impl.layers.ExternalErrorsFunction;
import org.nd4j.linalg.factory.Nd4j;
import org.nd4j.common.primitives.Pair;

import java.util.*;

public class SameDiffGraphVertex extends BaseGraphVertex {

    protected SameDiffVertex config;
    protected SameDiff sameDiff;
    protected SDVariable outputVar;
    protected ExternalErrorsFunction fn;
    protected String outputKey;
    protected Map<String,SDVariable> inputVars;
    protected INDArray[] maskArrays;

    protected INDArray params;
    protected INDArray gradients;
    protected Map<String,INDArray> paramTable;
    protected Map<String,INDArray> gradTable;
    private MaskState currentMaskState;
    private int minibatchSize;

    public SameDiffGraphVertex(SameDiffVertex config, ComputationGraph graph, String name, int vertexIndex,
                                  INDArray paramsView, boolean initParams, DataType dataType) {
        super(graph, name, vertexIndex, null, null, dataType);
        this.config = config;
        SDVertexParams vp = config.getVertexParams();
        paramTable = SameDiffParamInitializer.getInstance().subsetAndReshape(vp.getParameterKeys(),
                vp.getParamShapes(), paramsView, null, config);
        if(initParams){
            config.initializeParameters(paramTable);
        }
        this.params = paramsView;
    }

    @Override
    public String toString() {
        return null;
    }

    @Override
    public boolean hasLayer() {
        return false;
    }

    @Override
    public Layer getLayer() {
        return null;
    }

    @Override
    public INDArray doForward(boolean training, LayerWorkspaceMgr workspaceMgr) {
        try(MemoryWorkspace ws = Nd4j.getWorkspaceManager().scopeOutOfWorkspaces()) {
            if (sameDiff == null) {
                doInit();
            }
        }

        Map<String,INDArray> phMap = new HashMap<>();
        config.validateInput(inputs);
        for(int i = 0; i < inputs.length; i++) {
            String name = config.getVertexParams().getInputs().get(i);
            final String maskName = name + "_mask";
            phMap.put(name, inputs[i]);
            if(maskArrays != null && maskArrays[i] != null) {
                phMap.put(maskName, maskArrays[i]);
            }else{
                phMap.put(maskName, createMask(dataType, inputs[i].shape()));
            }
        }


        //Configure memory management for SameDiff instance - use DL4J workspaces
        String wsNameWorking = workspaceMgr.getWorkspaceName(ArrayType.FF_WORKING_MEM);
        String wsNameOutput = workspaceMgr.getWorkspaceName(ArrayType.ACTIVATIONS);
        WorkspaceConfiguration confWorking = workspaceMgr.getConfiguration(ArrayType.FF_WORKING_MEM);
        WorkspaceConfiguration confOutput = workspaceMgr.getConfiguration(ArrayType.ACTIVATIONS);
        boolean actScopedOut = workspaceMgr.isScopedOut(ArrayType.ACTIVATIONS);
        Preconditions.checkState(actScopedOut || wsNameOutput != null, "Activations must have a workspace or must be scoped out");
        SessionMemMgr mmgr = new DL4JSameDiffMemoryMgr(wsNameWorking, wsNameOutput, confWorking, confOutput);

        InferenceSession is = sameDiff.getSessions().get(Thread.currentThread().getId());
        if(is == null) {
            is = SameDiff.getInferenceFactory().create(sameDiff);
            sameDiff.getSessions().put(Thread.currentThread().getId(), is);
        }

        is.setMmgr(mmgr);

        INDArray result = sameDiff.outputSingle(phMap, outputKey);

        //Edge case: "vertex" is just an identity activation, for example
        //TODO there may be a cleaner way to do this...
        if(!actScopedOut && !result.data().getParentWorkspace().getId().equals(wsNameOutput)){
            result = workspaceMgr.dup(ArrayType.ACTIVATIONS, result);
        } else if(actScopedOut && result.isAttached()) {
            result = result.detach();
        }

        //Clear placeholders and op inputs to ensure no out-of-scope arrays are still referenced anywhere
        sameDiff.clearPlaceholders(true);
        sameDiff.clearOpInputs();
        return workspaceMgr.dup(ArrayType.ACTIVATIONS, result);
    }

