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

291 lines
12 KiB
C++
Raw Normal View History

/*******************************************************************************
* Copyright (c) 2019 Konduit K.K.
*
* 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 sgazeos@gmail.com
//
#include <ops/declarable/helpers/random.h>
//#include <vector>
#include <memory>
//#include <graph/Context.h>
#include <helpers/ShapeUtils.h>
#include <helpers/RandomLauncher.h>
Oleh multinomial (#163) * libnd4j: Multinomial op #8570 first raw step of multinomial random data generator implementation Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j: Multinomial op #8570 next step of multinomial random categories generator implementation on both cpu and cuda, need corrections and code clean up before review and testing * libnd4j: Multinomial op #8570 code clean up and fixed issues data selecting, moved from coords to tads * libnd4j: Multinomial op #8570 fixed cuda build add reference for math materials that was used for implementation * libnd4j: Multinomial op #8570 fixed several bugs, added several tests and improved cuda version. current implementation works, need testing of reproduction with the same seed * libnd4j: Multinomial op #8570 fixes and optimization after discussion in both cuda and cpu * libnd4j: Multinomial op #8570 add corrections after review, removed tads, replace 2D parallel loop by 3D Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j: Multinomial op fixed declaration and add tests need discussion * libnd4j: Multinomial op fix in test * libnd4j: Multinomial op corrected behavior to get reproducible results, fixed issue in uniform value getting, tests added, need cuda review and cuda testing Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j: Multinomial op fixed indexing on uniform calculation Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j: Multinomial op some corrections in max min declaration Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j: Multinomial op fixed index calculation, added rewind, corrected input declaration, added stats tests, both cuda and cpu. cuda need testing * libnd4j: Multinomial op fixed bugs on cuda nad cpu. need review Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j: Multinomial op corrected tests to handle different orders Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j: Multinomial op some improvements after code review Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j: Multinomial op more corrections after review Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j: Multinomial op fixed seed usage, update tests, fixed cuda based on comments, fixed bug of rewind, removed one behavior, minor corrections. Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j: Multinomial op minor corrections Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j: Multinomial op rise the bound of fluctuation for random cases Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j: Multinomial op modified operation inputs and update implementation and tests on both cpu and cuda * libnd4j: Multinomial op corrected data types according ops.proto Co-authored-by: raver119 <raver119@gmail.com>
2020-01-06 20:35:05 +01:00
#include <execution/Threads.h>
#include <helpers/ConstantTadHelper.h>
namespace sd {
namespace ops {
namespace helpers {
2020-06-08 12:14:22 +02:00
/**
* gammaLess - compute gamma distributed value for shapes (alpha) from 0 to 1
* @tparam T - any float types are acceptable
* @param rng - random generator for uniformly vals
* @param alpha - shape of distribution
* @param beta - scale of distributed values
* @return gamma distributed value
*/
template <typename T>
T gammaLess(graph::RandomGenerator& rng, T const alpha, T const beta) {
auto d = T(1.0334f) - T(0.0766f) * math::p_exp(T(2.2942f) * alpha);
auto a = math::p_pow(T(2.f), alpha) * math::p_pow(T(1.f) - math::p_exp(-d * T(0.5f)), alpha);
auto b = alpha * math::p_pow(d, alpha - T(1.f)) * exp(-d);
auto c = a + b;
T rawX;
static auto index = 0LL;
const T underAlpha = T(1.f) / alpha;
const T powerAlpha = math::p_pow(T(2.f), alpha - T(1.f));
for (;;) {
auto u = rng.relativeT<T>(index++, T(0.f), T(1.f));
if (u <= a / c) rawX = -T(2.f) * math::p_log(T(1.f) - T(0.5f) * math::p_pow(T(c * u), underAlpha));
else rawX = - math::p_log(c * (T(1.f) - u)/(alpha * math::p_pow(d, alpha - T(1.f))));
T v = rng.relativeT(index++, 0.f, 1.f);
if (rawX <= d) {
auto testVal = (math::p_pow(rawX, alpha - 1.f) * math::p_exp(-T(0.5f) * rawX)) / (powerAlpha * math::p_pow(T(1.f) - math::p_exp(-T(0.5f) * rawX), alpha - T(1.f)));
if (testVal < v) continue;
break;
}
else {
if (v <= math::p_pow(d / rawX, T(1.f) - alpha)) break;
continue;
}
}
return rawX / beta;
}
/**
* gammaGreat - generate gamma distributed value for shape (alpha) greater then 1
* @tparam T - given type (any float type is accepted.)
* @param rng - random generator
* @param alpha - shape of the gamma distribution (alpha)
* @param beta - scale of the gamma distribution (beta)
* @return - gamma distributed value with given params
*/
template <typename T>
T gammaGreat(graph::RandomGenerator& rng, T const alpha, T const beta) {
auto decreasedAlpha = alpha - T(1.f/3.f);
auto c = T(1.)/ math::p_sqrt(T(9.f) * decreasedAlpha);
static auto index = 0LL;
T x;
auto normalDistributed = [](graph::RandomGenerator& rng, Nd4jLong& index) {
auto v1 = rng.relativeT(index++, T(0.f), T(1.f));
auto v2 = rng.relativeT(index++, T(0.f), T(1.f));
return math::p_cos(T(2.f * 3.141592f) * v2) * math::p_sqrt(T(-2.f) * math::p_log(v1));
};
// const T underAlpha = T(1.f) / alpha;
// const T powerAlpha = math::p_pow(T(2.f), alpha - T(1.f));
float normalizedVar;
for(;;) {
do {
x = normalDistributed(rng, index); //printf("X = %f\n", x);
normalizedVar = T(1.f) + c * x;
} while(normalizedVar < T(0.f));
normalizedVar = normalizedVar * normalizedVar * normalizedVar; //v * v * v;
auto u = rng.relativeT<T>(index++, T(0.f), T(1.f)); //printf("UNI = %f\n", u);
if( u < T(1.f) - T(.0331f) * (x * x) * (x * x) )
break; //return (d * v / b);
if( log(u) < 0.5f * x * x + decreasedAlpha * (1. - normalizedVar + math::p_log(normalizedVar)) )
break;
}
return (decreasedAlpha * normalizedVar / beta);
}
template <typename T>
void fillRandomGamma_(LaunchContext* context, graph::RandomGenerator& rng, NDArray* alpha, NDArray* beta, NDArray* output) {
