org.nd4j.linalg.factory.Nd4j.gemm()方法的使用及代码示例

x33g5p2x  于2022-01-24 转载在 其他  
字(11.6k)|赞(0)|评价(0)|浏览(158)

本文整理了Java中org.nd4j.linalg.factory.Nd4j.gemm()方法的一些代码示例，展示了Nd4j.gemm()的具体用法。这些代码示例主要来源于Github/Stackoverflow/Maven等平台，是从一些精选项目中提取出来的代码，具有较强的参考意义，能在一定程度帮忙到你。Nd4j.gemm()方法的具体详情如下：
包路径：org.nd4j.linalg.factory.Nd4j
类名称：Nd4j
方法名：gemm

Nd4j.gemm介绍

[英]Matrix multiply: Implements c = alphaop(a)op(b) + betac where op(X) means transpose X (or not) depending on setting of arguments transposeA and transposeB.
Note that matrix c MUST be fortran order, have zero offset and have c.data().length == c.length(). An exception will be thrown otherwise.
Don't use this unless you know about level 3 blas and NDArray storage orders.
[中]矩阵乘法：实现c=alphaop（a）op（b）+betac，其中op（X）表示转置X（或不转置），具体取决于参数transpose和transposeB的设置。
注意，矩阵c必须是fortran顺序，偏移量为零，并且具有c.data（）。长度==c.长度（）。否则将引发异常。
除非你了解3级blas和NDArray存储订单，否则不要使用此功能。

代码示例

代码示例来源：origin: guoguibing/librec

  1. @Override
  2. public Pair<Gradient, INDArray> backpropGradient(INDArray epsilon) {
  3.   //If this layer is layer L, then epsilon is (w^(L+1)*(d^(L+1))^T) (or equivalent)
  4.   INDArray z = this.preOutput(true); //Note: using preOutput(INDArray) can't be used as this does a setInput(input) and resets the 'appliedDropout' flag
  5.   //INDArray activationDerivative = Nd4j.getExecutioner().execAndReturn(Nd4j.getOpFactory().createTransform(conf().getLayer().getActivationFunction(), z).derivative());
  6.   //        INDArray activationDerivative = conf().getLayer().getActivationFn().getGradient(z);
  7.   //        INDArray delta = epsilon.muli(activationDerivative);
  8.   INDArray delta = layerConf().getActivationFn().backprop(z, epsilon).getFirst(); //TODO handle activation function params
  9.   if (maskArray != null) {
  10.     applyMask(delta);
  11.   }
  12.   Gradient ret = new DefaultGradient();
  13.   INDArray weightGrad = gradientViews.get(CDAEParamInitializer.WEIGHT_KEY); //f order
  14.   Nd4j.gemm(input, delta, weightGrad, true, false, 1.0, 0.0);
  15.   INDArray userWeightGrad = gradientViews.get(CDAEParamInitializer.USER_WEIGHT_KEY); //f order
  16.   userWeightGrad.assign(delta);
  17.   INDArray biasGrad = gradientViews.get(CDAEParamInitializer.BIAS_KEY);
  18.   biasGrad.assign(delta.sum(0)); //biasGrad is initialized/zeroed first
  19.   ret.gradientForVariable().put(CDAEParamInitializer.WEIGHT_KEY, weightGrad);
  20.   ret.gradientForVariable().put(CDAEParamInitializer.BIAS_KEY, biasGrad);
  21.   ret.gradientForVariable().put(CDAEParamInitializer.USER_WEIGHT_KEY, userWeightGrad);
  22.   INDArray epsilonNext = params.get(CDAEParamInitializer.WEIGHT_KEY).mmul(delta.transpose()).transpose();
  23.   //epsilonNext = null;
  24.   return new Pair<>(ret, epsilonNext);
  25. }

代码示例来源：origin: deeplearning4j/nd4j

  1. /**
  2.  *
  3.  * matrix multiply: implements op(a)*op(b)
  4.  *
  5.  * where op(x) means transpose x (or not) depending on
  6.  * setting of arguments transposea and transposeb.<br>
  7.  * so gemm(a,b,false,false) == a.mmul(b), gemm(a,b,true,false) == a.transpose().mmul(b) etc.
  8.  * @param a first matrix
  9.  * @param b second matrix
  10.  * @param transposeA if true: transpose matrix a before mmul
  11.  * @param transposeB if true: transpose matrix b before mmul
  12.  * @return result
  13.  */
  14. public static INDArray gemm(INDArray a,
  15.               INDArray b,
  16.               boolean transposeA,
  17.               boolean transposeB) {
  18.   long cRows = (transposeA ? a.columns() : a.rows());
  19.   long cCols = (transposeB ? b.rows() : b.columns());
  20.   INDArray c = Nd4j.createUninitialized(new long[] {cRows, cCols}, 'f');
  21.   return gemm(a, b, c, transposeA, transposeB, 1.0, 0.0);
  22. }

