cavis/libnd4j/include/ops/declarable/generic/convo/pooling/maxpool2d.cpp

227 lines
10 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 raver119@gmail.com, created on 29/10/17.
// @author Yurii Shyrma (iuriish@yahoo.com), changed on 09.05.2018
//
#include <op_boilerplate.h>
#if NOT_EXCLUDED(OP_maxpool2d)
#include <ops/declarable/CustomOperations.h>
#include <ops/declarable/helpers/convolutions.h>
namespace nd4j {
namespace ops {
//////////////////////////////////////////////////////////////////////////
// maxpool2d corresponds to poolingMode=0
CUSTOM_OP_IMPL(maxpool2d, 1, 1, false, 0, 9) {
auto input = INPUT_VARIABLE(0);
REQUIRE_TRUE(input->rankOf() == 4, 0, "MAXPOOL2D OP: input array should have rank of 4, but got %i instead", input->rankOf());
// 0,1 - kernel Height/Width; 2,3 - stride Height/Width; 4,5 - pad Height/Width; 6,7 - dilation Height/Width; 8 - same mode;
auto output = OUTPUT_VARIABLE(0);
const int kH = INT_ARG(0);
const int kW = INT_ARG(1);
const int sH = INT_ARG(2);
const int sW = INT_ARG(3);
int pH = INT_ARG(4);
int pW = INT_ARG(5);
const int dH = INT_ARG(6);
const int dW = INT_ARG(7);
const bool isSameMode = INT_ARG(8);
REQUIRE_TRUE(dH != 0 && dW != 0, 0, "MAXPOOL2D op: dilation must not be zero, but got instead {%i, %i}", dH, dW);
int oH = 0;
int oW = 0;
int isNCHW = block.getIArguments()->size() > 10 ? !INT_ARG(10) : 1; // INT_ARG(10): 1-NHWC, 0-NCHW
const int iH = isNCHW ? input->sizeAt(2) : input->sizeAt(1);
const int iW = isNCHW ? input->sizeAt(3) : input->sizeAt(2);
if (!isNCHW) {
input = input->permute({0, 3, 1, 2}); // [bS, iH, iW, iC] -> [bS, iC, iH, iW]
output = output->permute({0, 3, 1, 2}); // [bS, oH, oW, iC] -> [bS, iC, oH, oW]
}
ConvolutionUtils::calcOutSizePool2D(oH, oW, kH, kW, sH, sW, pH, pW, dH, dW, iH, iW, isSameMode);
if (isSameMode)
ConvolutionUtils::calcPadding2D(pH, pW, oH, oW, iH, iW, kH, kW, sH, sW, dH, dW);
// 0,1 - kernel Height/Width; 2,3 - stride Height/Width; 4,5 - pad Height/Width; 6,7 - dilation Height/Width; poolingMode; 9 - divisor;
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::pooling2d(block, *input, *output, kH, kW, sH, sW, pH, pW, dH, dW, PoolingType::MAX_POOL, 1);
2019-06-06 14:21:15 +02:00
if (!isNCHW) {
delete input;
delete output;
}
return Status::OK();
}
DECLARE_SYN(MaxPool2D, maxpool2d);
DECLARE_SYN(MaxPool, maxpool2d);
DECLARE_SYN(maxpool, maxpool2d);
DECLARE_TYPES(maxpool2d) {
getOpDescriptor()
->setAllowedInputTypes(nd4j::DataType::ANY)
->setSameMode(true);
}
DECLARE_SHAPE_FN(maxpool2d) {
//NDArray<T> *x = block.getVariables().at(0)->getNDArray();
Nd4jLong* inShape = inputShape->at(0);
Nd4jLong* shapeOf = shape::shapeOf(inShape);
// 0 - number of dimensions; 1,2 - kernel Height/Width; 3,4 - stride Height/Width; 5,6 - pad Height/Width; 7,8 - dilation Height/Width; 9,10 - input Height/Width; 11 - batch size; 12 - input depth; 13 - same mode;
int kH = INT_ARG(0);
int kW = INT_ARG(1);
int sH = INT_ARG(2);
int sW = INT_ARG(3);
int pH = INT_ARG(4);
int pW = INT_ARG(5);
int dH = INT_ARG(6);
int dW = INT_ARG(7);
int isSameMode = INT_ARG(8);
int isNCHW = block.getIArguments()->size() > 10 ? !INT_ARG(10) : 1; // INT_ARG(10): 1-NHWC, 0-NCHW
REQUIRE_TRUE(dH != 0 && dW != 0, 0, "MAXPOOL2D op: dilation must not be zero, but got instead {%i, %i}", dH, dW);
int bS = shapeOf[0];
int iC = isNCHW ? shapeOf[1] : shapeOf[3];
int iH = isNCHW ? shapeOf[2] : shapeOf[1];
int iW = isNCHW ? shapeOf[3] : shapeOf[2];
char order = shape::order(inShape); // output order must be equal to input order
// calculate output Height/Width
int oH, oW;
ConvolutionUtils::calcOutSizePool2D(oH, oW, kH, kW, sH, sW, pH, pW, dH, dW, iH, iW, isSameMode);
// allocate memory for new shape
Nd4jLong newShape[4];
newShape[0] = bS;
if (isNCHW) {
newShape[1] = iC;
newShape[2] = oH;
newShape[3] = oW;
} else {
newShape[1] = oH;
newShape[2] = oW;
newShape[3] = iC;
}
return SHAPELIST(ConstantShapeHelper::getInstance()->createShapeInfo(ShapeDescriptor(ArrayOptions::dataType(inShape), order, newShape, 4)));
}
DECLARE_TYPES(maxpool2d_bp) {
getOpDescriptor()
->setAllowedInputTypes(nd4j::DataType::ANY)
