cavis/libnd4j/include/ops/declarable/generic/convo/deconv2d_tf.cpp

153 lines
9.2 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
// @author Yurii Shyrma
//
#include <op_boilerplate.h>
#if NOT_EXCLUDED(OP_deconv2d)
#include <ops/declarable/CustomOperations.h>
#include <ops/declarable/helpers/convolutions.h>
namespace nd4j {
namespace ops {
//////////////////////////////////////////////////////////////////////////
CUSTOM_OP_IMPL(deconv2d_tf, 3, 1, false, 0, 9) {
auto gradO = INPUT_VARIABLE(2); // [bS, oH, oW, oC] (NHWC) or [bS, oC, oH, oW] (NCHW), epsilon_next
auto weights = INPUT_VARIABLE(1); // [kH, kW, iC, oC] always
auto gradIShape = INPUT_VARIABLE(0); // [4] - shape of input of conv2d (that is shape of gradI)
auto gradI = OUTPUT_VARIABLE(0); // [bS, iH, iW, iC] (NHWC) or [bS, iC, iH, iW] (NCHW), epsilon
int kH = INT_ARG(0) > 0 ? INT_ARG(0) : static_cast<int>(weights->sizeAt(0));// filter(kernel) height
int kW = INT_ARG(1) > 0 ? INT_ARG(1) : static_cast<int>(weights->sizeAt(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() > 9 ? !INT_ARG(9) : 1; // INT_ARG(9): 1-NHWC, 0-NCHW
const int rank = gradO->rankOf();
REQUIRE_TRUE(weights->rankOf() == rank, 0, "CUSTOM DECONV2D_TF OP: rank of weights array must be equal to 4, but got %i instead !", weights->rankOf());
REQUIRE_TRUE(gradIShape->rankOf() == 1, 0, "CUSTOM DECONV2D_TF OP: rank of array with output shape must be equal to 1, but got %i instead !", gradIShape->rankOf());
REQUIRE_TRUE(gradIShape->lengthOf() == rank, 0, "CUSTOM DECONV2D_TF OP: length of array with output shape must be equal to 4, but got %i instead !", gradIShape->lengthOf());
// create empty conv2d input array
NDArray input(gradO->ordering(), gradIShape->asVectorT<Nd4jLong>(), gradO->dataType(), block.launchContext());
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);
int trueoH, trueoW; // true output height, width
ConvolutionUtils::calcOutSizePool2D(trueoH, trueoW, kH, kW, sH, sW, pH, pW, dH, dW, iH, iW, isSameMode);
std::string expectedGradOShape = ShapeUtils::shapeAsString(ShapeUtils::composeShapeUsingDimsAndIdx({bS,oC,trueoH,trueoW, 0,indIOioC,indOoH,indOoH+1}));
std::string expectedWeightsShape = ShapeUtils::shapeAsString({kH, kW, iC, oC});
REQUIRE_TRUE(expectedGradOShape == ShapeUtils::shapeAsString(gradO), 0, "CUSTOM DECONV2D_TF OP: wrong shape of input array, basing on array with output shape expected is %s, but got %s instead !", expectedGradOShape.c_str(), ShapeUtils::shapeAsString(gradO).c_str());
REQUIRE_TRUE(expectedWeightsShape == ShapeUtils::shapeAsString(weights), 0, "CUSTOM DECONV2D_TF OP: wrong shape of weights array, expected is %s, but got %s instead !", expectedWeightsShape.c_str(), ShapeUtils::shapeAsString(weights).c_str());
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::conv2dBP(block, &input, weights, nullptr, gradO, gradI, nullptr, nullptr, kH,kW,sH,sW,pH,pW,dH,dW,isSameMode,isNCHW);
2019-06-06 14:21:15 +02:00
return Status::OK();
}
DECLARE_TYPES(deconv2d_tf) {
getOpDescriptor()
->setAllowedInputTypes(nd4j::DataType::ANY)
->setAllowedOutputTypes({ALL_FLOATS});
}
DECLARE_SHAPE_FN(deconv2d_tf) {
auto gradOShapeInfo = inputShape->at(2); // [bS, oH, oW, oC] (NHWC) or [bS, oC, oH, oW] (NCHW), epsilon_next