代码示例来源：origin: deeplearning4j/nd4j

  1. cCopy2.assign(c);
  2. INDArray out = Nd4j.gemm(a, b, c, transposeA, transposeB, alpha, beta);
  3. if (out != c) {
  4.   System.out.println("Returned different array than c");
  5.   INDArray aCopy = Shape.toOffsetZeroCopy(a);
  6.   INDArray bCopy = Shape.toOffsetZeroCopy(b);
  7.   INDArray onCopies = Nd4j.gemm(aCopy, bCopy, cCopy1, transposeA, transposeB, alpha, beta);
  8.   printGemmFailureDetails(a, b, cCopy2, transposeA, transposeB, alpha, beta, rmExpected, out, onCopies);

代码示例来源：origin: org.deeplearning4j/deeplearning4j-nn

  1. /** Returns tuple: {Gradient,Delta,Output} given preOut */
  2. private Pair<Gradient, INDArray> getGradientsAndDelta(INDArray preOut) {
  3.   ILossFunction lossFunction = layerConf().getLossFn();
  4.   INDArray labels2d = getLabels2d();
  5.   //INDArray delta = lossFunction.computeGradient(labels2d, preOut, layerConf().getActivationFunction(), maskArray);
  6.   INDArray delta = lossFunction.computeGradient(labels2d, preOut, layerConf().getActivationFn(), maskArray);
  7.   Gradient gradient = new DefaultGradient();
  8.   INDArray weightGradView = gradientViews.get(DefaultParamInitializer.WEIGHT_KEY);
  9.   INDArray biasGradView = gradientViews.get(DefaultParamInitializer.BIAS_KEY);
  10.   Nd4j.gemm(input, delta, weightGradView, true, false, 1.0, 0.0); //Equivalent to:  weightGradView.assign(input.transpose().mmul(delta));
  11.   delta.sum(biasGradView, 0); //biasGradView is initialized/zeroed first in sum op
  12.   gradient.gradientForVariable().put(DefaultParamInitializer.WEIGHT_KEY, weightGradView);
  13.   gradient.gradientForVariable().put(DefaultParamInitializer.BIAS_KEY, biasGradView);
  14.   return new Pair<>(gradient, delta);
  15. }

代码示例来源：origin: org.deeplearning4j/deeplearning4j-nn

  1. @Override
  2. public Pair<Gradient, INDArray> backpropGradient(INDArray epsilon) {
  3.   //If this layer is layer L, then epsilon is (w^(L+1)*(d^(L+1))^T) (or equivalent)
  4.   INDArray z = preOutput(input, true);
  5.   INDArray activationDerivative = propUpDerivative(z);
  6.   INDArray delta = epsilon.muli(activationDerivative);
  7.   if (maskArray != null) {
  8.     delta.muliColumnVector(maskArray);
  9.   }
  10.   Gradient ret = new DefaultGradient();
  11.   INDArray weightGrad = gradientViews.get(DefaultParamInitializer.WEIGHT_KEY); //f order
  12.   Nd4j.gemm(input, delta, weightGrad, true, false, 1.0, 0.0);
  13.   INDArray biasGrad = gradientViews.get(DefaultParamInitializer.BIAS_KEY);
  14.   delta.sum(biasGrad, 0); //biasGrad is initialized/zeroed first in sum op
  15.   INDArray vBiasGradient = gradientViews.get(PretrainParamInitializer.VISIBLE_BIAS_KEY);
  16.   ret.gradientForVariable().put(DefaultParamInitializer.WEIGHT_KEY, weightGrad);
  17.   ret.gradientForVariable().put(DefaultParamInitializer.BIAS_KEY, biasGrad);
  18.   ret.gradientForVariable().put(PretrainParamInitializer.VISIBLE_BIAS_KEY, vBiasGradient);
  19.   INDArray epsilonNext = params.get(DefaultParamInitializer.WEIGHT_KEY).mmul(delta.transpose()).transpose();
  20.   return new Pair<>(ret, epsilonNext);
  21. }

代码示例来源：origin: org.nd4j/nd4j-api

  1. /**
  2.  *
  3.  * matrix multiply: implements op(a)*op(b)
  4.  *
  5.  * where op(x) means transpose x (or not) depending on
  6.  * setting of arguments transposea and transposeb.<br>
  7.  * so gemm(a,b,false,false) == a.mmul(b), gemm(a,b,true,false) == a.transpose().mmul(b) etc.
  8.  * @param a first matrix
  9.  * @param b second matrix
  10.  * @param transposeA if true: transpose matrix a before mmul
  11.  * @param transposeB if true: transpose matrix b before mmul
  12.  * @return result
  13.  */
  14. public static INDArray gemm(INDArray a, INDArray b, boolean transposeA, boolean transposeB) {
  15.   int cRows = (transposeA ? a.columns() : a.rows());
  16.   int cCols = (transposeB ? b.rows() : b.columns());
  17.   INDArray c = Nd4j.createUninitialized(new int[] {cRows, cCols}, 'f');
  18.   return gemm(a, b, c, transposeA, transposeB, 1.0, 0.0);
  19. }

