cavis/libnd4j/include/ops/declarable/generic/convo/pooling/maxpool3d.cpp

231 lines
14 KiB
C++
Raw Normal View History

2019-06-06 14:21:15 +02:00
/*******************************************************************************
* 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
******************************************************************************/
//
// @author Yurii Shyrma (iuriish@yahoo.com), created on 19.02.2018
//
#include <op_boilerplate.h>
#if NOT_EXCLUDED(OP_maxpool3dnew)
#include <ops/declarable/CustomOperations.h>
#include <ops/declarable/helpers/convolutions.h>
namespace nd4j {
namespace ops {
//////////////////////////////////////////////////////////////////////////
CUSTOM_OP_IMPL(maxpool3dnew, 1, 1, false, 0, 14) {
auto input = INPUT_VARIABLE(0); // [bS, iD, iH, iW, iC] (NDHWC) or [bS, iC, iD, iH, iW] (NCDHW)
auto output = OUTPUT_VARIABLE(0); // [bS, oD, oH, oW, iC] (NDHWC) or [bS, iC, oD, oH, oW] (NCDHW)
int kD = INT_ARG(0); // filter(kernel) depth
int kH = INT_ARG(1); // filter(kernel) height
int kW = INT_ARG(2); // filter(kernel) width
int sD = INT_ARG(3); // strides depth
int sH = INT_ARG(4); // strides height
int sW = INT_ARG(5); // strides width
int pD = INT_ARG(6); // paddings depth
int pH = INT_ARG(7); // paddings height
int pW = INT_ARG(8); // paddings width
int dD = INT_ARG(9); // dilations depth
int dH = INT_ARG(10); // dilations height
int dW = INT_ARG(11); // dilations width
int isSameMode = INT_ARG(12); // 1-SAME, 0-VALID
// int extraParam0 = INT_ARG(13); // unnecessary for max case, required only for avg and pnorm cases
int isNCDHW = block.getIArguments()->size() > 14 ? !INT_ARG(14) : 1; // 1-NDHWC, 0-NCDHW
REQUIRE_TRUE(input->rankOf() == 5, 0, "MAXPOOL3DNEW OP: rank of input array must be equal to 5, but got %i instead !", input->rankOf());
REQUIRE_TRUE(dD != 0 && dH != 0 && dW != 0, 0, "MAXPOOL3DNEW op: dilation must not be zero, but got instead {%i, %i, %i}", dD, dH, dW);
int bS, iC, iD, iH, iW, oC, oD, oH, oW; // batch size, input channels, input depth/height/width, output channels, output depth/height/width;
int indIOioC, indIOioD, indWoC, indWiC, indWkD; // corresponding indexes
ConvolutionUtils::getSizesAndIndexesConv3d(isNCDHW, *input, *output, bS, iC, iD, iH, iW, oC, oD, oH, oW, indIOioC, indIOioD, indWiC, indWoC, indWkD);
std::string expectedOutputShape = ShapeUtils::shapeAsString(ShapeUtils::composeShapeUsingDimsAndIdx({bS,iC,oD,oH,oW, 0,indIOioC,indIOioD,indIOioD+1,indIOioD+2}));
REQUIRE_TRUE(expectedOutputShape == ShapeUtils::shapeAsString(output), 0, "MAXPOOL3D op: wrong shape of output array, expected is %s, but got %s instead !", expectedOutputShape.c_str(), ShapeUtils::shapeAsString(output).c_str());
// REQUIRE_TRUE(iD >= kD && iH >= kH && iW >= kW, 0, "MAXPOOL3D OP: the input depth/height/width must be greater or equal to kernel(filter) depth/height/width, but got [%i, %i, %i] and [%i, %i, %i] correspondingly !", iD,iH,iW, kD,kH,kW);
// REQUIRE_TRUE(kD/2 >= pD && kH/2 >= pH && kW/2 >= pW, 0, "MAXPOOL3D OP: pad depth/height/width must not be greater than half of kernel depth/height/width, but got [%i, %i, %i] and [%i, %i, %i] correspondingly !", pD,pH,pW, kD,kH,kW);
if(!isNCDHW) {
input = input->permute({0, 4, 1, 2, 3}); // [bS, iD, iH, iW, iC] -> [bS, iC, iD, iH, iW]
output = output->permute({0, 4, 1, 2, 3}); // [bS, oD, oH, oW, iC] -> [bS, iC, oD, oH, oW]
}
if(isSameMode) // SAME
ConvolutionUtils::calcPadding3D(pD, pH, pW, oD, oH, oW, iD, iH, iW, kD, kH, kW, sD, sH, sW, dD, dH, dW);