Legacy API changes (#441) * initial commit Signed-off-by: raver119@gmail.com <raver119@gmail.com> * another initial commit Signed-off-by: raver119@gmail.com <raver119@gmail.com> * another initial commit Signed-off-by: raver119@gmail.com <raver119@gmail.com> * one more initial commit Signed-off-by: raver119@gmail.com <raver119@gmail.com> * next step Signed-off-by: raver119@gmail.com <raver119@gmail.com> * next step Signed-off-by: raver119@gmail.com <raver119@gmail.com> * next step Signed-off-by: raver119@gmail.com <raver119@gmail.com> * next step Signed-off-by: raver119@gmail.com <raver119@gmail.com> * Refactored buffer() and shapeInfo() methods usage with NDArray class. Signed-off-by: shugeo <sgazeos@gmail.com> * Adopt Graph class methods to use const shapes. Signed-off-by: shugeo <sgazeos@gmail.com> * Adopt choose op to use constant shapes. Signed-off-by: shugeo <sgazeos@gmail.com> * Adopt where op shape method to use constant shapes. Signed-off-by: shugeo <sgazeos@gmail.com> * Adopt lstsq op to use constant empty shapes. Signed-off-by: shugeo <sgazeos@gmail.com> * Adopt matrix_diag_part op shape routine to use constant shapes. Signed-off-by: shugeo <sgazeos@gmail.com> * Adopt determinant ops to use constant shapes. Signed-off-by: shugeo <sgazeos@gmail.com> * Adopt mean_pairwssqerr_loss ops to use constant shapes. Signed-off-by: shugeo <sgazeos@gmail.com> * Adopt ops shape methods. Signed-off-by: shugeo <sgazeos@gmail.com> * Adopt shape methods for loss ops. Signed-off-by: shugeo <sgazeos@gmail.com> * Adopt log_loss op shape method. Signed-off-by: shugeo <sgazeos@gmail.com> * Adopt shape methods for ops. Signed-off-by: shugeo <sgazeos@gmail.com> * Adopt dilation2d ops shape methods. Signed-off-by: shugeo <sgazeos@gmail.com> * Adopted deconv2d ops shape methods. Signed-off-by: shugeo <sgazeos@gmail.com> * Adopted dynamicRNN op shape method. Signed-off-by: shugeo <sgazeos@gmail.com> * Adopted shape methods for ops. Signed-off-by: shugeo <sgazeos@gmail.com> * Adopted shape methods for lstm layer ops. Signed-off-by: shugeo <sgazeos@gmail.com> * few updates Signed-off-by: raver119@gmail.com <raver119@gmail.com> * first cuda tweak Signed-off-by: raver119@gmail.com <raver119@gmail.com> * Adopt constant shapes for sconv2d ops. Signed-off-by: shugeo <sgazeos@gmail.com> * Adopt constant shapes for gru ops. Signed-off-by: shugeo <sgazeos@gmail.com> * Adopt constant shapes with shape methods for segment ops and so on. Signed-off-by: shugeo <sgazeos@gmail.com> * Adopted constant shapes with unsorted_segment_* ops. Signed-off-by: shugeo <sgazeos@gmail.com> * Adopted constant shapes with gamma op shape method. Signed-off-by: shugeo <sgazeos@gmail.com> * Adopted shape methods of reduce_stddev ops. Signed-off-by: shugeo <sgazeos@gmail.com> * Adopted shape methods for reduce_* ops. Signed-off-by: shugeo <sgazeos@gmail.com> * Adopt shape method for squeeze op. Signed-off-by: shugeo <sgazeos@gmail.com> * Adopt strided_slice shape method. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored concat op shape method to adopt constant shapes. Signed-off-by: shugeo <sgazeos@gmail.com> * Adopted shape method for mirror_pad op. Signed-off-by: shugeo <sgazeos@gmail.com> * Adopted split op shape method. Signed-off-by: shugeo <sgazeos@gmail.com> * Adopted tile ops shape methods. Signed-off-by: shugeo <sgazeos@gmail.com> * Added const cast for mkldnn routines handles. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored logSoftMaxForVector_ routine to conform with proper data and shape pointer casts. Signed-off-by: shugeo <sgazeos@gmail.com> * Cosmetic changes to proper usage of constant pointers. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored a couple shape comparators for strides and addBias helpers to proper use data pointers with inplace option. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored depthToSpace helpers. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored histogram helpers. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored im2col helpers. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored gather and gatherND helpers. Signed-off-by: shugeo <sgazeos@gmail.com> * Fixed buffer usage on percentile helper. Signed-off-by: shugeo <sgazeos@gmail.com> * Fixed gather shape with helpers and range buffer usage. Signed-off-by: shugeo <sgazeos@gmail.com> * Fixed buffer usage with space to depth helpers. Signed-off-by: shugeo <sgazeos@gmail.com> * Fixed buffer usage and constant shapes. Signed-off-by: shugeo <sgazeos@gmail.com> * Fixed buffer usage with LUP decomposition> Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored onehot_ helper. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored pad and prefix to use constant shapes. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactoed softmax helpers. Signed-off-by: shugeo <sgazeos@gmail.com> * Fixed space to batch helpers to use buffers properly. Signed-off-by: shugeo <sgazeos@gmail.com> * Fixed stack and split helpers. Signed-off-by: shugeo <sgazeos@gmail.com> * Fixed buffer usage with sparse to dense helpers. Signed-off-by: shugeo <sgazeos@gmail.com> * Fixed buffer usage with mindistance_ helpers. Signed-off-by: shugeo <sgazeos@gmail.com> * Fixed buffer usage with tile helper. Signed-off-by: shugeo <sgazeos@gmail.com> * Fixed constant shape usage. Signed-off-by: shugeo <sgazeos@gmail.com> * Fixed constant shape usage with legacy pairwise bool ops. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored a couple of methods to adopt constant shape usage. Signed-off-by: shugeo <sgazeos@gmail.com> * Fixed broadcasting with constant shape." Signed-off-by: shugeo <sgazeos@gmail.com> * Fixed const usage with inplace reverse and constant shapes with legacy reduction. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored legacy ops with const shapes. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored sort to adopt constant shapes. Signed-off-by: shugeo <sgazeos@gmail.com> * Corrected sort for constant shape usage. Signed-off-by: shugeo <sgazeos@gmail.com> * Fixed constant shape usage with special methods. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored Context to conform with constant shape usage. Signed-off-by: shugeo <sgazeos@gmail.com> * CUDA broadcasting headers Signed-off-by: raver119@gmail.com <raver119@gmail.com> * pairwise/indexreduce/random headers Signed-off-by: raver119@gmail.com <raver119@gmail.com> * Refactored native ops to adopt constant shapes. Signed-off-by: shugeo <sgazeos@gmail.com> * legacy reduce3/scalar headers Signed-off-by: raver119@gmail.com <raver119@gmail.com> * Corrected pullRow signature and tests. Signed-off-by: shugeo <sgazeos@gmail.com> * Corrected routines to proper use of constant shapes. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored tests to use constant shapes properly. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored legacy ops tests to use constant shapes properly. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored buffer usage with NDArray tests. Signed-off-by: shugeo <sgazeos@gmail.com> * Fixed native ops tests. Signed-off-by: shugeo <sgazeos@gmail.com> * Fixed special concat routine. Signed-off-by: shugeo <sgazeos@gmail.com> * Fixed buffer usage with test. Signed-off-by: shugeo <sgazeos@gmail.com> * Fixed buffer usage with a test. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored TAD.h and tests. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored calcStrides* routines to use constant shapes. Signed-off-by: shugeo <sgazeos@gmail.com> * Fixed miscelaneous errors with constant shapes. Signed-off-by: shugeo <sgazeos@gmail.com> * NativeOps const changes Signed-off-by: raver119@gmail.com <raver119@gmail.com> * Corrected definitions for declared functions. Signed-off-by: shugeo <sgazeos@gmail.com> * NativeOps const changes Signed-off-by: raver119@gmail.com <raver119@gmail.com> * few more const changes Signed-off-by: raver119@gmail.com <raver119@gmail.com> * Fixed const shapes with shape routines. Signed-off-by: shugeo <sgazeos@gmail.com> * few more const changes Signed-off-by: raver119@gmail.com <raver119@gmail.com> * Fixed shape method for broadcastable case. Signed-off-by: shugeo <sgazeos@gmail.com> * few more const changes Signed-off-by: raver119@gmail.com <raver119@gmail.com> * xw_plus_b BP shape fn restored Signed-off-by: raver119@gmail.com <raver119@gmail.com> * Fixed signatures with broadcasting. Signed-off-by: shugeo <sgazeos@gmail.com> * Repaired backprops shape methods for a set of operations. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored broadcast bool for cuda. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored methods for 3 args with const qualifier. Signed-off-by: shugeo <sgazeos@gmail.com> * Fixed a couple of kernel signatures for broadcasting. Signed-off-by: shugeo <sgazeos@gmail.com> * Fixed kernels signatures for const buffers and shapes. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored pairwise methods to persistent buffers and shapes usage. Signed-off-by: shugeo <sgazeos@gmail.com> * Adopt const to buffers and shapes with kernels. Signed-off-by: shugeo <sgazeos@gmail.com> * Adopt const to buffers and shapes with scalar kernels. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored indexreduce kernels signatures to use const buffers and shapes. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored pairwise kernels to adopt cons shapes and buffers. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored pairwise bool kernels to adopt cons shapes and buffers. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored random special ops to conform with const shapes and buffers. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored native ops to conform with const shapes and buffers under cuda platform. Signed-off-by: shugeo <sgazeos@gmail.com> * Cosmetical changes only. Signed-off-by: shugeo <sgazeos@gmail.com> * Fixed const shapes and buffers error. Signed-off-by: shugeo <sgazeos@gmail.com> * Corrected start pos routine. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored methods to conform with const shapes and buffers. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored helpers to use proper methods instead. Signed-off-by: shugeo <sgazeos@gmail.com> * bunch of changes Signed-off-by: raver119@gmail.com <raver119@gmail.com> * next bunch of changes Signed-off-by: raver119@gmail.com <raver119@gmail.com> * next bunch of changes Signed-off-by: raver119@gmail.com <raver119@gmail.com> * Fixed execScalar declaration. Signed-off-by: shugeo <sgazeos@gmail.com> * Fixed execScalar declaration. Signed-off-by: shugeo <sgazeos@gmail.com> * Corrected const shape cases with sort and so on. Signed-off-by: shugeo <sgazeos@gmail.com> * Fixed const shapes for sort. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored kernel declarations to adopt const shapes. Signed-off-by: shugeo <sgazeos@gmail.com> * Fixed kernels declarations to adopt const shapes. Signed-off-by: shugeo <sgazeos@gmail.com> * Corrected kernel declarations to adopt const shapes and buffers. Signed-off-by: shugeo <sgazeos@gmail.com> * Fixed kernels declarations to adopt const shapes. Signed-off-by: shugeo <sgazeos@gmail.com> * Fixed segment helpers kernels declarations and so on to adopt const shapes. Signed-off-by: shugeo <sgazeos@gmail.com> * Fixed const shape usage with segment and solve helpers. Signed-off-by: shugeo <sgazeos@gmail.com> * Fixed kernel declaration with adjustWeight helper. Signed-off-by: shugeo <sgazeos@gmail.com> * Fixed cuda implementations for constant shape helpers. Signed-off-by: shugeo <sgazeos@gmail.com> * Adopted const shape usage with kernels. Signed-off-by: shugeo <sgazeos@gmail.com> * Adopted top_k kernels to use const shapes and buffers. Signed-off-by: shugeo <sgazeos@gmail.com> * Corrected kernels declarations to adopt const shapes with helpers. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored NDArray definitions to adopt const shapes and buffers. Signed-off-by: shugeo <sgazeos@gmail.com> * Fixed const shapes with image suppression helpers. Signed-off-by: shugeo <sgazeos@gmail.com> * Slight improvement with buffers. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored buffer usage. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored buffer usage with tests. Signed-off-by: shugeo <sgazeos@gmail.com> * Fixed const shape usage with definitions. Signed-off-by: shugeo <sgazeos@gmail.com> * minor updates on cpu side Signed-off-by: raver119@gmail.com <raver119@gmail.com> * Refactored const shape usage with ConstantDescritor and native ops with cuda platform. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored tear and tile kernels to adopt with const shapes. Signed-off-by: shugeo <sgazeos@gmail.com> * softmax_loop fix Signed-off-by: raver119 <raver119@gmail.com> * update missing signature Signed-off-by: raver119@gmail.com <raver119@gmail.com> * softmax again Signed-off-by: raver119@gmail.com <raver119@gmail.com> * few more missing consts Signed-off-by: raver119 <raver119@gmail.com> * new methods updated Signed-off-by: raver119@gmail.com <raver119@gmail.com> Co-authored-by: shugeo <sgazeos@gmail.com>
2020-05-09 07:06:14 +02:00
auto broadcasted = alpha->shapeInfo();
if (beta != nullptr) {
const Nd4jLong* broadcastedShape = nullptr;
ShapeUtils::evalBroadcastShapeInfo(*alpha, *beta, true, broadcastedShape, context->getWorkspace());
broadcasted = broadcastedShape;
}
auto step = shape::length(broadcasted);
auto shift = output->lengthOf() / step;
auto copyAlpha = alpha;
auto copyBeta = beta;
if (beta != nullptr) {
NDArray alphaBroadcasted(broadcasted, alpha->dataType(), false, context);
NDArray betaBroadcasted(broadcasted, beta->dataType(), false, context);
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
copyAlpha = new NDArray(alphaBroadcasted.applyTrueBroadcast(BroadcastOpsTuple::Assign(), *alpha));
copyBeta = new NDArray(betaBroadcasted.applyTrueBroadcast(BroadcastOpsTuple::Assign(), *beta));
}
bool directOutput = output->ews() == 1 && output->ordering() == 'c';
T* outputBuf = output->dataBuffer()->primaryAsT<T>();
PRAGMA_OMP_PARALLEL_FOR
for (Nd4jLong k = 0; k < shift; k++) {
auto pos = k * step;
for (Nd4jLong e = 0; e < step; e++)
if (directOutput) {
2020-06-08 12:14:22 +02:00
outputBuf[pos + e] = copyAlpha->t<T>(e) <= 1? gammaLess(rng, copyAlpha->t<T>(e), beta?copyBeta->t<T>(e):T(1.f)):gammaGreat(rng, copyAlpha->t<T>(e), beta?copyBeta->t<T>(e):T(1.f));
}
else {
2020-06-08 12:14:22 +02:00
output->r<T>(pos + e) = copyAlpha->t<T>(e) <= 1? gammaLess(rng, copyAlpha->t<T>(e), beta?copyBeta->t<T>(e):T(1.f)):gammaGreat(rng, copyAlpha->t<T>(e), beta?copyBeta->t<T>(e):T(1.f));
}
}
if (beta != nullptr) {
delete copyAlpha;
delete copyBeta;
//delete broadcasted;
}
}
void fillRandomGamma(LaunchContext* context, graph::RandomGenerator& rng, NDArray* alpha, NDArray* beta, NDArray* output) {
BUILD_SINGLE_SELECTOR(output->dataType(), fillRandomGamma_, (context, rng, alpha, beta, output), FLOAT_NATIVE);
}
BUILD_SINGLE_TEMPLATE(template void fillRandomGamma_, (LaunchContext* context,
graph::RandomGenerator& rng, NDArray* alpha, NDArray* beta, NDArray* output), FLOAT_NATIVE);
/*
* algorithm Poisson generator based upon the inversion by sequential search:[48]:505
init:
Let x 0, p eλ, s p.