    @Override
    public Pair<Gradient, INDArray[]> doBackward(boolean tbptt, LayerWorkspaceMgr workspaceMgr) {
        Gradient g = new DefaultGradient();

        try(MemoryWorkspace ws = Nd4j.getWorkspaceManager().scopeOutOfWorkspaces()) {
            if (sameDiff == null) {
                doInit();
            }
        }

        List<String> inputNames = config.getVertexParams().getInputs();
        if(!sameDiff.hasGradientFunction()) {
            //Create when scoped out, to ensure any arrays are not in WS
            String[] inArr = inputNames.toArray(new String[inputNames.size()]);
            sameDiff.createGradFunction(inArr);
        }
        config.validateInput(inputs);

        //Configure memory management for SameDiff instance - use DL4J workspaces
        Map<Long,InferenceSession> sessionMap = sameDiff.getFunction("grad").getSessions();
        if(!sessionMap.containsKey(Thread.currentThread().getId())){
            sessionMap.put(Thread.currentThread().getId(), SameDiff.getInferenceFactory().create(sameDiff.getFunction("grad")));
        }
        String wsNameWorking = workspaceMgr.getWorkspaceName(ArrayType.BP_WORKING_MEM);
        String wsNameActGrad = workspaceMgr.getWorkspaceName(ArrayType.ACTIVATION_GRAD);
        WorkspaceConfiguration confWorking = workspaceMgr.getConfiguration(ArrayType.BP_WORKING_MEM);
        WorkspaceConfiguration confOutput = workspaceMgr.getConfiguration(ArrayType.ACTIVATION_GRAD);

        boolean actGradScopedOut = workspaceMgr.isScopedOut(ArrayType.ACTIVATION_GRAD);
        Preconditions.checkState(actGradScopedOut || wsNameActGrad != null, "Activation gradients must have a workspace or be scoped out");
        SessionMemMgr mmgr = new DL4JSameDiffMemoryMgr(wsNameWorking, wsNameActGrad, confWorking, confOutput);
        sessionMap.get(Thread.currentThread().getId()).setMmgr(mmgr);



        Map<String,INDArray> phMap = new HashMap<>();
        List<String> inputs = config.getVertexParams().getInputs();
        int i=0;
        for(String s : inputs){
            phMap.put(s, this.inputs[i++]);
        }
        for( int j=0; j<this.inputs.length; j++ ){
            String name = inputs.get(j);
            final String maskName = name + "_mask";
            if(maskArrays != null && maskArrays[j] != null) {
                phMap.put(maskName, maskArrays[j]);
            }else{
                phMap.put(maskName, createMask(dataType, this.inputs[j].shape()));
            }
        }
        String epsName = fn.getGradPlaceholderName();
        phMap.put(epsName, epsilon);

        List<String> required = new ArrayList<>(config.getVertexParams().getInputs());     //Ensure that the input placeholder gradients are calculated
        required.addAll(paramTable.keySet());

        Map<String,INDArray> gradsMap = sameDiff.calculateGradients(phMap, required);
        for(String s : paramTable.keySet() ){
            INDArray sdGrad = gradsMap.get(s);
            INDArray dl4jGrad = gradTable.get(s);
            dl4jGrad.assign(sdGrad);                                            //TODO OPTIMIZE THIS
            g.gradientForVariable().put(s, dl4jGrad);
        }

        INDArray[] dLdIns = new INDArray[inputs.size()];
        String fnName = fn.getGradPlaceholderName();
        for(int j=0; j<inputs.size(); j++ ){
            String name = inputs.get(j);
            dLdIns[j] = sameDiff.grad(name).getArr();

            String gradName = sameDiff.grad(inputNames.get(j)).name();
            if(dLdIns[j] == null && fnName.equals(gradName)){
                //Edge case with lambda vertices like identity: SameDiff doesn't store the placeholders
                // So, this getArr() can be trying to get placeholder from SameDiff instance, when it's available here
                dLdIns[j] = epsilon;
            }

            //Edge case: "vertex" is just an identity activation, for example
            //TODO there may be a cleaner way to do this...
            if(!actGradScopedOut && !dLdIns[j].data().getParentWorkspace().getId().equals(wsNameActGrad)){
                dLdIns[j] = workspaceMgr.dup(ArrayType.ACTIVATION_GRAD, dLdIns[j]);
            } else if(actGradScopedOut && dLdIns[j].isAttached()){
                dLdIns[j] = dLdIns[j].detach();
            }
        }