auto weightsShapeInfo = inputShape->at(1); // [kH, kW, iC, oC] always
auto gradIShapeShapeInfo = inputShape->at(0); // [4]
const int rank = 4;
REQUIRE_TRUE(weightsShapeInfo[0] == rank, 0, "CUSTOM DECONV2D_TF OP: rank of weights array must be equal to %i, but got %i instead !", rank, weightsShapeInfo[0]);
REQUIRE_TRUE(gradOShapeInfo[0] == rank, 0, "CUSTOM DECONV2D_TF OP: rank of input array must be equal to %i, but got %i instead !", rank, gradOShapeInfo[0]);
REQUIRE_TRUE(gradIShapeShapeInfo[0] == 1, 0, "CUSTOM DECONV2D_TF OP: rank of array with output shape must be equal to %i, but got %i instead !", 1, gradIShapeShapeInfo[0]);
const int kH = INT_ARG(0) > 0 ? INT_ARG(0) : static_cast<int>(shape::sizeAt(weightsShapeInfo, 0));// filter(kernel) height
const int kW = INT_ARG(1) > 0 ? INT_ARG(1) : static_cast<int>(shape::sizeAt(weightsShapeInfo, 1));// filter(kernel) width
const int sH = INT_ARG(2); // strides height
const int sW = INT_ARG(3); // strides width
const int pH = INT_ARG(4); // paddings height
const int pW = INT_ARG(5); // paddings width
const int dH = INT_ARG(6); // dilations height
const int dW = INT_ARG(7); // dilations width
const int isSameMode = INT_ARG(8); // 0-VALID, 1-SAME
const int isNCHW = block.getIArguments()->size() > 9 ? !INT_ARG(9) : 1; // INT_ARG(9): 1-NHWC, 0-NCHW
int indIOioC, indIiH, indWoC(3), indOoH;
if(!isNCHW) {
indIOioC = 3; indIiH = 1; indOoH = 1;
}
else {
indIOioC = 1; indIiH = 2; indOoH = 2;
}
std::vector<Nd4jLong> gradIShape = INPUT_VARIABLE(0)->template asVectorT<Nd4jLong>();
const int bS = gradIShape[0]; // batch size
const int iH = gradIShape[indIiH]; // input height
const int iW = gradIShape[indIiH+1]; // input width
const int iC = gradIShape[indIOioC]; // input channels
const int oC = weightsShapeInfo[indWoC+1]; // output channels
const int oH = gradOShapeInfo[indOoH+1]; // input height
const int oW = gradOShapeInfo[indOoH+2]; // input width
int trueiH, trueiW; // output height, width
ConvolutionUtils::calcOutSizeDeconv2D(trueiH, trueiW, kH, kW, sH, sW, pH, pW, dH, dW, oH, oW, isSameMode);
std::string expectedGradIShape = ShapeUtils::shapeAsString(ShapeUtils::composeShapeUsingDimsAndIdx({bS,iC,trueiH,trueiW, 0,indIOioC,indIiH,indIiH+1}));
std::string expectedWeightsShape = ShapeUtils::shapeAsString({kH, kW, iC, oC});
REQUIRE_TRUE(expectedGradIShape == ShapeUtils::shapeAsString(gradIShape), 0, "CUSTOM DECONV2D_TF OP: wrong shape of array with output shape, expected is %s, but got %s instead !", expectedGradIShape.c_str(), ShapeUtils::shapeAsString(gradIShape).c_str());
REQUIRE_TRUE(expectedWeightsShape == ShapeUtils::shapeAsString(weightsShapeInfo), 0, "CUSTOM DECONV2D_TF OP: wrong shape of weights array, expected is %s, but got %s instead !", expectedWeightsShape.c_str(), ShapeUtils::shapeAsString(weightsShapeInfo).c_str());
Nd4jLong shape[4];
shape[0] = bS;
if (isNCHW) {
shape[1] = iC;
shape[2] = iH;
shape[3] = iW;
} else {
shape[1] = iH;
shape[2] = iW;
shape[3] = iC;
}
return SHAPELIST(ConstantShapeHelper::getInstance()->createShapeInfo(ArrayOptions::dataType(weightsShapeInfo), shape::order(gradOShapeInfo), 4, shape));
}
}
}
#endif