Generate uniform random number u in [0,1].
while u > s do:
x x + 1.
p p * λ / x.
s s + p.
return x.
* */
template <typename T>
void fillRandomPoisson_(LaunchContext* context, graph::RandomGenerator& rng, NDArray* lambda, NDArray* output) {
auto shift = output->lengthOf() / lambda->lengthOf();
auto step = lambda->lengthOf();
T* lambdaBuf = lambda->dataBuffer()->primaryAsT<T>();
T* outputBuf = output->dataBuffer()->primaryAsT<T>();
bool directLa = lambda->ews() == 1 && lambda->ordering() == 'c';
bool directOut = output->ews() == 1 && output->ordering() == 'c';
PRAGMA_OMP_PARALLEL_FOR
for (Nd4jLong k = 0; k < shift; k++) {
auto pos = k * step;
auto u = rng.relativeT<T>(k, 0., 1.);
for (Nd4jLong e = 0; e < step; e++) {
auto p = math::nd4j_exp<T, T>(-lambda->t<T>(e));
auto s = p;
auto x = T(0.f);
while (u > s) {
x += 1.f;
p *= directLa?lambdaBuf[e]/x:lambda->t<T>(e) / x;
s += p;
}
if (directOut)
outputBuf[pos + e] = x;
else
Shyrma sqrtm (#429) * - start working on implementation of sqrtm op Signed-off-by: Yurii <iuriish@yahoo.com> * - improving householder procedure Signed-off-by: Yurii <iuriish@yahoo.com> * - further polishing householder stuff Signed-off-by: Yurii <iuriish@yahoo.com> * - polishing hh pivoting qr procedure Signed-off-by: Yurii <iuriish@yahoo.com> * - polishing BiDiagonalUp procedure Signed-off-by: Yurii <iuriish@yahoo.com> * - polishing householder sequence class Signed-off-by: Yurii <iuriish@yahoo.com> * - polishing jacobi svd class Signed-off-by: Yurii <iuriish@yahoo.com> * - polishing svd stuff 1 Signed-off-by: Yurii <iuriish@yahoo.com> * - polishing svd stuff 2 Signed-off-by: Yurii <iuriish@yahoo.com> * - implementation and testing class which performs Hessenberg decomposition of square matrix Signed-off-by: Yurii <iuriish@yahoo.com> * - add static method to JacobiSVD class which makes the continuous Givens rotation generation algorithm Signed-off-by: Yurii <iuriish@yahoo.com> * - implementation and testing auxiliary methods of Schur decomp class Signed-off-by: Yurii <iuriish@yahoo.com> * some references here and there Signed-off-by: raver119 <raver119@gmail.com> * - trying figure out difference between eigen and our Schur alg Signed-off-by: Yurii <iuriish@yahoo.com> * - testing fixing bugs in Schur decomposition op Signed-off-by: Yurii <iuriish@yahoo.com> * - start to implement class which performs calculation of eigen values and vectors Signed-off-by: Yurii <iuriish@yahoo.com> * - add to EigenValsAndVecs method which calculates complex eigen vectors Signed-off-by: Yurii <iuriish@yahoo.com> * - testing and fixing bugs in EigenValsAndVecs class Signed-off-by: Yurii <iuriish@yahoo.com> * - implementation and testing triangularSolver class Signed-off-by: Yurii <iuriish@yahoo.com> * Added a 2D routine for triangular systems solve. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored triangularSolve2D routine and tests. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored another test for triangularSolve2D. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored test for triangularSolve for vector-bar case. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored triangularSolve2D routine and tests. Signed-off-by: shugeo <sgazeos@gmail.com> * - implementation of FullPivLU class Signed-off-by: Yurii <iuriish@yahoo.com> * - fix bugs in FullPivLU::solve method Signed-off-by: Yurii <iuriish@yahoo.com> * - correct permutation vector in FullPivLU::solve Signed-off-by: Yurii <iuriish@yahoo.com> * - correct include headers Signed-off-by: Yurii <iuriish@yahoo.com> * - implementation of Sqrtm class Signed-off-by: Yurii <iuriish@yahoo.com> * - testing and fixing bugs in Sqrtm class Signed-off-by: Yurii <iuriish@yahoo.com> * - include sqrtm classes to cuda folder, investigate in what places synchronization doesn't work Signed-off-by: Yurii <iuriish@yahoo.com> * Added implementation for cuda triangularSolve2D and also refactored triangularSolve2D for cpu. Signed-off-by: shugeo <sgazeos@gmail.com> * Eliminated waste implementations. Signed-off-by: shugeo <sgazeos@gmail.com> * - make offset calculation faster in t<> methods Signed-off-by: Yurii <iuriish@yahoo.com> * - rename refference T& NDArray::t<> method Signed-off-by: Yurii <iuriish@yahoo.com> * - further work on cuda sqrtm Signed-off-by: Yurii <iuriish@yahoo.com> * - provide correct synchronization to device in Sqrtm class Signed-off-by: Yurii <iuriish@yahoo.com> * - add tests for sqrtm op Signed-off-by: Yurii <iuriish@yahoo.com> * - correct fails which appeared while testing on jenkins Signed-off-by: Yurii <iuriish@yahoo.com> * - trying to find out mistake in svd::deflation method Signed-off-by: Yurii <iuriish@yahoo.com> * Revert "- trying to find out mistake in svd::deflation method" This reverts commit 19d37baddbc509028e4bc67bc932fe7449becdb6. * Revert "- trying to find out mistake in svd::deflation method" This reverts commit 19d37baddbc509028e4bc67bc932fe7449becdb6. Signed-off-by: Yurii <iuriish@yahoo.com> * - change call semantic of r<> and t<> methods Signed-off-by: Yurii <iuriish@yahoo.com> * - ged rid of ambiguity in * operator overloads for windows buikd Signed-off-by: Yurii <iuriish@yahoo.com> * - get rid of ambiguity in * operator overloads for windows build 2 Signed-off-by: Yurii <iuriish@yahoo.com> * - get rid of ambiguity in * operator overloads for windows build 3 Signed-off-by: Yurii <iuriish@yahoo.com> * - resolve conflicts with master Signed-off-by: Yurii <iuriish@yahoo.com> * cmakelists updated Signed-off-by: raver119@gmail.com <raver119@gmail.com> * - minor fix in merge cpu helper - make use of reference getter Signed-off-by: Yurii <iuriish@yahoo.com> Co-authored-by: raver119 <raver119@gmail.com> Co-authored-by: shugeo <sgazeos@gmail.com>
2020-05-14 17:06:13 +02:00
output->r<T>(pos + e) = x;
}
}
}
void fillRandomPoisson(LaunchContext* context, graph::RandomGenerator& rng, NDArray* lambda, NDArray* output) {
BUILD_SINGLE_SELECTOR(output->dataType(), fillRandomPoisson_, (context, rng, lambda, output), FLOAT_NATIVE);
}
BUILD_SINGLE_TEMPLATE(template void fillRandomPoisson_, (LaunchContext* context,
graph::RandomGenerator& rng, NDArray* lambda, NDArray* output), FLOAT_TYPES);
template <typename T>
void fillRandomUniform_(LaunchContext* context, graph::RandomGenerator& rng, NDArray* min, NDArray* max, NDArray* output) {
T minVal = T(0);
T maxVal = DataTypeUtils::max<T>();
if (min)
minVal = min->t<T>(0);
if (max)
maxVal = max->t<T>(0);
if (output->isR())
RandomLauncher::fillUniform(context, rng, output, minVal, maxVal);
else {
PRAGMA_OMP_PARALLEL_FOR
for (Nd4jLong i = 0; i < output->lengthOf(); i++) {