代码示例来源：origin: org.deeplearning4j/deeplearning4j-nn

  1. @Override
  2. public Pair<Gradient, INDArray> backpropGradient(INDArray epsilon) {
  3.   //If this layer is layer L, then epsilon is (w^(L+1)*(d^(L+1))^T) (or equivalent)
  4.   INDArray z = preOutput(true); //Note: using preOutput(INDArray) can't be used as this does a setInput(input) and resets the 'appliedDropout' flag
  5.   //INDArray activationDerivative = Nd4j.getExecutioner().execAndReturn(Nd4j.getOpFactory().createTransform(conf().getLayer().getActivationFunction(), z).derivative());
  6.   //        INDArray activationDerivative = conf().getLayer().getActivationFn().getGradient(z);
  7.   //        INDArray delta = epsilon.muli(activationDerivative);
  8.   INDArray delta = layerConf().getActivationFn().backprop(z, epsilon).getFirst(); //TODO handle activation function params
  9.   if (maskArray != null) {
  10.     applyMask(delta);
  11.   }
  12.   Gradient ret = new DefaultGradient();
  13.   INDArray weightGrad = gradientViews.get(DefaultParamInitializer.WEIGHT_KEY); //f order
  14.   Nd4j.gemm(input, delta, weightGrad, true, false, 1.0, 0.0);
  15.   INDArray biasGrad = gradientViews.get(DefaultParamInitializer.BIAS_KEY);
  16.   delta.sum(biasGrad, 0); //biasGrad is initialized/zeroed first
  17.   ret.gradientForVariable().put(DefaultParamInitializer.WEIGHT_KEY, weightGrad);
  18.   ret.gradientForVariable().put(DefaultParamInitializer.BIAS_KEY, biasGrad);
  19.   INDArray epsilonNext = params.get(DefaultParamInitializer.WEIGHT_KEY).mmul(delta.transpose()).transpose();
  20.   return new Pair<>(ret, epsilonNext);
  21. }

代码示例来源：origin: org.deeplearning4j/deeplearning4j-nn

  1. Nd4j.gemm(lastEncoderActivation, currentDelta, dLdMeanW, true, false, 1.0, 0.0);
  2. INDArray dLdMeanB = gradientViews.get(VariationalAutoencoderParamInitializer.PZX_MEAN_B);
  3. currentDelta.sum(dLdMeanB, 0); //dLdMeanB is initialized/zeroed first in sum op
  4.     actInput = fwd.encoderActivations[i - 1];
  5.   Nd4j.gemm(actInput, currentDelta, dLdW, true, false, 1.0, 0.0);
  6.   currentDelta.sum(dLdB, 0); //dLdB is initialized/zeroed first in sum op

代码示例来源：origin: org.deeplearning4j/deeplearning4j-nn

  1. INDArray dLdxzb = gradientViews.get(VariationalAutoencoderParamInitializer.PXZ_B);
  2. INDArray lastDecActivations = decoderActivations[decoderActivations.length - 1];
  3. Nd4j.gemm(lastDecActivations, dpdpxz, dLdxzw, true, false, scaleFactor, gemmCConstant);
  4. if (l == 0) {
  5.   dpdpxz.sum(dLdxzb, 0); //dLdxzb array is initialized/zeroed first in sum op
  6.   Nd4j.gemm(actInput, currentDelta, dLdW, true, false, scaleFactor, gemmCConstant);
  7. INDArray dLdZXMeanW = gradientViews.get(VariationalAutoencoderParamInitializer.PZX_MEAN_W);
  8. INDArray dLdZXLogStdev2W = gradientViews.get(VariationalAutoencoderParamInitializer.PZX_LOGSTD2_W);
  9. Nd4j.gemm(lastEncoderActivation, dLdPreMu, dLdZXMeanW, true, false, scaleFactor, gemmCConstant);
  10. Nd4j.gemm(lastEncoderActivation, dLdPreLogSigma2, dLdZXLogStdev2W, true, false, scaleFactor, gemmCConstant);
  11. epsilon = Nd4j.gemm(dLdPreMu, eZXMeanW, false, true); //Equivalent to: epsilon = eZXMeanW.mmul(dLdPreMu.transpose()).transpose(); using   (AxB^T)^T = BxA^T
  12. Nd4j.gemm(dLdPreLogSigma2, eZXLogStdev2W, epsilon, false, true, 1.0, 1.0);
  13.     actInput = fwd.encoderActivations[i - 1];
  14.   Nd4j.gemm(actInput, currentDelta, dLdW, true, false, scaleFactor, gemmCConstant);
  15.   if (l == 0) {
  16.     currentDelta.sum(dLdB, 0);

