cavis/libnd4j/include/ops/declarable/helpers/cpu/activations.cpp

401 lines
16 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 19.04.2018
// @author raver119@gmail.com
//
#include <ops/declarable/helpers/activations.h>
#include <ShapeUtils.h>
#include <numeric>
#include <ConstantTadHelper.h>
#include <execution/Threads.h>
2019-06-06 14:21:15 +02:00
namespace nd4j {
namespace ops {
namespace helpers {
template <typename T>
static void softMaxForVector_(void *input, Nd4jLong *inShapeInfo, void *output, Nd4jLong *outShapeInfo) {
T* inBuff = reinterpret_cast<T *>(input);
T* outBuff = reinterpret_cast<T *>(output);
T max = -DataTypeUtils::max<T>();
T sum = 0.;
int inEWS = shape::elementWiseStride(inShapeInfo);
int outEWS = shape::elementWiseStride(outShapeInfo);
int length = shape::length(inShapeInfo);
if (inEWS >= 1 && outEWS >= 1) {
if (inEWS == 1 && outEWS == 1) {
for (int i = 0; i < length; i++)
max = nd4j::math::nd4j_max<T>(max, inBuff[i]);
for (int i = 0; i < length; i++) {
outBuff[i] = nd4j::math::nd4j_exp<T, T>(inBuff[i] - max);
sum += outBuff[i];
}
PRAGMA_OMP_SIMD
for (int i = 0; i < length; i++)
outBuff[i] /= sum;
}
else {
for (int i = 0; i < length; i++)
max = nd4j::math::nd4j_max<T>(max, inBuff[i * inEWS]);
for (int i = 0; i < length; i++) {
T r = nd4j::math::nd4j_exp<T, T>(inBuff[i * inEWS] - max);
outBuff[i * outEWS] = r;
sum += r;
}
PRAGMA_OMP_SIMD
for (int i = 0; i < length; i++)
outBuff[i * outEWS] /= sum;
}
}
}
///////////////////////////////////////////////////////////////////
template <typename T>
void static _softMaxDerivForVector(nd4j::LaunchContext * context, const void *input, const Nd4jLong *inShapeInfo, void *output) {
const T* inBuff = reinterpret_cast<const T *>(input);
T* outBuff = reinterpret_cast<T *>(output);
T max = -DataTypeUtils::max<T>();
T sum = 0.;
int length = shape::length(inShapeInfo);
for (int i = 0; i < length; i++) {
const Nd4jLong offset = shape::getIndexOffset(i, inShapeInfo);
2019-06-06 14:21:15 +02:00
max = nd4j::math::nd4j_max<T>(max, inBuff[offset]);
}
for (int i = 0; i < length; i++) {
const Nd4jLong offset = shape::getIndexOffset(i, inShapeInfo);
2019-06-06 14:21:15 +02:00
outBuff[offset] = nd4j::math::nd4j_exp<T, T>(inBuff[offset] - max);
sum += outBuff[offset];
}
2019-06-06 14:21:15 +02:00
for (int i = 0; i < length; i++) {
const Nd4jLong offset = shape::getIndexOffset(i, inShapeInfo);
2019-06-06 14:21:15 +02:00
outBuff[offset] /= sum;
outBuff[offset] *= (1.f - outBuff[offset]); // derivative
}
}
///////////////////////////////////////////////////////////////////
void softmaxDerivative(nd4j::LaunchContext * context, const NDArray& input, NDArray& output, const int dimension) {
const int rank = input.rankOf();
int temp;
if(shape::isCommonVector(input.getShapeInfo(), temp)) {
BUILD_SINGLE_SELECTOR(input.dataType(), _softMaxDerivForVector, (context, input.getBuffer(), input.getShapeInfo(), output.buffer()), FLOAT_TYPES);
}
else {
Shyrma temp (#131) * - specifying template instantiation for certain types in float16 and bloat16 Signed-off-by: Yurii <iuriish@yahoo.com> * - polishing bfloat16 and float16 member functions template specialization Signed-off-by: Yurii <iuriish@yahoo.com> * - rewrite and overload array +-*/ scalar and scalar +-*/ arr in NDAray class Signed-off-by: Yurii <iuriish@yahoo.com> * - make corrections which have to do with and rvalue lvalue conversions Signed-off-by: Yurii <iuriish@yahoo.com> * - provide move semantic in NDArray operators array +-/* array Signed-off-by: Yurii <iuriish@yahoo.com> * float16/bfloat16 tweaks Signed-off-by: raver119 <raver119@gmail.com> * one more tweak Signed-off-by: raver119 <raver119@gmail.com> * - make float16 and bfloat16 to compile successfully on cuda Signed-off-by: Yurii <iuriish@yahoo.com> * - do not use resources of view-like arrays when move semantics is applied Signed-off-by: Yurii <iuriish@yahoo.com> * - get rid of pointers in signatures NDArray methods 1 Signed-off-by: Yurii <iuriish@yahoo.com> * - correction of signature of NDArray::dup method Signed-off-by: Yurii <iuriish@yahoo.com> * - correction of signature of NDArray::reduceAlongDimension method Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of NDArray::applyIndexReduce and applyTrueBroadcast methods Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of NDArray::applyReduce3 and varianceAlongDimension methods Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of NDArray::tensorsAlongDimension and diagonal methods Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of NDArray::allTensorsAlongDimension Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of NDArray::reduceAlongDimension 2 Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of NDArray::applyTransform 2 Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of NDArray::applyPairwiseTransform 2 Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of NDArray::applyBroadcast 2 Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of NDArray::applyTrueBroadcast 2 Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of NDArray::applyScalar and applyScalarArr Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of NDArray::lambda methods Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of NDArray::reduce3 methods 2 Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of following NDArray methods: add/sub/mul/div row/column and fillAsTriangular Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of NDArray::tileToShape methods Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of NDArray::isShapeSameStrict method Signed-off-by: Yurii <iuriish@yahoo.com> * minor corrections in tests Signed-off-by: Yurii <iuriish@yahoo.com> * - replace reduce op in batchnorm mkldnn Signed-off-by: Yurii <iuriish@yahoo.com> * - add explicit templates instantiations for operator+(NDArray&&. const scalar) Signed-off-by: Yurii <iuriish@yahoo.com> * - corrections of casts in float16/bfloat16 Signed-off-by: Yurii <iuriish@yahoo.com> * - provide move semantics in following NDArray methods: transform, applyTrueBroadcast, transpose, reshape, permute Signed-off-by: Yurii <iuriish@yahoo.com> * - get rid of input array A duplicate in svd cuda op Signed-off-by: Yurii <iuriish@yahoo.com> * - avoid available bug in svd cuda API Signed-off-by: Yurii <iuriish@yahoo.com> * - add temporary global memory buffer in svd cuda when calcUV = false and m != n Signed-off-by: Yurii <iuriish@yahoo.com> * - remove test with blfoat16 type for betainC Signed-off-by: Yurii <iuriish@yahoo.com> * - resolve conflicts after master has been merged in Signed-off-by: Yurii <iuriish@yahoo.com> * - changed type of affected input array in fused_batch_norm Signed-off-by: Yurii <iuriish@yahoo.com> * - add several explicit type castings Signed-off-by: Yurii <iuriish@yahoo.com> * - add ND4J_EXPORT to operators Signed-off-by: Yurii <iuriish@yahoo.com> * - add explicit template types in instantiations of template arithm operators of NDArray class Signed-off-by: Yurii <iuriish@yahoo.com> * - one more test fix Signed-off-by: Yurii <iuriish@yahoo.com> Co-authored-by: raver119 <raver119@gmail.com>
2019-12-20 20:35:39 +01:00
auto maxAlongDim = const_cast<NDArray&>(input).reduceAlongDimension(reduce::Max, {dimension}, true);
(input - maxAlongDim).applyTransform(transform::Exp, output); // output contains exponents temporarily
auto sumAlongDim = output.reduceAlongDimension(reduce::Sum, {dimension}, true);
2019-06-06 14:21:15 +02:00
output /= sumAlongDim;
output *= (1.f - output); // derivative
}
}
///////////////////////////////////////////////////////////////////
void softMaxForVector(nd4j::LaunchContext * context, const NDArray& input, NDArray& output) {
if(!input.isVector() || !output.isVector())