Shyrma sqrtm (#429) * - start working on implementation of sqrtm op Signed-off-by: Yurii <iuriish@yahoo.com> * - improving householder procedure Signed-off-by: Yurii <iuriish@yahoo.com> * - further polishing householder stuff Signed-off-by: Yurii <iuriish@yahoo.com> * - polishing hh pivoting qr procedure Signed-off-by: Yurii <iuriish@yahoo.com> * - polishing BiDiagonalUp procedure Signed-off-by: Yurii <iuriish@yahoo.com> * - polishing householder sequence class Signed-off-by: Yurii <iuriish@yahoo.com> * - polishing jacobi svd class Signed-off-by: Yurii <iuriish@yahoo.com> * - polishing svd stuff 1 Signed-off-by: Yurii <iuriish@yahoo.com> * - polishing svd stuff 2 Signed-off-by: Yurii <iuriish@yahoo.com> * - implementation and testing class which performs Hessenberg decomposition of square matrix Signed-off-by: Yurii <iuriish@yahoo.com> * - add static method to JacobiSVD class which makes the continuous Givens rotation generation algorithm Signed-off-by: Yurii <iuriish@yahoo.com> * - implementation and testing auxiliary methods of Schur decomp class Signed-off-by: Yurii <iuriish@yahoo.com> * some references here and there Signed-off-by: raver119 <raver119@gmail.com> * - trying figure out difference between eigen and our Schur alg Signed-off-by: Yurii <iuriish@yahoo.com> * - testing fixing bugs in Schur decomposition op Signed-off-by: Yurii <iuriish@yahoo.com> * - start to implement class which performs calculation of eigen values and vectors Signed-off-by: Yurii <iuriish@yahoo.com> * - add to EigenValsAndVecs method which calculates complex eigen vectors Signed-off-by: Yurii <iuriish@yahoo.com> * - testing and fixing bugs in EigenValsAndVecs class Signed-off-by: Yurii <iuriish@yahoo.com> * - implementation and testing triangularSolver class Signed-off-by: Yurii <iuriish@yahoo.com> * Added a 2D routine for triangular systems solve. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored triangularSolve2D routine and tests. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored another test for triangularSolve2D. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored test for triangularSolve for vector-bar case. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored triangularSolve2D routine and tests. Signed-off-by: shugeo <sgazeos@gmail.com> * - implementation of FullPivLU class Signed-off-by: Yurii <iuriish@yahoo.com> * - fix bugs in FullPivLU::solve method Signed-off-by: Yurii <iuriish@yahoo.com> * - correct permutation vector in FullPivLU::solve Signed-off-by: Yurii <iuriish@yahoo.com> * - correct include headers Signed-off-by: Yurii <iuriish@yahoo.com> * - implementation of Sqrtm class Signed-off-by: Yurii <iuriish@yahoo.com> * - testing and fixing bugs in Sqrtm class Signed-off-by: Yurii <iuriish@yahoo.com> * - include sqrtm classes to cuda folder, investigate in what places synchronization doesn't work Signed-off-by: Yurii <iuriish@yahoo.com> * Added implementation for cuda triangularSolve2D and also refactored triangularSolve2D for cpu. Signed-off-by: shugeo <sgazeos@gmail.com> * Eliminated waste implementations. Signed-off-by: shugeo <sgazeos@gmail.com> * - make offset calculation faster in t<> methods Signed-off-by: Yurii <iuriish@yahoo.com> * - rename refference T& NDArray::t<> method Signed-off-by: Yurii <iuriish@yahoo.com> * - further work on cuda sqrtm Signed-off-by: Yurii <iuriish@yahoo.com> * - provide correct synchronization to device in Sqrtm class Signed-off-by: Yurii <iuriish@yahoo.com> * - add tests for sqrtm op Signed-off-by: Yurii <iuriish@yahoo.com> * - correct fails which appeared while testing on jenkins Signed-off-by: Yurii <iuriish@yahoo.com> * - trying to find out mistake in svd::deflation method Signed-off-by: Yurii <iuriish@yahoo.com> * Revert "- trying to find out mistake in svd::deflation method" This reverts commit 19d37baddbc509028e4bc67bc932fe7449becdb6. * Revert "- trying to find out mistake in svd::deflation method" This reverts commit 19d37baddbc509028e4bc67bc932fe7449becdb6. Signed-off-by: Yurii <iuriish@yahoo.com> * - change call semantic of r<> and t<> methods Signed-off-by: Yurii <iuriish@yahoo.com> * - ged rid of ambiguity in * operator overloads for windows buikd Signed-off-by: Yurii <iuriish@yahoo.com> * - get rid of ambiguity in * operator overloads for windows build 2 Signed-off-by: Yurii <iuriish@yahoo.com> * - get rid of ambiguity in * operator overloads for windows build 3 Signed-off-by: Yurii <iuriish@yahoo.com> * - resolve conflicts with master Signed-off-by: Yurii <iuriish@yahoo.com> * cmakelists updated Signed-off-by: raver119@gmail.com <raver119@gmail.com> * - minor fix in merge cpu helper - make use of reference getter Signed-off-by: Yurii <iuriish@yahoo.com> Co-authored-by: raver119 <raver119@gmail.com> Co-authored-by: shugeo <sgazeos@gmail.com>
2020-05-14 17:06:13 +02:00
output->r<T>(i) = rng.relativeT<T>(i, minVal, maxVal);
}
}
}
void fillRandomUniform(LaunchContext* context, graph::RandomGenerator& rng, NDArray* min, NDArray* max, NDArray* output) {
BUILD_SINGLE_SELECTOR(output->dataType(), fillRandomUniform_, (context, rng, min, max, output), NUMERIC_TYPES);
}
Oleh multinomial (#163) * libnd4j: Multinomial op #8570 first raw step of multinomial random data generator implementation Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j: Multinomial op #8570 next step of multinomial random categories generator implementation on both cpu and cuda, need corrections and code clean up before review and testing * libnd4j: Multinomial op #8570 code clean up and fixed issues data selecting, moved from coords to tads * libnd4j: Multinomial op #8570 fixed cuda build add reference for math materials that was used for implementation * libnd4j: Multinomial op #8570 fixed several bugs, added several tests and improved cuda version. current implementation works, need testing of reproduction with the same seed * libnd4j: Multinomial op #8570 fixes and optimization after discussion in both cuda and cpu * libnd4j: Multinomial op #8570 add corrections after review, removed tads, replace 2D parallel loop by 3D Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j: Multinomial op fixed declaration and add tests need discussion * libnd4j: Multinomial op fix in test * libnd4j: Multinomial op corrected behavior to get reproducible results, fixed issue in uniform value getting, tests added, need cuda review and cuda testing Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j: Multinomial op fixed indexing on uniform calculation Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j: Multinomial op some corrections in max min declaration Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j: Multinomial op fixed index calculation, added rewind, corrected input declaration, added stats tests, both cuda and cpu. cuda need testing * libnd4j: Multinomial op fixed bugs on cuda nad cpu. need review Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j: Multinomial op corrected tests to handle different orders Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j: Multinomial op some improvements after code review Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j: Multinomial op more corrections after review Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j: Multinomial op fixed seed usage, update tests, fixed cuda based on comments, fixed bug of rewind, removed one behavior, minor corrections. Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j: Multinomial op minor corrections Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j: Multinomial op rise the bound of fluctuation for random cases Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j: Multinomial op modified operation inputs and update implementation and tests on both cpu and cuda * libnd4j: Multinomial op corrected data types according ops.proto Co-authored-by: raver119 <raver119@gmail.com>
2020-01-06 20:35:05 +01:00
// used https://en.wikipedia.org/wiki/Categorical_distribution
// methods: gumbel trick + softmax + argmax
template <typename Tx, typename Tz>
void fillRandomMultiNomial_(LaunchContext* context, graph::RandomGenerator& rng, NDArray& input, NDArray& output, const Nd4jLong numOfSamples, const int dimC) {
Shyrma sqrtm (#429) * - start working on implementation of sqrtm op Signed-off-by: Yurii <iuriish@yahoo.com> * - improving householder procedure Signed-off-by: Yurii <iuriish@yahoo.com> * - further polishing householder stuff Signed-off-by: Yurii <iuriish@yahoo.com> * - polishing hh pivoting qr procedure Signed-off-by: Yurii <iuriish@yahoo.com> * - polishing BiDiagonalUp procedure Signed-off-by: Yurii <iuriish@yahoo.com> * - polishing householder sequence class Signed-off-by: Yurii <iuriish@yahoo.com> * - polishing jacobi svd class Signed-off-by: Yurii <iuriish@yahoo.com> * - polishing svd stuff 1 Signed-off-by: Yurii <iuriish@yahoo.com> * - polishing svd stuff 2 Signed-off-by: Yurii <iuriish@yahoo.com> * - implementation and testing class which performs Hessenberg decomposition of square matrix Signed-off-by: Yurii <iuriish@yahoo.com> * - add static method to JacobiSVD class which makes the continuous Givens rotation generation algorithm Signed-off-by: Yurii <iuriish@yahoo.com> * - implementation and testing auxiliary methods of Schur decomp class Signed-off-by: Yurii <iuriish@yahoo.com> * some references here and there Signed-off-by: raver119 <raver119@gmail.com> * - trying figure out difference between eigen and our Schur alg Signed-off-by: Yurii <iuriish@yahoo.com> * - testing fixing bugs in Schur decomposition op Signed-off-by: Yurii <iuriish@yahoo.com> * - start to implement class which performs calculation of eigen values and vectors Signed-off-by: Yurii <iuriish@yahoo.com> * - add to EigenValsAndVecs method which calculates complex eigen vectors Signed-off-by: Yurii <iuriish@yahoo.com> * - testing and fixing bugs in EigenValsAndVecs class Signed-off-by: Yurii <iuriish@yahoo.com> * - implementation and testing triangularSolver class Signed-off-by: Yurii <iuriish@yahoo.com> * Added a 2D routine for triangular systems solve. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored triangularSolve2D routine and tests. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored another test for triangularSolve2D. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored test for triangularSolve for vector-bar case. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored triangularSolve2D routine and tests. Signed-off-by: shugeo <sgazeos@gmail.com> * - implementation of FullPivLU class Signed-off-by: Yurii <iuriish@yahoo.com> * - fix bugs in FullPivLU::solve method Signed-off-by: Yurii <iuriish@yahoo.com> * - correct permutation vector in FullPivLU::solve Signed-off-by: Yurii <iuriish@yahoo.com> * - correct include headers Signed-off-by: Yurii <iuriish@yahoo.com> * - implementation of Sqrtm class Signed-off-by: Yurii <iuriish@yahoo.com> * - testing and fixing bugs in Sqrtm class Signed-off-by: Yurii <iuriish@yahoo.com> * - include sqrtm classes to cuda folder, investigate in what places synchronization doesn't work Signed-off-by: Yurii <iuriish@yahoo.com> * Added implementation for cuda triangularSolve2D and also refactored triangularSolve2D for cpu. Signed-off-by: shugeo <sgazeos@gmail.com> * Eliminated waste implementations. Signed-off-by: shugeo <sgazeos@gmail.com> * - make offset calculation faster in t<> methods Signed-off-by: Yurii <iuriish@yahoo.com> * - rename refference T& NDArray::t<> method Signed-off-by: Yurii <iuriish@yahoo.com> * - further work on cuda sqrtm Signed-off-by: Yurii <iuriish@yahoo.com> * - provide correct synchronization to device in Sqrtm class Signed-off-by: Yurii <iuriish@yahoo.com> * - add tests for sqrtm op Signed-off-by: Yurii <iuriish@yahoo.com> * - correct fails which appeared while testing on jenkins Signed-off-by: Yurii <iuriish@yahoo.com> * - trying to find out mistake in svd::deflation method Signed-off-by: Yurii <iuriish@yahoo.com> * Revert "- trying to find out mistake in svd::deflation method" This reverts commit 19d37baddbc509028e4bc67bc932fe7449becdb6. * Revert "- trying to find out mistake in svd::deflation method" This reverts commit 19d37baddbc509028e4bc67bc932fe7449becdb6. Signed-off-by: Yurii <iuriish@yahoo.com> * - change call semantic of r<> and t<> methods Signed-off-by: Yurii <iuriish@yahoo.com> * - ged rid of ambiguity in * operator overloads for windows buikd Signed-off-by: Yurii <iuriish@yahoo.com> * - get rid of ambiguity in * operator overloads for windows build 2 Signed-off-by: Yurii <iuriish@yahoo.com> * - get rid of ambiguity in * operator overloads for windows build 3 Signed-off-by: Yurii <iuriish@yahoo.com> * - resolve conflicts with master Signed-off-by: Yurii <iuriish@yahoo.com> * cmakelists updated Signed-off-by: raver119@gmail.com <raver119@gmail.com> * - minor fix in merge cpu helper - make use of reference getter Signed-off-by: Yurii <iuriish@yahoo.com> Co-authored-by: raver119 <raver119@gmail.com> Co-authored-by: shugeo <sgazeos@gmail.com>
2020-05-14 17:06:13 +02:00