        //Clear placeholders and op inputs to ensure no out-of-scope arrays are still referenced anywhere
        sameDiff.clearPlaceholders(true);
        sameDiff.clearOpInputs();
        return new Pair<>(g, dLdIns);
    }

    @Override
    public void setBackpropGradientsViewArray(INDArray backpropGradientsViewArray) {
        SDVertexParams vp = config.getVertexParams();
        gradTable = SameDiffParamInitializer.getInstance().subsetAndReshape(vp.getParameterKeys(),
                vp.getParamShapes(), backpropGradientsViewArray, null, config);
    }

    @Override
    public Pair<INDArray, MaskState> feedForwardMaskArrays(INDArray[] maskArrays, MaskState currentMaskState, int minibatchSize) {
        this.maskArrays = maskArrays;
        this.currentMaskState = currentMaskState;

        return config.feedForwardMaskArrays(maskArrays, currentMaskState, minibatchSize);
    }


    protected void doInit(){
        try(MemoryWorkspace ws = Nd4j.getWorkspaceManager().scopeOutOfWorkspaces()) {
            sameDiff = SameDiff.create();
            //Use SingleThreadArrayHolder so we can use views (also don't need multithreading here, DL4J is not thread safe)
            sameDiff.setArrayHolders(new SingleThreadArrayHolder(), new SingleThreadArrayHolder(), false);

            inputVars = new LinkedHashMap<>();
            LinkedHashMap<String, SDVariable> maskVars = new LinkedHashMap<>();
            int i = 0;
            for(String s : config.getVertexParams().getInputs()) {
                val inputShape = inputs[i++].shape().clone();
                INDArray maskTemp = createMask(dataType, inputShape);
                inputShape[0] = -1;
                SDVariable inputVar = sameDiff.placeHolder(s, dataType, inputShape);
                inputVars.put(s, inputVar);
                long[] maskShape = maskTemp.shape().clone();
                maskShape[0] = -1;
                SDVariable maskVar = sameDiff.placeHolder(s + "_mask", maskTemp.dataType(), maskShape);
                maskVars.put(s, maskVar);
            }

            Map<String, long[]> paramShapes = config.getVertexParams().getParamShapes();
            Map<String, SDVariable> params = new LinkedHashMap<>();
            for (String s : paramShapes.keySet()) {
                val ps = paramShapes.get(s);
                SDVariable v = sameDiff.var(s, dataType, ps);
                params.put(s, v);
            }
            SDVariable layerOutput = config.defineVertex(sameDiff, inputVars, params, maskVars);
            Preconditions.checkNotNull(layerOutput, "Invalid output: layer output is null");
            outputVar = layerOutput;

            for (Map.Entry<String, INDArray> e : paramTable.entrySet()) {
                sameDiff.associateArrayWithVariable(e.getValue(), sameDiff.getVariable(e.getKey()));
            }

            //Define the function for external errors:
            fn = SameDiffUtils.externalErrors(sameDiff, null, layerOutput);
            fn.outputVariable();

            this.outputKey = outputVar.name();
        }
    }

    @Override
    public void clearVertex() {
        clear();
    }

    @Override
    public Map<String, INDArray> paramTable(boolean backpropOnly) {
        return paramTable;
    }

    @Override
    public TrainingConfig getConfig() {
        return config;
    }

    @Override
    public INDArray params() {
        return params;
    }

    @Override
    public INDArray getGradientsViewArray() {
        return gradients;
    }

    //Package private
    static INDArray createMask(DataType dataType, long[] shape){
        switch (shape.length){
            case 2: // FF-Type input
                return Nd4j.ones(dataType,shape[0], 1);
            case 3: // RNN-Type input
                return Nd4j.ones(dataType, shape[0], shape[2]);
            case 4: //CNN input
                return Nd4j.ones(dataType, shape[0], 1, 1, 1);
            default:
                Preconditions.throwEx("Can not create all-ones-mask for given input shape %s.", Arrays.toString(shape));
                return null;
        }
    }
}