throw std::runtime_error("ops::helpers::softMaxForVector function: input and output arrays must be vectors !");
auto xType = input.dataType();
BUILD_SINGLE_SELECTOR(xType, softMaxForVector_, (input.getBuffer(), input.getShapeInfo(), output.buffer(), output.shapeInfo()), FLOAT_TYPES);
}
Merge master to upstream (#7945) * Shugeo strided slice zeros (#14) * Modified strided_slice op to properly work with empty-like shapes. * Fixed test for reduce_mean with empty-like input. * [WIP] Last merge (#15) * 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 * [WIP] Fixing outstanding issues for NLP (#9) * Avoid using not-inited objects * Test fixed. * Redundant method avoided for models like FastText * KMeans++ implementation * KMeans++ implementation * Disable parallel execution * KMeans++ * Tests * Dev branch merge (#16) * SameDiff: convertDataType and gradient check util improvements (#12) * GradCheck util improvements * StopGradient constructor + test * SameDiff: Add datatype conversion * Javadoc and add DataType.isNumerical() * Small fix * Fix SameDiff TF import test cases intermediate naming (workaround for bad default) * TFGraphTestAllHelper: check intermediates in execution order * Add missing debug listener * [WIP] lstmBlock fix + other changes (#13) - fixes lstmBlock issue - changes NDArray method reshape(), permute(), transpose() by making them return instance instead of pointer - CheckNumerics op - fixes for ReduceBool IsInfOrNan & IsFinite * Small test fix * CheckNumerics op wrapper * Fix some issues on master (#17) * Fix DataVec test issue * Fix issue with dl4j SameDiff output layer * Dtype fix for lambda layers * #7912 BertIterator dtype fix (use float32 not global default) * [WIP] Next set of CUDA stuff (#7) New CUDA implementations and improvements * bad file * Dev branch master merge (#23) * SameDiff: convertDataType and gradient check util improvements (#12) * GradCheck util improvements * StopGradient constructor + test * SameDiff: Add datatype conversion * Javadoc and add DataType.isNumerical() * Small fix * Fix SameDiff TF import test cases intermediate naming (workaround for bad default) * TFGraphTestAllHelper: check intermediates in execution order * Add missing debug listener * [WIP] lstmBlock fix + other changes (#13) - fixes lstmBlock issue - changes NDArray method reshape(), permute(), transpose() by making them return instance instead of pointer - CheckNumerics op - fixes for ReduceBool IsInfOrNan & IsFinite * Small test fix * CheckNumerics op wrapper * Compatibility of deserialization (#18) Signed-off-by: Alexander Stoyakin <alexander.stoyakin@gmail.com> * SameDiff: add activation gradient checking support for debugging (#19) * SameDiff gradient checker: first pass on activation gradient checks * Fixes + tests for activation gradient checking * Javadoc * [WIP] Some nd4j data type corrections (#20) * Adjust data type * Set correct Data type. * Size of proper data type. * fix averaged cpu load (#22) * SameDiff ops, TF import and fixes (#24) * CheckNumerics tests + fixes + misc fixes Signed-off-by: AlexDBlack <blacka101@gmail.com> * Fake quant Signed-off-by: AlexDBlack <blacka101@gmail.com> * Fixes Signed-off-by: AlexDBlack <blacka101@gmail.com> * FakeQuantWithMinMaxArgs Signed-off-by: AlexDBlack <blacka101@gmail.com> * CheckNumerics fix Signed-off-by: AlexDBlack <blacka101@gmail.com> * Fix libnd4j ALL_INTS and ALL_FLOATS declaration (uint and bfloat types) Signed-off-by: AlexDBlack <blacka101@gmail.com> * Small fix Signed-off-by: AlexDBlack <blacka101@gmail.com> * Javadoc Signed-off-by: AlexDBlack <blacka101@gmail.com> * Exception tweak Signed-off-by: AlexDBlack <blacka101@gmail.com> * fix Signed-off-by: AlexDBlack <blacka101@gmail.com> * Fix for out of scope stack allocated var use Signed-off-by: AlexDBlack <blacka101@gmail.com> * Ignores Signed-off-by: AlexDBlack <blacka101@gmail.com> * Ignore for known failing test (already logged issue) Signed-off-by: AlexDBlack <blacka101@gmail.com> * Merge upstream to fork (#25) * Add thousand-separator commas to TotalParams (#7915) * Add thousand-separator commas to TotalParams The number of parameters can be quite large, and it would help the reading of the summary printout to have the TotalParams column & values at the bottom have thousand-separator-commas in them. * Add thousand-separator commas to MultiLayerNetwork Corresponding change to MultiLayerNetwork Signed-off-by: Jxtps Jxtps <jxtps435@gmail.com> * Update contributing and issue/PR templates (#7934) Signed-off-by: AlexDBlack <blacka101@gmail.com> * Fix link to AdaDelta paper (#7942) Fix link to AdaDelta paper hosted on matthewzeiler.com Signed-off-by: Jxtps * Fixes, and ignores for known/logged failing issues (#7943) Signed-off-by: AlexDBlack <blacka101@gmail.com> * SameDiff + DL4J/SameDiff: Multiple fixes (#28) * #7919 HDF5 attribute buffer length fix Signed-off-by: AlexDBlack <blacka101@gmail.com> * #7909 Arbiter constructor exception ux improvements Signed-off-by: AlexDBlack <blacka101@gmail.com> * #7925 RNN output layer length checks Signed-off-by: AlexDBlack <blacka101@gmail.com> * #7939 Add listener for validating inputs are not incorrectly modified Signed-off-by: AlexDBlack <blacka101@gmail.com> * #7939 Integrate NonInplaceValidationListener into tests * #7844 DL4J SameDiff fixes for variable minibatch size * DL4J SameDiff fixes - ensure gradient for input placeholder is available Signed-off-by: AlexDBlack <blacka101@gmail.com> * Tweaks to ExternalErrorsFunction - use placeholders, make more robust * Another fix * More fixes * More SameDiff/DL4J fixes * Scope out scalar array creation in BaseScalarOp * Remove debug code Signed-off-by: AlexDBlack <blacka101@gmail.com> * [WIP] Final dev branch merge (#29) * SameDiff: convertDataType and gradient check util improvements (#12) * GradCheck util improvements * StopGradient constructor + test * SameDiff: Add datatype conversion * Javadoc and add DataType.isNumerical() * Small fix * Fix SameDiff TF import test cases intermediate naming (workaround for bad default) * TFGraphTestAllHelper: check intermediates in execution order * Add missing debug listener * [WIP] lstmBlock fix + other changes (#13) - fixes lstmBlock issue - changes NDArray method reshape(), permute(), transpose() by making them return instance instead of pointer - CheckNumerics op - fixes for ReduceBool IsInfOrNan & IsFinite * Small test fix * CheckNumerics op wrapper * Compatibility of deserialization (#18) Signed-off-by: Alexander Stoyakin <alexander.stoyakin@gmail.com> * SameDiff: add activation gradient checking support for debugging (#19) * SameDiff gradient checker: first pass on activation gradient checks * Fixes + tests for activation gradient checking * Javadoc * [WIP] Some nd4j data type corrections (#20) * Adjust data type * Set correct Data type. * Size of proper data type. * fix averaged cpu load (#22) * [WIP] Multiple dataset iterators (#27) * Splitting dataset into arbitrary number * Fixes * Multiple split of iterator * Test * Test * Some fixes * signature change * one more tweak Signed-off-by: raver119 <raver119@gmail.com> * one more test for sequential use of DataSetIteratorSplitter Signed-off-by: raver119 <raver119@gmail.com> * Fixes * Fixes * one more test for Alexander Signed-off-by: raver119 <raver119@gmail.com> * Some fixes * Some fixes * one more test for Alexander Signed-off-by: raver119 <raver119@gmail.com> * minor test fix Signed-off-by: raver119 <raver119@gmail.com> * Some fixes * Some fixes * couple of assertions tweaked Signed-off-by: raver119 <raver119@gmail.com> * MDS splitter test :/ Signed-off-by: raver119 <raver119@gmail.com> * Minor refactoring * Multi dataset * Some fixes * More tests * Small number of test fixes/improvements (failures on CI) (#31) Signed-off-by: AlexDBlack <blacka101@gmail.com> * [WIP] More CUDA stuff (#26) * initial commit Signed-off-by: raver119 <raver119@gmail.com> * LRN BP CUDA Signed-off-by: raver119 <raver119@gmail.com> * less memory Signed-off-by: raver119 <raver119@gmail.com> * Fixed bug with crop_and_resize op helper. * get rid of unnecessary index-calculation dunction Signed-off-by: Yurii <yurii@skymind.io> * Fixed sort with nth_element cuda-based helper. * Refactored nth_element. * Refactored nth_element op and tests. * Modified usage of dim array with sortTad routine. * Refactored main routine of helper for non_max_image_suppression op. * non_max_image_suppression op helper with cuda kernel implementation. Initial revision. * fix vol2col cuda kernel * meh Signed-off-by: raver119 <raver119@gmail.com> * topK concept Signed-off-by: raver119 <raver119@gmail.com> * unsorted topK with scanWitdh of 1 Signed-off-by: raver119 <raver119@gmail.com> * correct vol2col tests * sorted/unsorted topK Signed-off-by: raver119 <raver119@gmail.com> * implementation and fixing col2im/col2vol * Corrected usage flags with input/output with reverse op. * dup is const now Signed-off-by: raver119 <raver119@gmail.com> * percentile op Signed-off-by: raver119 <raver119@gmail.com> * group tests for mapool2d Signed-off-by: Yurii <yurii@skymind.io> * special test for george Signed-off-by: raver119 <raver119@gmail.com> * less threads for sortTad Signed-off-by: raver119 <raver119@gmail.com> * provide conv2d for cuda Signed-off-by: Yurii <yurii@skymind.io> * remove auther in sort tad kernel code Signed-off-by: Yurii <yurii@skymind.io> * provide depthwise_conv2d for cuda Signed-off-by: Yurii <yurii@skymind.io> * - max_pooling_with_argmax - null check for special use Signed-off-by: raver119 <raver119@gmail.com> * dts cuda Signed-off-by: raver119 <raver119@gmail.com> * provide sconv2d for cuda Signed-off-by: Yurii <yurii@skymind.io> * std cuda Signed-off-by: raver119 <raver119@gmail.com> * Refactored non_max_suppression op to conform TF implementation. * Improved suppression helper. * provide pooling3d for cuda Signed-off-by: Yurii <yurii@skymind.io> * minor lstm rearrangements Signed-off-by: raver119 <raver119@gmail.com> * more of minor lstm rearrangements Signed-off-by: raver119 <raver119@gmail.com> * (bi)dynamic_rnn Signed-off-by: raver119 <raver119@gmail.com> * templates init order Signed-off-by: raver119 <raver119@gmail.com> * Refactored non_max_suppression op. * Added cuda kernel for non_max_suppression. * CPU sort by key/value Signed-off-by: raver119 <raver119@gmail.com> * CPU sort TAD by key/value Signed-off-by: raver119 <raver119@gmail.com> * CPU sort TAD by key/value tests Signed-off-by: raver119 <raver119@gmail.com> * Eliminate compiler error with cuda implementation. * - repaired gradCheck in cuda - provide conv2d_bp for cuda Signed-off-by: Yurii <yurii@skymind.io> * missed signature Signed-off-by: raver119 <raver119@gmail.com> * provide depthwise_conv2d_bp for cuda Signed-off-by: Yurii <yurii@skymind.io> * Implementation of lup helper with cuda kernel. Initial commit. * further work on backprops for convolutions Signed-off-by: Yurii <yurii@skymind.io> * CUDA linear sort by key/val Signed-off-by: raver119 <raver119@gmail.com> * CUDA tad sort by key/val Signed-off-by: raver119 <raver119@gmail.com> * start providing of backprop for pooling2d/3d Signed-off-by: Yurii <yurii@skymind.io> * Added atomicAdd for bool datatype. * dynamic partition concept Signed-off-by: raver119 <raver119@gmail.com> * dynamic partition concept Signed-off-by: raver119 <raver119@gmail.com> * dynamic partition scalar CUDA Signed-off-by: raver119 <raver119@gmail.com> * important comment Signed-off-by: raver119 <raver119@gmail.com> * fix pooling2d/3d backprop helpers Signed-off-by: Yurii <yurii@skymind.io> * Added non-linear test with dynamic_partition. * Improved test for dynamic_partition. * dynamic_partition TAD concept Signed-off-by: raver119 <raver119@gmail.com> * - dynamic_partition TAD CUDA impl - dynamic_partition TAD CPU fix Signed-off-by: raver119 <raver119@gmail.com> * - rewrite cpu code for usampling2d/3d - write cuda code for usampling2d/3d Signed-off-by: Yurii <yurii@skymind.io> * dynamic_stitch CUDA vector case Signed-off-by: raver119 <raver119@gmail.com> * dynamic_stitch CUDA TAD case concept Signed-off-by: raver119 <raver119@gmail.com> * dynamic_stitch CUDA TAD case impl Signed-off-by: raver119 <raver119@gmail.com> * Added tests for dynamic_stitch 3D-4D cases. * minor tests tweaks Signed-off-by: raver119 <raver119@gmail.com> * Fixed type check for dynamic stitch. * min/max bp Signed-off-by: raver119 <raver119@gmail.com> * rewrite code for upsampling2d/3d cpu Signed-off-by: Yurii <yurii@skymind.io> * reduce min/max/norm_max bp Signed-off-by: raver119 <raver119@gmail.com> * lup implementation. Additional enhancements. * provide code for upsamling2d/3d backprop Signed-off-by: Yurii <yurii@skymind.io> * weightedCrossEntropyWithLogits Signed-off-by: raver119 <raver119@gmail.com> * Fixed template math atomicMul for 64bit ints. * Refactored dynamic_partition_bp op. * inverseBroadcast fix Signed-off-by: raver119 <raver119@gmail.com> * DynamicPartitionBP test datatype fixed. * - nd4j_atomicMul Windows fix - cpu/NDArrayLambda.hpp excluded from CUDA Signed-off-by: raver119 <raver119@gmail.com>
2019-06-27 17:37:04 +02:00
///////////////////////////////////////////////////////////////////
2019-06-06 14:21:15 +02:00
template <typename T>
void logSoftMaxForVector_(void *input, Nd4jLong *inShapeInfo, void *output, Nd4jLong *outShapeInfo) {
auto inBuff = reinterpret_cast<T *>(input);
auto outBuff = reinterpret_cast<T *>(output);
T max = -DataTypeUtils::max<T>();
T sum = 0;
auto inEWS = shape::elementWiseStride(inShapeInfo);
auto length = shape::length(inShapeInfo);
if (inEWS == 1) {
for (int i = 0; i < length; i++)
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
max = nd4j::math::nd4j_max<T>(max, inBuff[i]);
2019-06-06 14:21:15 +02:00
PRAGMA_OMP_SIMD_SUM(sum)
for (int i = 0; i < length; i++) {
outBuff[i] = nd4j::math::nd4j_exp<T,T>(inBuff[i] - max);
sum += outBuff[i];
}
PRAGMA_OMP_SIMD
for (int i = 0; i < length; i++) {
outBuff[i] /= sum;
outBuff[i] = nd4j::math::nd4j_log<T,T>(outBuff[i]);
}
}
else if (inEWS > 1) {
PRAGMA_OMP_SIMD_MAX(max)
for (int i = 0; i < length; i++)
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
max = nd4j::math::nd4j_max<T>(max, inBuff[i * inEWS]);
2019-06-06 14:21:15 +02:00
PRAGMA_OMP_SIMD_SUM(sum)
for (int i = 0; i < length; i++) {
outBuff[i * inEWS] = nd4j::math::nd4j_exp<T,T>(inBuff[i * inEWS] - max);
sum += outBuff[i * inEWS];
}
PRAGMA_OMP_SIMD
for (int i = 0; i < length; i++) {
outBuff[i * inEWS] /= sum;
outBuff[i * inEWS] = nd4j::math::nd4j_log<T, T>(outBuff[i * inEWS]);
}
}
}
///////////////////////////////////////////////////////////////////
void logSoftMaxForVector(nd4j::LaunchContext * context, const NDArray& input, NDArray& output) {
if(!input.isVector() || !output.isVector())
throw std::runtime_error("ops::helpers::logSoftMaxForVector function input and output arrays must be vectors !");
auto xType = input.dataType();
BUILD_SINGLE_SELECTOR(xType, logSoftMaxForVector_, (input.getBuffer(), input.getShapeInfo(), output.buffer(), output.shapeInfo()), FLOAT_TYPES);
}
//////////////////////////////////////////////////////////////////////////
template <typename T>
static void softmax_(nd4j::LaunchContext * context, const NDArray& input, NDArray& output, const int dimension) {
const int rank = input.rankOf();
if(input.isVector()) {
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
2019-06-06 14:21:15 +02:00