Dev branch merge: dev_20190606 (#7904) * correct logsoftmax looss (#2) * Small SameDiff listener fix (#4) * Various fixes (#6) * #7839 Fix for asXMatrix and tests * #7866 EmbeddingSequenceLayer dtype fix + test * #7856 SameDiff save/load stream methods * #7859 RegressionEvaluation rank 4 fix + tests + axis configuration * EvaluationBinary 3d/4d * More evaluation 3d/4d tests * #7847 Evaluation empty checks * Small test ifx * #7848 Fix median edge case * Improve DL4J samediff layer tests * [WIP] FastText wrapper implemented (#8) * FastText implemented * Some fixes * Fix shapes for wordsNearest * Validation of input vectors * Fixes * Fixed test * Thread tagged * Some tweaks * setContextClassLoader for DeallocatorServiceThread * Numpy format tests (#1) * Various fixes (#11) * #7852 SameDiff gather fix * #7892 SameDiff placeholder to constant conversion * #7890 validate input rank for MLN/CG init methods * Fix broken permute shape calculation * Permute and gather fixes * Tests * #7850 LogSumExp fix + test * Handful of test fixes * Empty arrays with non-scalar shapes (#10) * minor rearrangements for lambdas * empty tensors with non-scalar shapes * numpy empty tensors with non-scalar shapes * few more empty tweaks * Small fixes * conv3d signature update * micro fix in batchnorm mkldnn * Import fixes * Fix * MKL-DNN update * Small fill fix * fill with empty input + test * Fixes * Small error improvement * Fix * one special test * couple of fixes for lstm * Rewrite TFGraphMapper.getNDArrayFromTensor to be maintainable and less error prone * Fixes * FP16 * Unsigned * BFloat16 * Fill op - empty tweaks * - couple of fixes for empty arrays construction - stack updated * strided slice fix * one transform test * provide method for reducing shapeInfo in case of input array is empty * Fixed reduceAlongDimensions to use empty input properly. * couple of broadcast tests * couple of tests broadcast tests + tweak to make them pass * add check of non-empty to methods producing sub-arrays * Fixed reshapeC with zeros in shape. * complete empty check in reduce_... legacy ops * Concat and cumsum/prod * Tweak to empty shape inference on import * add empty check to the rest of reduce legacy ops * one more test * correct typo in evalReduceShapeInfoEmpty * Added tests for reduce_* ops to tests with zero shapes. * few more tests for empty reductions * Fixed strided_slice op with empty case and tests. * one more empty reduction test * Fixed strided_slice test. * add empty check to NDArray::reshapei * infOrMax * empty min/max with infinity tests * made unstack working correctly with empty arrays * few IndexReduce tests + tweaks for empty shapes * add test for empty concat * few tests fixed * Validation fix for reductions on empty shapes * Reverse fix * Reduction shape calc fixes * SameDiff.generateOutputVariable: don't use shape function to determine number of outputs * Range fix * - NDArray constructor updated for scalars/empty arrays - few tests fixed * More fixes * Empty creator fixes * concat fix * concat fix * TF import tests: allow 'both all NaN' and 'both all inf' to pass * Slice, zero fraction, and reshape fixes * transpose, gather * Zero fraction * scalar cast fix * Empty reduction axis support * few more tests fixed * Fixed input checks conforming with TF for concat op and tests. * few tests fixed * matmul scalar shape fix * Fixed checkout for data type and scalarity with concat to allow non-empty scalars with vector concats. * broadcast bool fix * few more tests * few more tests * correct evalReduceShapeInfoEmpty * argmax/argmin + tests * one more empty edge case + one more test * argmax/argmin/realdiv_bp tweaks * empty reshape test + fix * Helper fixes * Small fixes * Gather test fix * Gather test fix * Small fixes * reduce scalar zero values * scalar mean workaround * Remove debug code * along dim mean workaround * one more test * - equalsTo() tweak for empty arrays - one more test * broadcast tweaks