Oleh multinomial (#163) * libnd4j: Multinomial op #8570 first raw step of multinomial random data generator implementation Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j: Multinomial op #8570 next step of multinomial random categories generator implementation on both cpu and cuda, need corrections and code clean up before review and testing * libnd4j: Multinomial op #8570 code clean up and fixed issues data selecting, moved from coords to tads * libnd4j: Multinomial op #8570 fixed cuda build add reference for math materials that was used for implementation * libnd4j: Multinomial op #8570 fixed several bugs, added several tests and improved cuda version. current implementation works, need testing of reproduction with the same seed * libnd4j: Multinomial op #8570 fixes and optimization after discussion in both cuda and cpu * libnd4j: Multinomial op #8570 add corrections after review, removed tads, replace 2D parallel loop by 3D Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j: Multinomial op fixed declaration and add tests need discussion * libnd4j: Multinomial op fix in test * libnd4j: Multinomial op corrected behavior to get reproducible results, fixed issue in uniform value getting, tests added, need cuda review and cuda testing Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j: Multinomial op fixed indexing on uniform calculation Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j: Multinomial op some corrections in max min declaration Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j: Multinomial op fixed index calculation, added rewind, corrected input declaration, added stats tests, both cuda and cpu. cuda need testing * libnd4j: Multinomial op fixed bugs on cuda nad cpu. need review Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j: Multinomial op corrected tests to handle different orders Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j: Multinomial op some improvements after code review Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j: Multinomial op more corrections after review Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j: Multinomial op fixed seed usage, update tests, fixed cuda based on comments, fixed bug of rewind, removed one behavior, minor corrections. Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j: Multinomial op minor corrections Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j: Multinomial op rise the bound of fluctuation for random cases Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j: Multinomial op modified operation inputs and update implementation and tests on both cpu and cuda * libnd4j: Multinomial op corrected data types according ops.proto Co-authored-by: raver119 <raver119@gmail.com>
2020-01-06 20:35:05 +01:00
const Tx* x = input.bufferAsT<Tx>();
Tz* z = output.bufferAsT<Tz>();
Shyrma sqrtm (#429) * - start working on implementation of sqrtm op Signed-off-by: Yurii <iuriish@yahoo.com> * - improving householder procedure Signed-off-by: Yurii <iuriish@yahoo.com> * - further polishing householder stuff Signed-off-by: Yurii <iuriish@yahoo.com> * - polishing hh pivoting qr procedure Signed-off-by: Yurii <iuriish@yahoo.com> * - polishing BiDiagonalUp procedure Signed-off-by: Yurii <iuriish@yahoo.com> * - polishing householder sequence class Signed-off-by: Yurii <iuriish@yahoo.com> * - polishing jacobi svd class Signed-off-by: Yurii <iuriish@yahoo.com> * - polishing svd stuff 1 Signed-off-by: Yurii <iuriish@yahoo.com> * - polishing svd stuff 2 Signed-off-by: Yurii <iuriish@yahoo.com> * - implementation and testing class which performs Hessenberg decomposition of square matrix Signed-off-by: Yurii <iuriish@yahoo.com> * - add static method to JacobiSVD class which makes the continuous Givens rotation generation algorithm Signed-off-by: Yurii <iuriish@yahoo.com> * - implementation and testing auxiliary methods of Schur decomp class Signed-off-by: Yurii <iuriish@yahoo.com> * some references here and there Signed-off-by: raver119 <raver119@gmail.com> * - trying figure out difference between eigen and our Schur alg Signed-off-by: Yurii <iuriish@yahoo.com> * - testing fixing bugs in Schur decomposition op Signed-off-by: Yurii <iuriish@yahoo.com> * - start to implement class which performs calculation of eigen values and vectors Signed-off-by: Yurii <iuriish@yahoo.com> * - add to EigenValsAndVecs method which calculates complex eigen vectors Signed-off-by: Yurii <iuriish@yahoo.com> * - testing and fixing bugs in EigenValsAndVecs class Signed-off-by: Yurii <iuriish@yahoo.com> * - implementation and testing triangularSolver class Signed-off-by: Yurii <iuriish@yahoo.com> * Added a 2D routine for triangular systems solve. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored triangularSolve2D routine and tests. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored another test for triangularSolve2D. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored test for triangularSolve for vector-bar case. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored triangularSolve2D routine and tests. Signed-off-by: shugeo <sgazeos@gmail.com> * - implementation of FullPivLU class Signed-off-by: Yurii <iuriish@yahoo.com> * - fix bugs in FullPivLU::solve method Signed-off-by: Yurii <iuriish@yahoo.com> * - correct permutation vector in FullPivLU::solve Signed-off-by: Yurii <iuriish@yahoo.com> * - correct include headers Signed-off-by: Yurii <iuriish@yahoo.com> * - implementation of Sqrtm class Signed-off-by: Yurii <iuriish@yahoo.com> * - testing and fixing bugs in Sqrtm class Signed-off-by: Yurii <iuriish@yahoo.com> * - include sqrtm classes to cuda folder, investigate in what places synchronization doesn't work Signed-off-by: Yurii <iuriish@yahoo.com> * Added implementation for cuda triangularSolve2D and also refactored triangularSolve2D for cpu. Signed-off-by: shugeo <sgazeos@gmail.com> * Eliminated waste implementations. Signed-off-by: shugeo <sgazeos@gmail.com> * - make offset calculation faster in t<> methods Signed-off-by: Yurii <iuriish@yahoo.com> * - rename refference T& NDArray::t<> method Signed-off-by: Yurii <iuriish@yahoo.com> * - further work on cuda sqrtm Signed-off-by: Yurii <iuriish@yahoo.com> * - provide correct synchronization to device in Sqrtm class Signed-off-by: Yurii <iuriish@yahoo.com> * - add tests for sqrtm op Signed-off-by: Yurii <iuriish@yahoo.com> * - correct fails which appeared while testing on jenkins Signed-off-by: Yurii <iuriish@yahoo.com> * - trying to find out mistake in svd::deflation method Signed-off-by: Yurii <iuriish@yahoo.com> * Revert "- trying to find out mistake in svd::deflation method" This reverts commit 19d37baddbc509028e4bc67bc932fe7449becdb6. * Revert "- trying to find out mistake in svd::deflation method" This reverts commit 19d37baddbc509028e4bc67bc932fe7449becdb6. Signed-off-by: Yurii <iuriish@yahoo.com> * - change call semantic of r<> and t<> methods Signed-off-by: Yurii <iuriish@yahoo.com> * - ged rid of ambiguity in * operator overloads for windows buikd Signed-off-by: Yurii <iuriish@yahoo.com> * - get rid of ambiguity in * operator overloads for windows build 2 Signed-off-by: Yurii <iuriish@yahoo.com> * - get rid of ambiguity in * operator overloads for windows build 3 Signed-off-by: Yurii <iuriish@yahoo.com> * - resolve conflicts with master Signed-off-by: Yurii <iuriish@yahoo.com> * cmakelists updated Signed-off-by: raver119@gmail.com <raver119@gmail.com> * - minor fix in merge cpu helper - make use of reference getter Signed-off-by: Yurii <iuriish@yahoo.com> Co-authored-by: raver119 <raver119@gmail.com> Co-authored-by: shugeo <sgazeos@gmail.com>