if(rank == 1 || input.sizeAt(dimension) != 1)
softMaxForVector_<T>(input.getBuffer(), input.getShapeInfo(), output.buffer(), output.getShapeInfo());
else
output = 1.;
}
Shyrma temp (#131) * - specifying template instantiation for certain types in float16 and bloat16 Signed-off-by: Yurii <iuriish@yahoo.com> * - polishing bfloat16 and float16 member functions template specialization Signed-off-by: Yurii <iuriish@yahoo.com> * - rewrite and overload array +-*/ scalar and scalar +-*/ arr in NDAray class Signed-off-by: Yurii <iuriish@yahoo.com> * - make corrections which have to do with and rvalue lvalue conversions Signed-off-by: Yurii <iuriish@yahoo.com> * - provide move semantic in NDArray operators array +-/* array Signed-off-by: Yurii <iuriish@yahoo.com> * float16/bfloat16 tweaks Signed-off-by: raver119 <raver119@gmail.com> * one more tweak Signed-off-by: raver119 <raver119@gmail.com> * - make float16 and bfloat16 to compile successfully on cuda Signed-off-by: Yurii <iuriish@yahoo.com> * - do not use resources of view-like arrays when move semantics is applied Signed-off-by: Yurii <iuriish@yahoo.com> * - get rid of pointers in signatures NDArray methods 1 Signed-off-by: Yurii <iuriish@yahoo.com> * - correction of signature of NDArray::dup method Signed-off-by: Yurii <iuriish@yahoo.com> * - correction of signature of NDArray::reduceAlongDimension method Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of NDArray::applyIndexReduce and applyTrueBroadcast methods Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of NDArray::applyReduce3 and varianceAlongDimension methods Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of NDArray::tensorsAlongDimension and diagonal methods Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of NDArray::allTensorsAlongDimension Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of NDArray::reduceAlongDimension 2 Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of NDArray::applyTransform 2 Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of NDArray::applyPairwiseTransform 2 Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of NDArray::applyBroadcast 2 Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of NDArray::applyTrueBroadcast 2 Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of NDArray::applyScalar and applyScalarArr Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of NDArray::lambda methods Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of NDArray::reduce3 methods 2 Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of following NDArray methods: add/sub/mul/div row/column and fillAsTriangular Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of NDArray::tileToShape methods Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of NDArray::isShapeSameStrict method Signed-off-by: Yurii <iuriish@yahoo.com> * minor corrections in tests Signed-off-by: Yurii <iuriish@yahoo.com> * - replace reduce op in batchnorm mkldnn Signed-off-by: Yurii <iuriish@yahoo.com> * - add explicit templates instantiations for operator+(NDArray&&. const scalar) Signed-off-by: Yurii <iuriish@yahoo.com> * - corrections of casts in float16/bfloat16 Signed-off-by: Yurii <iuriish@yahoo.com> * - provide move semantics in following NDArray methods: transform, applyTrueBroadcast, transpose, reshape, permute Signed-off-by: Yurii <iuriish@yahoo.com> * - get rid of input array A duplicate in svd cuda op Signed-off-by: Yurii <iuriish@yahoo.com> * - avoid available bug in svd cuda API Signed-off-by: Yurii <iuriish@yahoo.com> * - add temporary global memory buffer in svd cuda when calcUV = false and m != n Signed-off-by: Yurii <iuriish@yahoo.com> * - remove test with blfoat16 type for betainC Signed-off-by: Yurii <iuriish@yahoo.com> * - resolve conflicts after master has been merged in Signed-off-by: Yurii <iuriish@yahoo.com> * - changed type of affected input array in fused_batch_norm Signed-off-by: Yurii <iuriish@yahoo.com> * - add several explicit type castings Signed-off-by: Yurii <iuriish@yahoo.com> * - add ND4J_EXPORT to operators Signed-off-by: Yurii <iuriish@yahoo.com> * - add explicit template types in instantiations of template arithm operators of NDArray class Signed-off-by: Yurii <iuriish@yahoo.com> * - one more test fix Signed-off-by: Yurii <iuriish@yahoo.com> Co-authored-by: raver119 <raver119@gmail.com>
2019-12-20 20:35:39 +01:00
else if(input.isSameShapeStrict(output)) {
2019-06-06 14:21:15 +02:00
TadPack tadPack = nd4j::ConstantTadHelper::getInstance()->tadForDimensions(input.getShapeInfo(), dimension);
2019-06-06 14:21:15 +02:00
Nd4jLong* tadShapeInfo = tadPack.primaryShapeInfo();
Nd4jLong* tadOffsets = tadPack.primaryOffsets();
const uint numOfSubArrs = tadPack.numberOfTads();
const uint tadLen = shape::length(tadShapeInfo);
if(shape::elementWiseStride(tadShapeInfo) == 1){
auto func = PRAGMA_THREADS_FOR {
for (auto i = start; i < stop; i += increment) {
T *inBuff = input.bufferAsT<T>() + tadOffsets[i];
T *outBuff = output.bufferAsT<T>() + tadOffsets[i];
2019-06-06 14:21:15 +02:00
T max = -DataTypeUtils::max<T>();
T sum = 0;
2019-06-06 14:21:15 +02:00
for (uint j = 0; j < tadLen; ++j)
max = nd4j::math::nd4j_max<T>(max, inBuff[j]);
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
for (uint j = 0; j < tadLen; ++j) {
T temp = nd4j::math::nd4j_exp<T, T>(inBuff[j] - max);
outBuff[j] = temp;
sum += temp;
}
2019-06-06 14:21:15 +02:00
for (uint j = 0; j < tadLen; ++j)
outBuff[j] /= sum;
2019-06-06 14:21:15 +02:00
}
};
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
samediff::Threads::parallel_tad(func,0, numOfSubArrs);
2019-06-06 14:21:15 +02:00
}
else {
uint inShapeInfoCast[MAX_RANK];
bool canCast = nd4j::DataTypeUtils::castShapeInfo(tadShapeInfo, inShapeInfoCast);
auto offsets = new Nd4jLong[tadLen];
shape::calcOffsets(tadShapeInfo, offsets);
auto func = PRAGMA_THREADS_FOR {
for (auto i = start; i < stop; i += increment) {
auto inBuff = input.bufferAsT<T>() + tadOffsets[i];
auto outBuff = output.bufferAsT<T>() + tadOffsets[i];
2019-06-06 14:21:15 +02:00
T max = -DataTypeUtils::max<T>();
T sum = 0.f;
2019-06-06 14:21:15 +02:00
for (uint j = 0; j < tadLen; ++j)
max = nd4j::math::nd4j_max<T>(max, inBuff[offsets[j]]);
2019-06-06 14:21:15 +02:00
for (uint j = 0; j < tadLen; ++j) {
T temp = nd4j::math::nd4j_exp<T, T>(inBuff[offsets[j]] - max);
outBuff[offsets[j]] = temp;
sum += temp;
}
2019-06-06 14:21:15 +02:00
for (uint j = 0; j < tadLen; ++j)
outBuff[offsets[j]] /= sum;
2019-06-06 14:21:15 +02:00
}
};
samediff::Threads::parallel_tad(func, 0, numOfSubArrs);
2019-06-06 14:21:15 +02:00
delete []offsets;
}
}
else {