2019-06-15 13:34:34 +02:00
ConvolutionUtils::pooling3d(block, *input, *output, kD, kH, kW, sD, sH, sW, pD, pH, pW, dD, dH, dW, 0, 1);
2019-06-06 14:21:15 +02:00
if(!isNCDHW) {
delete input;
delete output;
}
return Status::OK();
}
DECLARE_TYPES(maxpool3dnew) {
getOpDescriptor()
->setAllowedInputTypes(nd4j::DataType::ANY)
->setSameMode(true);
}
DECLARE_SHAPE_FN(maxpool3dnew) {
int kD = INT_ARG(0); // filter(kernel) depth
int kH = INT_ARG(1); // filter(kernel) height
int kW = INT_ARG(2); // filter(kernel) width
int sD = INT_ARG(3); // strides depth
int sH = INT_ARG(4); // strides height
int sW = INT_ARG(5); // strides width
int pD = INT_ARG(6); // paddings depth
int pH = INT_ARG(7); // paddings height
int pW = INT_ARG(8); // paddings width
int dD = INT_ARG(9); // dilations depth
int dH = INT_ARG(10); // dilations height
int dW = INT_ARG(11); // dilations width
int isSameMode = INT_ARG(12); // 1-SAME, 0-VALID
// int extraParam0 = INT_ARG(13);
int isNCDHW = block.getIArguments()->size() > 14 ? !INT_ARG(14) : 1; // 1-NDHWC, 0-NCDHW
REQUIRE_TRUE(dD != 0 && dH != 0 && dW != 0, 0, "MAXPOOL3DNEW op: dilation must not be zero, but got instead {%i, %i, %i}", dD, dH, dW);
Nd4jLong* inputShapeInfo = inputShape->at(0);
int idxID, idxIC;
if(isNCDHW) { idxID = 2; idxIC = 1;}
else { idxID = 1; idxIC = 4;}
int bS = inputShapeInfo[1]; // batch size
int iC = inputShapeInfo[idxIC+1]; // input channels
int iD = inputShapeInfo[idxID+1]; // input depth
int iH = inputShapeInfo[idxID+2]; // input height
int iW = inputShapeInfo[idxID+3]; // input width
int oD, oH, oW; // output depth, height, width
ConvolutionUtils::calcOutSizePool3D(oD, oH, oW, kD, kH, kW, sD, sH, sW, pD, pH, pW, dD, dH, dW, iD, iH, iW, isSameMode);
Nd4jLong outputShape[5];
outputShape[0] = bS;
if (isNCDHW) {
outputShape[1] = iC;
outputShape[2] = oD;
outputShape[3] = oH;
outputShape[4] = oW;
} else {
outputShape[1] = oD;
outputShape[2] = oH;
outputShape[3] = oW;
outputShape[4] = iC;
}
return SHAPELIST(ConstantShapeHelper::getInstance()->createShapeInfo(ShapeDescriptor(ArrayOptions::dataType(inputShapeInfo), shape::order(inputShapeInfo), outputShape, 5)));
}
DECLARE_TYPES(maxpool3dnew_bp) {
getOpDescriptor()
->setAllowedInputTypes(nd4j::DataType::ANY)
->setAllowedOutputTypes({ALL_FLOATS});
}
//////////////////////////////////////////////////////////////////////////
CUSTOM_OP_IMPL(maxpool3dnew_bp, 2, 1, false, 0, 14) {
auto input = INPUT_VARIABLE(0); // [bS, iD, iH, iW, iC] (NDHWC) or [bS, iC, iD, iH, iW] (NCDHW)
auto gradO = INPUT_VARIABLE(1); // [bS, oD, oH, oW, oC] (NDHWC) or [bS, oC, oD, oH, oW] (NCDHW), epsilon_next
auto gradI = OUTPUT_VARIABLE(0); // [bS, iD, iH, iW, iC] (NDHWC) or [bS, iC, iD, iH, iW] (NCDHW), epsilon
const int kD = INT_ARG(0); // filter(kernel) depth
const int kH = INT_ARG(1); // filter(kernel) height
const int kW = INT_ARG(2); // filter(kernel) width
const int sD = INT_ARG(3); // strides depth
const int sH = INT_ARG(4); // strides height
const int sW = INT_ARG(5); // strides width
int pD = INT_ARG(6); // paddings depth
int pH = INT_ARG(7); // paddings height
int pW = INT_ARG(8); // paddings width
const int dD = INT_ARG(9); // dilations depth
const int dH = INT_ARG(10); // dilations height
const int dW = INT_ARG(11); // dilations width
const int isSameMode = INT_ARG(12); // 1-SAME, 0-VALID
// int extraParam0 = INT_ARG(13); // unnecessary for max case, required only for avg and pnorm cases
int isNCDHW = block.getIArguments()->size() > 14 ? !INT_ARG(14) : 1; // 1-NDHWC, 0-NCDHW