->setAllowedOutputTypes({ALL_FLOATS});
}
//////////////////////////////////////////////////////////////////////////
CUSTOM_OP_IMPL(maxpool2d_bp, 2, 1, false, 0, 10) {
auto input = INPUT_VARIABLE(0); // [bS, iH, iW, iC] (NHWC) or [bS, iC, iH, iW] (NCHW)
auto gradO = INPUT_VARIABLE(1); // [bS, oH, oW, oC] (NHWC) or [bS, oC, oH, oW] (NCHW), epsilon_next
auto gradI = OUTPUT_VARIABLE(0); // [bS, iH, iW, iC] (NHWC) or [bS, iC, iH, iW] (NCHW), epsilon
int kH = INT_ARG(0); // filter(kernel) height
int kW = INT_ARG(1); // filter(kernel) width
int sH = INT_ARG(2); // strides height
int sW = INT_ARG(3); // strides width
int pH = INT_ARG(4); // paddings height
int pW = INT_ARG(5); // paddings width
int dH = INT_ARG(6); // dilations height
int dW = INT_ARG(7); // dilations width
int isSameMode = INT_ARG(8); // 0-VALID, 1-SAME
int isNCHW = block.getIArguments()->size() > 10 ? !INT_ARG(10) : 1; // INT_ARG(10): 1-NHWC, 0-NCHW
REQUIRE_TRUE(input->rankOf() == 4, 0, "MAXPOOL2D_BP op: input should have rank of 4, but got %i instead", input->rankOf());
REQUIRE_TRUE(dH != 0 && dW != 0, 0, "MAXPOOL2D_BP op: dilation must not be zero, but got instead {%i, %i}", dH, dW);
int bS, iC, iH, iW, oC, oH, oW; // batch size, input channels, input height/width, output channels, output height/width;
int indIOioC, indIiH, indWoC, indWiC, indWkH, indOoH; // corresponding indexes
ConvolutionUtils::getSizesAndIndexesConv2d(isNCHW, *input, *gradO, bS, iC, iH, iW, oC, oH, oW, indIOioC, indIiH, indWiC, indWoC, indWkH, indOoH);
std::string expectedGradOShape = ShapeUtils::shapeAsString(ShapeUtils::composeShapeUsingDimsAndIdx({bS,iC,oH,oW, 0,indIOioC,indIiH,indIiH+1}));
std::string expectedGradIShape = ShapeUtils::shapeAsString(ShapeUtils::composeShapeUsingDimsAndIdx({bS,iC,iH,iW, 0,indIOioC,indIiH,indIiH+1}));
REQUIRE_TRUE(expectedGradOShape == ShapeUtils::shapeAsString(gradO), 0, "MAXPOOL2D_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, "MAXPOOL2D_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(!isNCHW) {
input = input->permute({0, 3, 1, 2}); // [bS, iH, iW, iC] -> [bS, iC, iH, iW]
gradI = gradI->permute({0, 3, 1, 2}); // [bS, iH, iW, iC] -> [bS, iC, iH, iW]
gradO = gradO->permute({0, 3, 1, 2}); // [bS, oH, oW, iC] -> [bS, iC, oH, oW]
}
if(isSameMode) // SAME
ConvolutionUtils::calcPadding2D(pH, pW, oH, oW, iH, iW, kH, kW, sH, sW, dH, dW);
// NDArray<T> columnsWrongShape(input->ordering(), {bS, iC, oH, oW, kH, kW}, input->getWorkspace());
// NDArray<T>* columns = columnsWrongShape.permute({0, 1, 4, 5, 2, 3}); // [bS, iC, oH, oW, kH, kW] -> [bS, iC, kH, kW, oH, oW]
// input->template applyTransform<simdOps::Im2col<T>>(columns, std::vector<T>({(T)kH, (T)kW, (T)sH, (T)sW, (T)pH, (T)pW, (T)dH, (T)dW, (T)0.f, (T)0.f}).data());
// NDArray<T>* columns2d = columnsWrongShape.reshape('c', {bS*iC*oH*oW, kH*kW});
// NDArray<T>* gradOVector = gradO->reshape('c', {(int) gradO->lengthOf(), 1});
// columns2d->template applyTransform<simdOps::IsMax<T>>(std::vector<T>({(T)1., (T)1.}).data());
// columns2d->muliColumnVector(gradOVector);
// columns->template applyTransform<simdOps::Col2Im<T>>(gradI, std::vector<T>({(T)sH, (T)sW, (T)pH, (T)pW, (T)iH, (T)iW, (T)dH, (T)dW}).data());
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::pooling2dBP(block, *input, *gradO, *gradI, kH, kW, sH, sW, pH, pW, dH, dW, 0., 1.);
2019-06-06 14:21:15 +02:00
if(!isNCHW) {
delete input;
delete gradI;
delete gradO;
}
// delete columns;
// delete columns2d;
// delete gradOVector;
return Status::OK();
}
DECLARE_SYN(MaxPool2D_bp, maxpool2d_bp);
DECLARE_SYN(MaxPool_bp, maxpool2d_bp);
DECLARE_SHAPE_FN(maxpool2d_bp) {
REQUIRE_TRUE(inputShape->at(0)[0] == 4, 0, "MAXPOOL2D_BP op: input array must be 4D, but got %i instead!", inputShape->at(0)[0]);
REQUIRE_TRUE(inputShape->at(1)[0] == 4, 0, "MAXPOOL2D_BP op: output's gradient array (next epsilon) must be 4D, but got %i instead!", inputShape->at(1)[0]);
return SHAPELIST(ConstantShapeHelper::getInstance()->createShapeInfo(ShapeDescriptor(inputShape->at(0), ArrayOptions::dataType(inputShape->at(1)))));
}
}
}
#endif