Shyrma temp (#131) * - specifying template instantiation for certain types in float16 and bloat16 Signed-off-by: Yurii <iuriish@yahoo.com> * - polishing bfloat16 and float16 member functions template specialization Signed-off-by: Yurii <iuriish@yahoo.com> * - rewrite and overload array +-*/ scalar and scalar +-*/ arr in NDAray class Signed-off-by: Yurii <iuriish@yahoo.com> * - make corrections which have to do with and rvalue lvalue conversions Signed-off-by: Yurii <iuriish@yahoo.com> * - provide move semantic in NDArray operators array +-/* array Signed-off-by: Yurii <iuriish@yahoo.com> * float16/bfloat16 tweaks Signed-off-by: raver119 <raver119@gmail.com> * one more tweak Signed-off-by: raver119 <raver119@gmail.com> * - make float16 and bfloat16 to compile successfully on cuda Signed-off-by: Yurii <iuriish@yahoo.com> * - do not use resources of view-like arrays when move semantics is applied Signed-off-by: Yurii <iuriish@yahoo.com> * - get rid of pointers in signatures NDArray methods 1 Signed-off-by: Yurii <iuriish@yahoo.com> * - correction of signature of NDArray::dup method Signed-off-by: Yurii <iuriish@yahoo.com> * - correction of signature of NDArray::reduceAlongDimension method Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of NDArray::applyIndexReduce and applyTrueBroadcast methods Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of NDArray::applyReduce3 and varianceAlongDimension methods Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of NDArray::tensorsAlongDimension and diagonal methods Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of NDArray::allTensorsAlongDimension Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of NDArray::reduceAlongDimension 2 Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of NDArray::applyTransform 2 Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of NDArray::applyPairwiseTransform 2 Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of NDArray::applyBroadcast 2 Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of NDArray::applyTrueBroadcast 2 Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of NDArray::applyScalar and applyScalarArr Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of NDArray::lambda methods Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of NDArray::reduce3 methods 2 Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of following NDArray methods: add/sub/mul/div row/column and fillAsTriangular Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of NDArray::tileToShape methods Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of NDArray::isShapeSameStrict method Signed-off-by: Yurii <iuriish@yahoo.com> * minor corrections in tests Signed-off-by: Yurii <iuriish@yahoo.com> * - replace reduce op in batchnorm mkldnn Signed-off-by: Yurii <iuriish@yahoo.com> * - add explicit templates instantiations for operator+(NDArray&&. const scalar) Signed-off-by: Yurii <iuriish@yahoo.com> * - corrections of casts in float16/bfloat16 Signed-off-by: Yurii <iuriish@yahoo.com> * - provide move semantics in following NDArray methods: transform, applyTrueBroadcast, transpose, reshape, permute Signed-off-by: Yurii <iuriish@yahoo.com> * - get rid of input array A duplicate in svd cuda op Signed-off-by: Yurii <iuriish@yahoo.com> * - avoid available bug in svd cuda API Signed-off-by: Yurii <iuriish@yahoo.com> * - add temporary global memory buffer in svd cuda when calcUV = false and m != n Signed-off-by: Yurii <iuriish@yahoo.com> * - remove test with blfoat16 type for betainC Signed-off-by: Yurii <iuriish@yahoo.com> * - resolve conflicts after master has been merged in Signed-off-by: Yurii <iuriish@yahoo.com> * - changed type of affected input array in fused_batch_norm Signed-off-by: Yurii <iuriish@yahoo.com> * - add several explicit type castings Signed-off-by: Yurii <iuriish@yahoo.com> * - add ND4J_EXPORT to operators Signed-off-by: Yurii <iuriish@yahoo.com> * - add explicit template types in instantiations of template arithm operators of NDArray class Signed-off-by: Yurii <iuriish@yahoo.com> * - one more test fix Signed-off-by: Yurii <iuriish@yahoo.com> Co-authored-by: raver119 <raver119@gmail.com>
2019-12-20 20:35:39 +01:00
NDArray max = input.reduceAlongDimension(nd4j::reduce::Max, {dimension}, true);
input.applyTrueBroadcast(nd4j::BroadcastOpsTuple::Subtract(), max, output, false);
output.applyTransform(nd4j::transform::Exp, output);
NDArray sum = output.reduceAlongDimension(nd4j::reduce::Sum, {dimension}, true);
2019-06-06 14:21:15 +02:00
output /= sum;
}
}
///////////////////////////////////////////////////////////////////
void softmax(nd4j::LaunchContext * context, const NDArray& input, NDArray& output, const int dimension) {
BUILD_SINGLE_SELECTOR(input.dataType(), softmax_, (context, input, output, dimension), FLOAT_TYPES);
}
[WIP] more CUDA stuff (#57) * initial commit Signed-off-by: raver119 <raver119@gmail.com> * Added gradcheck test for dynamic_partition_bp op. * - implementation of dilation op (cpu and cuda) Signed-off-by: Yurii <yurii@skymind.io> * Fixed broadcast_dynamic_shape 1D case and tests. * Fixed usage of default integer arguments. * Fixed dynamic_partition_bp op and tests. * Eliminated test with grad check for dynamic_partition_bp op. * start working on cuda svd - porting available corresponding api from cuSOLVER library Signed-off-by: Yurii <yurii@skymind.io> * provide prelu_bp Signed-off-by: Yurii <yurii@skymind.io> * - provide gruCell_bp (old version ??) Signed-off-by: Yurii <yurii@skymind.io> * - polishing cumsum_bp and cumprod_bp tests Signed-off-by: Yurii <yurii@skymind.io> * provide sparseSoftmaxCrossEntropyWithLogits and sparseSoftmaxCrossEntropyWithLogits_grad Signed-off-by: Yurii <yurii@skymind.io> * Fixed atomicMul with float input/output * implementation of cuda kernel for triu_bp operation Signed-off-by: Yurii <yurii@skymind.io> * Refactored lup helper to add parrallel computing. * cusolver libraries Signed-off-by: raver119 <raver119@gmail.com> * uncomment cuSolver APIs in svd.cu Signed-off-by: Yurii <yurii@skymind.io> * cusolver var Signed-off-by: raver119 <raver119@gmail.com> * - further work on cuSolver svd Signed-off-by: Yurii <yurii@skymind.io> * Implement usage of cuda solver to LUP decomposition. * - correct naames in lup functions Signed-off-by: Yurii <yurii@skymind.io> * correct svdQR cuda Signed-off-by: Yurii <yurii@skymind.io> * - provide transpositions of input matrices in case of c order in svdCudaQR Signed-off-by: Yurii <yurii@skymind.io> * Fixed implementation issues with LUP usign cuda solver. * Implementation of matrix_determinant helper with cuda kernels. Working revision. * Implemented log_matrix_determinant helper with cuda kernels. * - implementation of batched cuda svd Signed-off-by: Yurii <yurii@skymind.io> * Refactored cholesky helper and implementation of cuda solver cholesky batch. * - implementation of cuda kernel for tile bp Signed-off-by: Yurii <yurii@skymind.io> * Implementation of cholesky and logdet with cuda kernels. * - implementation of cuda kernel for sru_bidirectional Signed-off-by: Yurii <yurii@skymind.io> * Fixed cholesky helper. * Cholesky op helper implementation. Working double-based cublas implementation. * bad import excluded Signed-off-by: raver119 <raver119@gmail.com> * Finished with cuda implementation of cholesky helper and tests. * - implementation of cuda kernel for sru_bidirectional_backprop operation Signed-off-by: Yurii <yurii@skymind.io> * Implementation of matrix_inverse op helper with cuda kernels. The first revision. * - start working on gruCell_bp Signed-off-by: Yurii <yurii@skymind.io> * Implementation of matrix_inverse helper. * - further work on new gruCell_bp Signed-off-by: Yurii <yurii@skymind.io> * cuBLAS related fixes Signed-off-by: raver119 <raver119@gmail.com> * calculateOutputShapes() now passes device buffers as well Signed-off-by: raver119 <raver119@gmail.com> * special concat/average/accumulate init host pointers now Signed-off-by: raver119 <raver119@gmail.com> * few more tweaks Signed-off-by: raver119 <raver119@gmail.com> * additional CudaDataBufferFactory signatures certain for data types Signed-off-by: raver119 <raver119@gmail.com> * cuSolver host buffer Signed-off-by: raver119 <raver119@gmail.com> * buffer to buffer memcpy host ptr allocation Signed-off-by: raver119 <raver119@gmail.com>
2019-07-12 10:51:51 +02:00
//////////////////////////////////////////////////////////////////////////
void prelu(nd4j::LaunchContext * context, const NDArray& input, const NDArray& alpha, NDArray& output) {
const Nd4jLong inputLen = input.lengthOf();
const Nd4jLong* inputShapeInfo = input.getShapeInfo();
const Nd4jLong* alphaShapeInfo = alpha.getShapeInfo();
auto func = PRAGMA_THREADS_FOR {
for (auto i = start; i < stop; i += increment) {
// FIXME: double!