REQUIRE_TRUE(input->rankOf() == 5, 0, "MAXPOOL3D_BP op: input should have rank of 5, but got %i instead", input->rankOf());
REQUIRE_TRUE(dD != 0 && dH != 0 && dW != 0, 0, "MAXPOOL3DNEW op: dilation must not be zero, but got instead {%i, %i, %i}", dD, dH, dW);
int bS, iC, iD, iH, iW, oC, oD, oH, oW; // batch size, input channels, input depth/height/width, output channels, output depth/height/width;
int indIOioC, indIOioD, indWoC, indWiC, indWkD; // corresponding indexes
ConvolutionUtils::getSizesAndIndexesConv3d(isNCDHW, *input, *gradO, bS, iC, iD, iH, iW, oC, oD, oH, oW, indIOioC, indIOioD, indWiC, indWoC, indWkD);
std::string expectedGradOShape = ShapeUtils::shapeAsString(ShapeUtils::composeShapeUsingDimsAndIdx({bS,iC,oD,oH,oW, 0,indIOioC,indIOioD,indIOioD+1,indIOioD+2}));
std::string expectedGradIShape = ShapeUtils::shapeAsString(ShapeUtils::composeShapeUsingDimsAndIdx({bS,iC,iD,iH,iW, 0,indIOioC,indIOioD,indIOioD+1,indIOioD+2}));
REQUIRE_TRUE(expectedGradOShape == ShapeUtils::shapeAsString(gradO), 0, "MAXPOOL3D_BP op: wrong shape of output's gradients array (next epsilon), expected is %s, but got %s instead !", expectedGradOShape.c_str(), ShapeUtils::shapeAsString(gradO).c_str());
REQUIRE_TRUE(expectedGradIShape == ShapeUtils::shapeAsString(gradI), 0, "MAXPOOL3D_BP op: wrong shape of input's gradients array (epsilon), expected is %s, but got %s instead !", expectedGradIShape.c_str(), ShapeUtils::shapeAsString(gradI).c_str());
if(!isNCDHW) {
input = input->permute({0, 4, 1, 2, 3}); // [bS, iD, iH, iW, iC] -> [bS, iC, iD, iH, iW]
gradI = gradI->permute({0, 4, 1, 2, 3}); // [bS, iD, iH, iW, iC] -> [bS, iC, iD, iH, iW]
gradO = gradO->permute({0, 4, 1, 2, 3}); // [bS, oD, oH, oW, iC] -> [bS, iC, oD, oH, oW]
}
if(isSameMode) // SAME
ConvolutionUtils::calcPadding3D(pD, pH, pW, oD, oH, oW, iD, iH, iW, kD, kH, kW, sD, sH, sW, dD, dH, dW);
// NDArray<T> columnsWrongShape(input->ordering(), {bS, iC, oD, oH, oW, kD, kH, kW}, input->getWorkspace());
// NDArray<T>* columns = columnsWrongShape.permute({0, 1, 5, 6, 7, 2, 3, 4}); // [bS, iC, oD, oH, oW, kD, kH, kW] -> [bS, iC, kD, kH, kW, oD, oH, oW]
// ConvolutionUtils<T>::vol2col(*input, *columns, sD, sH, sW, pD, pH, pW, dD, dH, dW); // [bS, iC, iD, iH, iW] is convoluted to [bS, iC, kD, kH, kW, oD, oH, oW]
// NDArray<T>* columns2d = columnsWrongShape.reshape('c', {bS*iC*oD*oH*oW, kD*kH*kW});
// NDArray<T>* gradOVector = gradO->reshape('c', {(int) gradO->lengthOf(), 1});
// T extraParams[] = {(T)1., (T)1.};
// columns2d->template applyTransform<simdOps::IsMax<T>>(extraParams);
// columns2d->muliColumnVector(gradOVector);
// ConvolutionUtils<T>::col2vol(*columns, *gradI, sD, sH, sW, pD, pH, pW, dD, dH, dW); // columns [bS, iC, kD, kH, kW, oD, oH, oW] is de-convoluted to [bS, iC, iD, iH, iW]
// 0,1 - kernel Height/Width; 2,3 - stride Height/Width; 4,5 - pad Height/Width; 6,7 - dilation Height/Width; 8 - poolingMode; 9 - unnecessary;
Dev branch merge: dev_20190606 (#7904) * correct logsoftmax looss (#2) * Small SameDiff listener fix (#4) * Various fixes (#6) * #7839 Fix for asXMatrix and tests * #7866 EmbeddingSequenceLayer dtype fix + test * #7856 SameDiff save/load stream methods * #7859 RegressionEvaluation rank 4 fix + tests + axis configuration * EvaluationBinary 3d/4d * More evaluation 3d/4d tests * #7847 Evaluation empty checks * Small test ifx * #7848 Fix median edge case * Improve DL4J samediff layer tests * [WIP] FastText wrapper implemented (#8) * FastText implemented * Some fixes * Fix shapes for wordsNearest * Validation of input vectors * Fixes * Fixed test * Thread tagged * Some tweaks * setContextClassLoader for DeallocatorServiceThread * Numpy format tests (#1) * Various fixes (#11) * #7852 SameDiff