代码示例来源：origin: org.nd4j/nd4j-api

  1. cCopy2.assign(c);
  2. INDArray out = Nd4j.gemm(a, b, c, transposeA, transposeB, alpha, beta);
  3. if (out != c) {
  4.   System.out.println("Returned different array than c");
  5.   INDArray aCopy = Shape.toOffsetZeroCopy(a);
  6.   INDArray bCopy = Shape.toOffsetZeroCopy(b);
  7.   INDArray onCopies = Nd4j.gemm(aCopy, bCopy, cCopy1, transposeA, transposeB, alpha, beta);
  8.   printGemmFailureDetails(a, b, cCopy2, transposeA, transposeB, alpha, beta, rmExpected, out, onCopies);

代码示例来源：origin: org.deeplearning4j/deeplearning4j-nn

  1. Nd4j.gemm(input, delta, weightGradView, true, false, 1.0, 0.0); //Equivalent to:  weightGradView.assign(input.transpose().mmul(delta));
  2. delta.sum(biasGradView, 0); //biasGradView is initialized/zeroed first in sum op

代码示例来源：origin: org.deeplearning4j/deeplearning4j-nn

  1. if (iTimeIndex != timeSeriesLength - 1) {
  2.   Nd4j.gemm(deltaifogNext, wIFOG, nablaOut, false, true, 1.0, 1.0);
  3.   Nd4j.gemm(prevLayerActivationSlice, deltaifogNext, iwGradientsOut, true, false, 1.0, 1.0);
  4. } else {
  5.   INDArray iwGradients_i =
  6.           iwGradientsOut.get(NDArrayIndex.all(), NDArrayIndex.interval(0, hiddenLayerSize));
  7.   Nd4j.gemm(prevLayerActivationSlice, deltai, iwGradients_i, true, false, 1.0, 1.0);
  8.   INDArray iwGradients_og = iwGradientsOut.get(NDArrayIndex.all(),
  9.           NDArrayIndex.interval(2 * hiddenLayerSize, 4 * hiddenLayerSize));
  10.   INDArray deltaog = deltaifogNext.get(NDArrayIndex.all(),
  11.           NDArrayIndex.interval(2 * hiddenLayerSize, 4 * hiddenLayerSize));
  12.   Nd4j.gemm(prevLayerActivationSlice, deltaog, iwGradients_og, true, false, 1.0, 1.0);
  13.   Nd4j.gemm(prevHiddenUnitActivation, deltaifogNext, rwGradientsIFOG, true, false, 1.0, 1.0);
  14. if (iTimeIndex > 0 || prevHiddenUnitActivation != null) {
  15.   Nd4j.gemm(deltaifogNext, inputWeights, epsilonNextSlice, false, true, 1.0, 1.0);
  16. } else {
  17.   Nd4j.gemm(deltai, wi, epsilonNextSlice, false, true, 1.0, 1.0);
  18.   INDArray deltaog = deltaifogNext.get(NDArrayIndex.all(),
  19.           NDArrayIndex.interval(2 * hiddenLayerSize, 4 * hiddenLayerSize));
  20.   INDArray wog = inputWeights.get(NDArrayIndex.all(),
  21.           NDArrayIndex.interval(2 * hiddenLayerSize, 4 * hiddenLayerSize));
  22.   Nd4j.gemm(deltaog, wog, epsilonNextSlice, false, true, 1.0, 1.0); //epsilonNextSlice.addi(deltao.mmul(woTranspose)).addi(deltag.mmul(wgTranspose));

代码示例来源：origin: org.deeplearning4j/deeplearning4j-nn

  1. Nd4j.gemm(im2col2d, delta2d, weightGradView2df, true, true, 1.0, 0.0);

代码示例来源：origin: org.deeplearning4j/deeplearning4j-nn

  1. .notifyScopeLeft();
  2. Nd4j.gemm(prevOutputActivations, recurrentWeightsIFOG, ifogActivations, false, false, 1.0, 1.0);
  3. ifogActivations.addiRowVector(biases);

内容来源于网络，如有侵权，请联系作者删除！

Nd4j

相关文章

    查看更多

    热门标签

    更多
    JavaquerypythonNode开发语言requestUtil数据库Table后端算法LoggerMessageElementParser

    最新文章

    更多

    Nd4j类方法

    • 蜀ICP备13028337号-1 大数据知识库 https://www.saoniuhuo.com © All rights reserved
    • 本站内容来源互联网,如果侵犯您的权益请联系我们删除, 联系方式：448109455@qq.com