double x = input.e<double>(i);
if (x < 0.0) {
// FIXME: double
output.p(i, (x * alpha.e<double>(shape::subArrayIndex(i, inputShapeInfo, alphaShapeInfo))));
} else
output.p(i, x);
}
};
samediff::Threads::parallel_for(func, 0, inputLen);
[WIP] more CUDA stuff (#57) * initial commit Signed-off-by: raver119 <raver119@gmail.com> * Added gradcheck test for dynamic_partition_bp op. * - implementation of dilation op (cpu and cuda) Signed-off-by: Yurii <yurii@skymind.io> * Fixed broadcast_dynamic_shape 1D case and tests. * Fixed usage of default integer arguments. * Fixed dynamic_partition_bp op and tests. * Eliminated test with grad check for dynamic_partition_bp op. * start working on cuda svd - porting available corresponding api from cuSOLVER library Signed-off-by: Yurii <yurii@skymind.io> * provide prelu_bp Signed-off-by: Yurii <yurii@skymind.io> * - provide gruCell_bp (old version ??) Signed-off-by: Yurii <yurii@skymind.io> * - polishing cumsum_bp and cumprod_bp tests Signed-off-by: Yurii <yurii@skymind.io> * provide sparseSoftmaxCrossEntropyWithLogits and sparseSoftmaxCrossEntropyWithLogits_grad Signed-off-by: Yurii <yurii@skymind.io> * Fixed atomicMul with float input/output * implementation of cuda kernel for triu_bp operation Signed-off-by: Yurii <yurii@skymind.io> * Refactored lup helper to add parrallel computing. * cusolver libraries Signed-off-by: raver119 <raver119@gmail.com> * uncomment cuSolver APIs in svd.cu Signed-off-by: Yurii <yurii@skymind.io> * cusolver var Signed-off-by: raver119 <raver119@gmail.com> * - further work on cuSolver svd Signed-off-by: Yurii <yurii@skymind.io> * Implement usage of cuda solver to LUP decomposition. * - correct naames in lup functions Signed-off-by: Yurii <yurii@skymind.io> * correct svdQR cuda Signed-off-by: Yurii <yurii@skymind.io> * - provide transpositions of input matrices in case of c order in svdCudaQR Signed-off-by: Yurii <yurii@skymind.io> * Fixed implementation issues with LUP usign cuda solver. * Implementation of matrix_determinant helper with cuda kernels. Working revision. * Implemented log_matrix_determinant helper with cuda kernels. * - implementation of batched cuda svd Signed-off-by: Yurii <yurii@skymind.io> * Refactored cholesky helper and implementation of cuda solver cholesky batch. * - implementation of cuda kernel for tile bp Signed-off-by: Yurii <yurii@skymind.io> * Implementation of cholesky and logdet with cuda kernels. * - implementation of cuda kernel for sru_bidirectional Signed-off-by: Yurii <yurii@skymind.io> * Fixed cholesky helper. * Cholesky op helper implementation. Working double-based cublas implementation. * bad import excluded Signed-off-by: raver119 <raver119@gmail.com> * Finished with cuda implementation of cholesky helper and tests. * - implementation of cuda kernel for sru_bidirectional_backprop operation Signed-off-by: Yurii <yurii@skymind.io> * Implementation of matrix_inverse op helper with cuda kernels. The first revision. * - start working on gruCell_bp Signed-off-by: Yurii <yurii@skymind.io> * Implementation of matrix_inverse helper. * - further work on new gruCell_bp Signed-off-by: Yurii <yurii@skymind.io> * cuBLAS related fixes Signed-off-by: raver119 <raver119@gmail.com> * calculateOutputShapes() now passes device buffers as well Signed-off-by: raver119 <raver119@gmail.com> * special concat/average/accumulate init host pointers now Signed-off-by: raver119 <raver119@gmail.com> * few more tweaks Signed-off-by: raver119 <raver119@gmail.com> * additional CudaDataBufferFactory signatures certain for data types Signed-off-by: raver119 <raver119@gmail.com> * cuSolver host buffer Signed-off-by: raver119 <raver119@gmail.com> * buffer to buffer memcpy host ptr allocation Signed-off-by: raver119 <raver119@gmail.com>
2019-07-12 10:51:51 +02:00
}
2019-06-06 14:21:15 +02:00
[WIP] more CUDA stuff (#57) * initial commit Signed-off-by: raver119 <raver119@gmail.com> * Added gradcheck test for dynamic_partition_bp op. * - implementation of dilation op (cpu and cuda) Signed-off-by: Yurii <yurii@skymind.io> * Fixed broadcast_dynamic_shape 1D case and tests. * Fixed usage of default integer arguments. * Fixed dynamic_partition_bp op and tests. * Eliminated test with grad check for dynamic_partition_bp op. * start working on cuda svd - porting available corresponding api from cuSOLVER library Signed-off-by: Yurii <yurii@skymind.io> * provide prelu_bp Signed-off-by: Yurii <yurii@skymind.io> * - provide gruCell_bp (old version ??) Signed-off-by: Yurii <yurii@skymind.io> * - polishing cumsum_bp and cumprod_bp tests Signed-off-by: Yurii <yurii@skymind.io> * provide sparseSoftmaxCrossEntropyWithLogits and sparseSoftmaxCrossEntropyWithLogits_grad Signed-off-by: Yurii <yurii@skymind.io> * Fixed atomicMul with float input/output * implementation of cuda kernel for triu_bp operation Signed-off-by: Yurii <yurii@skymind.io> * Refactored lup helper to add parrallel computing. * cusolver libraries Signed-off-by: raver119 <raver119@gmail.com> * uncomment cuSolver APIs in svd.cu Signed-off-by: Yurii <yurii@skymind.io> * cusolver var Signed-off-by: raver119 <raver119@gmail.com> * - further work on cuSolver svd Signed-off-by: Yurii <yurii@skymind.io> * Implement usage of cuda solver to LUP decomposition. * - correct naames in lup functions Signed-off-by: Yurii <yurii@skymind.io> * correct svdQR cuda Signed-off-by: Yurii <yurii@skymind.io> * - provide transpositions of input matrices in case of c order in svdCudaQR Signed-off-by: Yurii <yurii@skymind.io> * Fixed implementation issues with LUP usign cuda solver. * Implementation of matrix_determinant helper with cuda kernels. Working revision. * Implemented log_matrix_determinant helper with cuda kernels. * - implementation of batched cuda svd Signed-off-by: Yurii <yurii@skymind.io> * Refactored cholesky helper and implementation of cuda solver cholesky batch. * - implementation of cuda kernel for tile bp Signed-off-by: Yurii <yurii@skymind.io> * Implementation of cholesky and logdet with cuda kernels. * - implementation of cuda kernel for sru_bidirectional Signed-off-by: Yurii <yurii@skymind.io> * Fixed cholesky helper. * Cholesky op helper implementation. Working double-based cublas implementation. * bad import excluded Signed-off-by: raver119 <raver119@gmail.com> * Finished with cuda implementation of cholesky helper and tests. * - implementation of cuda kernel for sru_bidirectional_backprop operation Signed-off-by: Yurii <yurii@skymind.io> * Implementation of matrix_inverse op helper with cuda kernels. The first revision. * - start working on gruCell_bp Signed-off-by: Yurii <yurii@skymind.io> * Implementation of matrix_inverse helper. * - further work on new gruCell_bp Signed-off-by: Yurii <yurii@skymind.io> * cuBLAS related fixes Signed-off-by: raver119 <raver119@gmail.com> * calculateOutputShapes() now passes device buffers as well Signed-off-by: raver119 <raver119@gmail.com> * special concat/average/accumulate init host pointers now Signed-off-by: raver119 <raver119@gmail.com> * few more tweaks Signed-off-by: raver119 <raver119@gmail.com> * additional CudaDataBufferFactory signatures certain for data types Signed-off-by: raver119 <raver119@gmail.com> * cuSolver host buffer Signed-off-by: raver119 <raver119@gmail.com> * buffer to buffer memcpy host ptr allocation Signed-off-by: raver119 <raver119@gmail.com>
2019-07-12 10:51:51 +02:00
//////////////////////////////////////////////////////////////////////////
void preluBP(nd4j::LaunchContext * context, const NDArray& input, const NDArray& alpha, const NDArray& dLdO, NDArray& dLdI, NDArray& dLdA) {
const Nd4jLong inputLen = input.lengthOf();
const Nd4jLong* inputShapeInfo = input.getShapeInfo();
const Nd4jLong* alphaShapeInfo = alpha.getShapeInfo();
dLdA.assign(0.0f);
for(Nd4jLong i = 0; i < inputLen; ++i) {
// FIXME: double
double x = input.e<double>(i);
double grO = dLdO.e<double>(i);
if(x < 0.0) {
Nd4jLong alphaInd = shape::subArrayIndex(i, inputShapeInfo, alphaShapeInfo);
dLdI.p(i, grO * alpha.e<double>(alphaInd));
double prevVal = dLdA.e<double>(alphaInd);
prevVal += (grO * x);
dLdA.p(alphaInd, prevVal);
}
else
dLdI.p(i, grO);
2019-06-06 14:21:15 +02:00
}
}
bool checkAlphaShapeLen(std::vector<Nd4jLong> const& expectedShape, Nd4jLong shapeLen) {
Nd4jLong expectedAlphaLen = std::accumulate(expectedShape.cbegin(), expectedShape.cend(), 1, std::multiplies<Nd4jLong>());
return expectedAlphaLen == shapeLen;
}
template <typename T>
static void thresholdRelu_(NDArray const& input, double threshold, NDArray& output) {
auto routine = LAMBDA_T(_x, threshold) {
return _x > (T)threshold? _x: (T)0.f;
};