gather fix * #7892 SameDiff placeholder to constant conversion * #7890 validate input rank for MLN/CG init methods * Fix broken permute shape calculation * Permute and gather fixes * Tests * #7850 LogSumExp fix + test * Handful of test fixes * Empty arrays with non-scalar shapes (#10) * minor rearrangements for lambdas * empty tensors with non-scalar shapes * numpy empty tensors with non-scalar shapes * few more empty tweaks * Small fixes * conv3d signature update * micro fix in batchnorm mkldnn * Import fixes * Fix * MKL-DNN update * Small fill fix * fill with empty input + test * Fixes * Small error improvement * Fix * one special test * couple of fixes for lstm * Rewrite TFGraphMapper.getNDArrayFromTensor to be maintainable and less error prone * Fixes * FP16 * Unsigned * BFloat16 * Fill op - empty tweaks * - couple of fixes for empty arrays construction - stack updated * strided slice fix * one transform test * provide method for reducing shapeInfo in case of input array is empty * Fixed reduceAlongDimensions to use empty input properly. * couple of broadcast tests * couple of tests broadcast tests + tweak to make them pass * add check of non-empty to methods producing sub-arrays * Fixed reshapeC with zeros in shape. * complete empty check in reduce_... legacy ops * Concat and cumsum/prod * Tweak to empty shape inference on import * add empty check to the rest of reduce legacy ops * one more test * correct typo in evalReduceShapeInfoEmpty * Added tests for reduce_* ops to tests with zero shapes. * few more tests for empty reductions * Fixed strided_slice op with empty case and tests. * one more empty reduction test * Fixed strided_slice test. * add empty check to NDArray::reshapei * infOrMax * empty min/max with infinity tests * made unstack working correctly with empty arrays * few IndexReduce tests + tweaks for empty shapes * add test for empty concat * few tests fixed * Validation fix for reductions on empty shapes * Reverse fix * Reduction shape calc fixes * SameDiff.generateOutputVariable: don't use shape function to determine number of outputs * Range fix * - NDArray constructor updated for scalars/empty arrays - few tests fixed * More fixes * Empty creator fixes * concat fix * concat fix * TF import tests: allow 'both all NaN' and 'both all inf' to pass * Slice, zero fraction, and reshape fixes * transpose, gather * Zero fraction * scalar cast fix * Empty reduction axis support * few more tests fixed * Fixed input checks conforming with TF for concat op and tests. * few tests fixed * matmul scalar shape fix * Fixed checkout for data type and scalarity with concat to allow non-empty scalars with vector concats. * broadcast bool fix * few more tests * few more tests * correct evalReduceShapeInfoEmpty * argmax/argmin + tests * one more empty edge case + one more test * argmax/argmin/realdiv_bp tweaks * empty reshape test + fix * Helper fixes * Small fixes * Gather test fix * Gather test fix * Small fixes * reduce scalar zero values * scalar mean workaround * Remove debug code * along dim mean workaround * one more test * - equalsTo() tweak for empty arrays - one more test * broadcast tweaks
2019-06-15 13:34:34 +02:00
ConvolutionUtils::pooling3dBP(block, *input, *gradO, *gradI, kD, kH, kW, sD, sH, sW, pD, pH, pW, dD, dH, dW, 0, 1);
2019-06-06 14:21:15 +02:00
if(!isNCDHW) {
delete input;
delete gradI;
delete gradO;
}
// delete columns;
// delete columns2d;
// delete gradOVector;
return Status::OK();
}
DECLARE_SHAPE_FN(maxpool3dnew_bp) {
return SHAPELIST(ConstantShapeHelper::getInstance()->createShapeInfo(ShapeDescriptor(inputShape->at(0), ArrayOptions::dataType(inputShape->at(1)))));
}
}
}
#endif