Shyrma temp (#131) * - specifying template instantiation for certain types in float16 and bloat16 Signed-off-by: Yurii <iuriish@yahoo.com> * - polishing bfloat16 and float16 member functions template specialization Signed-off-by: Yurii <iuriish@yahoo.com> * - rewrite and overload array +-*/ scalar and scalar +-*/ arr in NDAray class Signed-off-by: Yurii <iuriish@yahoo.com> * - make corrections which have to do with and rvalue lvalue conversions Signed-off-by: Yurii <iuriish@yahoo.com> * - provide move semantic in NDArray operators array +-/* array Signed-off-by: Yurii <iuriish@yahoo.com> * float16/bfloat16 tweaks Signed-off-by: raver119 <raver119@gmail.com> * one more tweak Signed-off-by: raver119 <raver119@gmail.com> * - make float16 and bfloat16 to compile successfully on cuda Signed-off-by: Yurii <iuriish@yahoo.com> * - do not use resources of view-like arrays when move semantics is applied Signed-off-by: Yurii <iuriish@yahoo.com> * - get rid of pointers in signatures NDArray methods 1 Signed-off-by: Yurii <iuriish@yahoo.com> * - correction of signature of NDArray::dup method Signed-off-by: Yurii <iuriish@yahoo.com> * - correction of signature of NDArray::reduceAlongDimension method Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of NDArray::applyIndexReduce and applyTrueBroadcast methods Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of NDArray::applyReduce3 and varianceAlongDimension methods Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of NDArray::tensorsAlongDimension and diagonal methods Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of NDArray::allTensorsAlongDimension Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of NDArray::reduceAlongDimension 2 Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of NDArray::applyTransform 2 Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of NDArray::applyPairwiseTransform 2 Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of NDArray::applyBroadcast 2 Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of NDArray::applyTrueBroadcast 2 Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of NDArray::applyScalar and applyScalarArr Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of NDArray::lambda methods Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of NDArray::reduce3 methods 2 Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of following NDArray methods: add/sub/mul/div row/column and fillAsTriangular Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of NDArray::tileToShape methods Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of NDArray::isShapeSameStrict method Signed-off-by: Yurii <iuriish@yahoo.com> * minor corrections in tests Signed-off-by: Yurii <iuriish@yahoo.com> * - replace reduce op in batchnorm mkldnn Signed-off-by: Yurii <iuriish@yahoo.com> * - add explicit templates instantiations for operator+(NDArray&&. const scalar) Signed-off-by: Yurii <iuriish@yahoo.com> * - corrections of casts in float16/bfloat16 Signed-off-by: Yurii <iuriish@yahoo.com> * - provide move semantics in following NDArray methods: transform, applyTrueBroadcast, transpose, reshape, permute Signed-off-by: Yurii <iuriish@yahoo.com> * - get rid of input array A duplicate in svd cuda op Signed-off-by: Yurii <iuriish@yahoo.com> * - avoid available bug in svd cuda API Signed-off-by: Yurii <iuriish@yahoo.com> * - add temporary global memory buffer in svd cuda when calcUV = false and m != n Signed-off-by: Yurii <iuriish@yahoo.com> * - remove test with blfoat16 type for betainC Signed-off-by: Yurii <iuriish@yahoo.com> * - resolve conflicts after master has been merged in Signed-off-by: Yurii <iuriish@yahoo.com> * - changed type of affected input array in fused_batch_norm Signed-off-by: Yurii <iuriish@yahoo.com> * - add several explicit type castings Signed-off-by: Yurii <iuriish@yahoo.com> * - add ND4J_EXPORT to operators Signed-off-by: Yurii <iuriish@yahoo.com> * - add explicit template types in instantiations of template arithm operators of NDArray class Signed-off-by: Yurii <iuriish@yahoo.com> * - one more test fix Signed-off-by: Yurii <iuriish@yahoo.com> Co-authored-by: raver119 <raver119@gmail.com>
2019-12-20 20:35:39 +01:00
const_cast<NDArray&>(input).applyLambda<T>(routine, output);
2019-06-06 14:21:15 +02:00
}
void thresholdRelu(nd4j::LaunchContext * context, NDArray const& input, double threshold, NDArray& output) {
BUILD_SINGLE_SELECTOR(input.dataType(), thresholdRelu_, (input, threshold, output), FLOAT_TYPES);
}
template <typename T>
static void thresholdReluDerivative_(nd4j::LaunchContext * context, NDArray* input, double theta, NDArray* dLdO, NDArray* output) {
auto derivative = LAMBDA_TT(_x, grO, theta) {if (_x > theta) return grO; else return static_cast<T>(0); };
Shyrma temp (#131) * - specifying template instantiation for certain types in float16 and bloat16 Signed-off-by: Yurii <iuriish@yahoo.com> * - polishing bfloat16 and float16 member functions template specialization Signed-off-by: Yurii <iuriish@yahoo.com> * - rewrite and overload array +-*/ scalar and scalar +-*/ arr in NDAray class Signed-off-by: Yurii <iuriish@yahoo.com> * - make corrections which have to do with and rvalue lvalue conversions Signed-off-by: Yurii <iuriish@yahoo.com> * - provide move semantic in NDArray operators array +-/* array Signed-off-by: Yurii <iuriish@yahoo.com> * float16/bfloat16 tweaks Signed-off-by: raver119 <raver119@gmail.com> * one more tweak Signed-off-by: raver119 <raver119@gmail.com> * - make float16 and bfloat16 to compile successfully on cuda Signed-off-by: Yurii <iuriish@yahoo.com> * - do not use resources of view-like arrays when move semantics is applied Signed-off-by: Yurii <iuriish@yahoo.com> * - get rid of pointers in signatures NDArray methods 1 Signed-off-by: Yurii <iuriish@yahoo.com> * - correction of signature of NDArray::dup method Signed-off-by: Yurii <iuriish@yahoo.com> * - correction of signature of NDArray::reduceAlongDimension method Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of NDArray::applyIndexReduce and applyTrueBroadcast methods Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of NDArray::applyReduce3 and varianceAlongDimension methods Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of NDArray::tensorsAlongDimension and diagonal methods Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of NDArray::allTensorsAlongDimension Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of NDArray::reduceAlongDimension 2 Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of NDArray::applyTransform 2 Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of NDArray::applyPairwiseTransform 2 Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of NDArray::applyBroadcast 2 Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of NDArray::applyTrueBroadcast 2 Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of NDArray::applyScalar and applyScalarArr Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of NDArray::lambda methods Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of NDArray::reduce3 methods 2 Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of following NDArray methods: add/sub/mul/div row/column and fillAsTriangular Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of NDArray::tileToShape methods Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of NDArray::isShapeSameStrict method Signed-off-by: Yurii <iuriish@yahoo.com> * minor corrections in tests Signed-off-by: Yurii <iuriish@yahoo.com> * - replace reduce op in batchnorm mkldnn Signed-off-by: Yurii <iuriish@yahoo.com> * - add explicit templates instantiations for operator+(NDArray&&. const scalar) Signed-off-by: Yurii <iuriish@yahoo.com> * - corrections of casts in float16/bfloat16 Signed-off-by: Yurii <iuriish@yahoo.com> * - provide move semantics in following NDArray methods: transform, applyTrueBroadcast, transpose, reshape, permute Signed-off-by: Yurii <iuriish@yahoo.com> * - get rid of input array A duplicate in svd cuda op Signed-off-by: Yurii <iuriish@yahoo.com> * - avoid available bug in svd cuda API Signed-off-by: Yurii <iuriish@yahoo.com> * - add temporary global memory buffer in svd cuda when calcUV = false and m != n Signed-off-by: Yurii <iuriish@yahoo.com> * - remove test with blfoat16 type for betainC Signed-off-by: Yurii <iuriish@yahoo.com> * - resolve conflicts after master has been merged in Signed-off-by: Yurii <iuriish@yahoo.com> * - changed type of affected input array in fused_batch_norm Signed-off-by: Yurii <iuriish@yahoo.com> * - add several explicit type castings Signed-off-by: Yurii <iuriish@yahoo.com> * - add ND4J_EXPORT to operators Signed-off-by: Yurii <iuriish@yahoo.com> * - add explicit template types in instantiations of template arithm operators of NDArray class Signed-off-by: Yurii <iuriish@yahoo.com> * - one more test fix Signed-off-by: Yurii <iuriish@yahoo.com> Co-authored-by: raver119 <raver119@gmail.com>
2019-12-20 20:35:39 +01:00
input->applyPairwiseLambda<T>(*dLdO, derivative, *output);
2019-06-06 14:21:15 +02:00
}
void thresholdReluDerivative(nd4j::LaunchContext * context, NDArray* input, double threshold, NDArray* dLdO, NDArray* output) {
BUILD_SINGLE_SELECTOR(input->dataType(), thresholdReluDerivative_, (context, input, threshold, dLdO, output), FLOAT_TYPES);
}
///////////////////////////////////////////////////////////////////
void logSoftmax(nd4j::LaunchContext * context, const NDArray& input, NDArray& output, const int dimension) {
const int rank = input.rankOf();
if(input.isVector()) {
if(rank == 1 || input.sizeAt(dimension) != 1) {
BUILD_SINGLE_SELECTOR(input.dataType(), logSoftMaxForVector_, (input.getBuffer(), input.getShapeInfo(), output.buffer(), output.shapeInfo()), FLOAT_TYPES);
}
else
output = 0.;
}
else {
Shyrma temp (#131) * - specifying template instantiation for certain types in float16 and bloat16 Signed-off-by: Yurii <iuriish@yahoo.com> * - polishing bfloat16 and float16 member functions template specialization Signed-off-by: Yurii <iuriish@yahoo.com> * - rewrite and overload array +-*/ scalar and scalar +-*/ arr in NDAray class Signed-off-by: Yurii <iuriish@yahoo.com> * - make corrections which have to do with and rvalue lvalue conversions Signed-off-by: Yurii <iuriish@yahoo.com> * - provide move semantic in NDArray operators array +-/* array Signed-off-by: Yurii <iuriish@yahoo.com> * float16/bfloat16 tweaks Signed-off-by: raver119 <raver119@gmail.com> * one more tweak Signed-off-by: raver119 <raver119@gmail.com> * - make float16 and bfloat16 to compile successfully on cuda Signed-off-by: Yurii <iuriish@yahoo.com> * - do not use resources of view-like arrays when move semantics is applied Signed-off-by: Yurii <iuriish@yahoo.com> * - get rid of pointers in signatures NDArray methods 1 Signed-off-by: Yurii <iuriish@yahoo.com> * - correction of signature of NDArray::dup method Signed-off-by: Yurii <iuriish@yahoo.com> * - correction of signature of NDArray::reduceAlongDimension method Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of NDArray::applyIndexReduce and applyTrueBroadcast methods Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of NDArray::applyReduce3 and varianceAlongDimension methods Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of NDArray::tensorsAlongDimension and diagonal methods Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of NDArray::allTensorsAlongDimension Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of NDArray::reduceAlongDimension 2 Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of NDArray::applyTransform 2 Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of NDArray::applyPairwiseTransform 2 Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of NDArray::applyBroadcast 2 Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of NDArray::applyTrueBroadcast 2 Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of NDArray::applyScalar and applyScalarArr Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of NDArray::lambda methods Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of NDArray::reduce3 methods 2 Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of following NDArray methods: add/sub/mul/div row/column and fillAsTriangular Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of NDArray::tileToShape methods Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of NDArray::isShapeSameStrict method Signed-off-by: Yurii <iuriish@yahoo.com> * minor corrections in tests Signed-off-by: Yurii <iuriish@yahoo.com> * - replace reduce op in batchnorm mkldnn Signed-off-by: Yurii <iuriish@yahoo.com> * - add explicit templates instantiations for operator+(NDArray&&. const scalar) Signed-off-by: Yurii <iuriish@yahoo.com> * - corrections of casts in float16/bfloat16 Signed-off-by: Yurii <iuriish@yahoo.com> * - provide move semantics in following NDArray methods: transform, applyTrueBroadcast, transpose, reshape, permute Signed-off-by: Yurii <iuriish@yahoo.com> * - get rid of input array A duplicate in svd cuda op Signed-off-by: Yurii <iuriish@yahoo.com> * - avoid available bug in svd cuda API Signed-off-by: Yurii <iuriish@yahoo.com> * - add temporary global memory buffer in svd cuda when calcUV = false and m != n Signed-off-by: Yurii <iuriish@yahoo.com> * - remove test with blfoat16 type for betainC Signed-off-by: Yurii <iuriish@yahoo.com> * - resolve conflicts after master has been merged in Signed-off-by: Yurii <iuriish@yahoo.com> * - changed type of affected input array in fused_batch_norm Signed-off-by: Yurii <iuriish@yahoo.com> * - add several explicit type castings Signed-off-by: Yurii <iuriish@yahoo.com> * - add ND4J_EXPORT to operators Signed-off-by: Yurii <iuriish@yahoo.com> * - add explicit template types in instantiations of template arithm operators of NDArray class Signed-off-by: Yurii <iuriish@yahoo.com> * - one more test fix Signed-off-by: Yurii <iuriish@yahoo.com> Co-authored-by: raver119 <raver119@gmail.com>
2019-12-20 20:35:39 +01:00
auto maxAlongDim = const_cast<NDArray&>(input).reduceAlongDimension(reduce::Max, {dimension}, true);
(input - maxAlongDim).applyTransform(transform::Exp, output); // output contains exponents temporarily
auto sumAlongDim = output.reduceAlongDimension(reduce::Sum, {dimension}, true);
2019-06-06 14:21:15 +02:00
output /= sumAlongDim;
Shyrma temp (#131) * - specifying template instantiation for certain types in float16 and bloat16 Signed-off-by: Yurii <iuriish@yahoo.com> * - polishing bfloat16 and float16 member functions template specialization Signed-off-by: Yurii <iuriish@yahoo.com> * - rewrite and overload array +-*/ scalar and scalar +-*/ arr in NDAray class Signed-off-by: Yurii <iuriish@yahoo.com> * - make corrections which have to do with and rvalue lvalue conversions Signed-off-by: Yurii <iuriish@yahoo.com> * - provide move semantic in NDArray operators array +-/* array Signed-off-by: Yurii <iuriish@yahoo.com> * float16/bfloat16 tweaks Signed-off-by: raver119 <raver119@gmail.com> * one more tweak Signed-off-by: raver119 <raver119@gmail.com> * - make float16 and bfloat16 to compile successfully on cuda Signed-off-by: Yurii <iuriish@yahoo.com> * - do not use resources of view-like arrays when move semantics is applied Signed-off-by: Yurii <iuriish@yahoo.com> * - get rid of pointers in signatures NDArray methods 1 Signed-off-by: Yurii <iuriish@yahoo.com> * - correction of signature of NDArray::dup method Signed-off-by: Yurii <iuriish@yahoo.com> * - correction of signature of NDArray::reduceAlongDimension method Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of NDArray::applyIndexReduce and applyTrueBroadcast methods Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of NDArray::applyReduce3 and varianceAlongDimension methods Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of NDArray::tensorsAlongDimension and diagonal methods Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of NDArray::allTensorsAlongDimension Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of NDArray::reduceAlongDimension 2 Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of NDArray::applyTransform 2 Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of NDArray::applyPairwiseTransform 2 Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of NDArray::applyBroadcast 2 Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of NDArray::applyTrueBroadcast 2 Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of NDArray::applyScalar and applyScalarArr Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of NDArray::lambda methods Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of NDArray::reduce3 methods 2 Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of following NDArray methods: add/sub/mul/div row/column and fillAsTriangular Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of NDArray::tileToShape methods Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of NDArray::isShapeSameStrict method Signed-off-by: Yurii <iuriish@yahoo.com> * minor corrections in tests Signed-off-by: Yurii <iuriish@yahoo.com> * - replace reduce op in batchnorm mkldnn Signed-off-by: Yurii <iuriish@yahoo.com> * - add explicit templates instantiations for operator+(NDArray&&. const scalar) Signed-off-by: Yurii <iuriish@yahoo.com> * - corrections of casts in float16/bfloat16 Signed-off-by: Yurii <iuriish@yahoo.com> * - provide move semantics in following NDArray methods: transform, applyTrueBroadcast, transpose, reshape, permute Signed-off-by: Yurii <iuriish@yahoo.com> * - get rid of input array A duplicate in svd cuda op Signed-off-by: Yurii <iuriish@yahoo.com> * - avoid available bug in svd cuda API Signed-off-by: Yurii <iuriish@yahoo.com> * - add temporary global memory buffer in svd cuda when calcUV = false and m != n Signed-off-by: Yurii <iuriish@yahoo.com> * - remove test with blfoat16 type for betainC Signed-off-by: Yurii <iuriish@yahoo.com> * - resolve conflicts after master has been merged in Signed-off-by: Yurii <iuriish@yahoo.com> * - changed type of affected input array in fused_batch_norm Signed-off-by: Yurii <iuriish@yahoo.com> * - add several explicit type castings Signed-off-by: Yurii <iuriish@yahoo.com> * - add ND4J_EXPORT to operators Signed-off-by: Yurii <iuriish@yahoo.com> * - add explicit template types in instantiations of template arithm operators of NDArray class Signed-off-by: Yurii <iuriish@yahoo.com> * - one more test fix Signed-off-by: Yurii <iuriish@yahoo.com> Co-authored-by: raver119 <raver119@gmail.com>
2019-12-20 20:35:39 +01:00
output.applyTransform(transform::Log, output);
2019-06-06 14:21:15 +02:00
}
}
BUILD_SINGLE_TEMPLATE(template void thresholdReluDerivative_, (nd4j::LaunchContext * context, NDArray* input, double threshold, NDArray* dLdO, NDArray* output), FLOAT_TYPES);
BUILD_SINGLE_TEMPLATE(template void softmax_, (nd4j::LaunchContext * context, const NDArray& input, NDArray& output, const int dimension), FLOAT_TYPES);
BUILD_SINGLE_TEMPLATE(template void logSoftMaxForVector_, (void *input, Nd4jLong *inShapeInfo, void *output, Nd4jLong *outShapeInfo), FLOAT_TYPES);
BUILD_SINGLE_TEMPLATE(template void _softMaxDerivForVector, (nd4j::LaunchContext * context, const void *input, const Nd4jLong *inShapeInfo, void *output), FLOAT_TYPES);
}
}
}