2020-05-14 17:06:13 +02:00
Oleh multinomial (#163) * libnd4j: Multinomial op #8570 first raw step of multinomial random data generator implementation Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j: Multinomial op #8570 next step of multinomial random categories generator implementation on both cpu and cuda, need corrections and code clean up before review and testing * libnd4j: Multinomial op #8570 code clean up and fixed issues data selecting, moved from coords to tads * libnd4j: Multinomial op #8570 fixed cuda build add reference for math materials that was used for implementation * libnd4j: Multinomial op #8570 fixed several bugs, added several tests and improved cuda version. current implementation works, need testing of reproduction with the same seed * libnd4j: Multinomial op #8570 fixes and optimization after discussion in both cuda and cpu * libnd4j: Multinomial op #8570 add corrections after review, removed tads, replace 2D parallel loop by 3D Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j: Multinomial op fixed declaration and add tests need discussion * libnd4j: Multinomial op fix in test * libnd4j: Multinomial op corrected behavior to get reproducible results, fixed issue in uniform value getting, tests added, need cuda review and cuda testing Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j: Multinomial op fixed indexing on uniform calculation Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j: Multinomial op some corrections in max min declaration Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j: Multinomial op fixed index calculation, added rewind, corrected input declaration, added stats tests, both cuda and cpu. cuda need testing * libnd4j: Multinomial op fixed bugs on cuda nad cpu. need review Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j: Multinomial op corrected tests to handle different orders Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j: Multinomial op some improvements after code review Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j: Multinomial op more corrections after review Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j: Multinomial op fixed seed usage, update tests, fixed cuda based on comments, fixed bug of rewind, removed one behavior, minor corrections. Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j: Multinomial op minor corrections Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j: Multinomial op rise the bound of fluctuation for random cases Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j: Multinomial op modified operation inputs and update implementation and tests on both cpu and cuda * libnd4j: Multinomial op corrected data types according ops.proto Co-authored-by: raver119 <raver119@gmail.com>
2020-01-06 20:35:05 +01:00
Tx minVal = DataTypeUtils::min<Tx>();
Shyrma sqrtm (#429) * - start working on implementation of sqrtm op Signed-off-by: Yurii <iuriish@yahoo.com> * - improving householder procedure Signed-off-by: Yurii <iuriish@yahoo.com> * - further polishing householder stuff Signed-off-by: Yurii <iuriish@yahoo.com> * - polishing hh pivoting qr procedure Signed-off-by: Yurii <iuriish@yahoo.com> * - polishing BiDiagonalUp procedure Signed-off-by: Yurii <iuriish@yahoo.com> * - polishing householder sequence class Signed-off-by: Yurii <iuriish@yahoo.com> * - polishing jacobi svd class Signed-off-by: Yurii <iuriish@yahoo.com> * - polishing svd stuff 1 Signed-off-by: Yurii <iuriish@yahoo.com> * - polishing svd stuff 2 Signed-off-by: Yurii <iuriish@yahoo.com> * - implementation and testing class which performs Hessenberg decomposition of square matrix Signed-off-by: Yurii <iuriish@yahoo.com> * - add static method to JacobiSVD class which makes the continuous Givens rotation generation algorithm Signed-off-by: Yurii <iuriish@yahoo.com> * - implementation and testing auxiliary methods of Schur decomp class Signed-off-by: Yurii <iuriish@yahoo.com> * some references here and there Signed-off-by: raver119 <raver119@gmail.com> * - trying figure out difference between eigen and our Schur alg Signed-off-by: Yurii <iuriish@yahoo.com> * - testing fixing bugs in Schur decomposition op Signed-off-by: Yurii <iuriish@yahoo.com> * - start to implement class which performs calculation of eigen values and vectors Signed-off-by: Yurii <iuriish@yahoo.com> * - add to EigenValsAndVecs method which calculates complex eigen vectors Signed-off-by: Yurii <iuriish@yahoo.com> * - testing and fixing bugs in EigenValsAndVecs class Signed-off-by: Yurii <iuriish@yahoo.com> * - implementation and testing triangularSolver class Signed-off-by: Yurii <iuriish@yahoo.com> * Added a 2D routine for triangular systems solve. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored triangularSolve2D routine and tests. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored another test for triangularSolve2D. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored test for triangularSolve for vector-bar case. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored triangularSolve2D routine and tests. Signed-off-by: shugeo <sgazeos@gmail.com> * - implementation of FullPivLU class Signed-off-by: Yurii <iuriish@yahoo.com> * - fix bugs in FullPivLU::solve method Signed-off-by: Yurii <iuriish@yahoo.com> * - correct permutation vector in FullPivLU::solve Signed-off-by: Yurii <iuriish@yahoo.com> * - correct include headers Signed-off-by: Yurii <iuriish@yahoo.com> * - implementation of Sqrtm class Signed-off-by: Yurii <iuriish@yahoo.com> * - testing and fixing bugs in Sqrtm class Signed-off-by: Yurii <iuriish@yahoo.com> * - include sqrtm classes to cuda folder, investigate in what places synchronization doesn't work Signed-off-by: Yurii <iuriish@yahoo.com> * Added implementation for cuda triangularSolve2D and also refactored triangularSolve2D for cpu. Signed-off-by: shugeo <sgazeos@gmail.com> * Eliminated waste implementations. Signed-off-by: shugeo <sgazeos@gmail.com> * - make offset calculation faster in t<> methods Signed-off-by: Yurii <iuriish@yahoo.com> * - rename refference T& NDArray::t<> method Signed-off-by: Yurii <iuriish@yahoo.com> * - further work on cuda sqrtm Signed-off-by: Yurii <iuriish@yahoo.com> * - provide correct synchronization to device in Sqrtm class Signed-off-by: Yurii <iuriish@yahoo.com> * - add tests for sqrtm op Signed-off-by: Yurii <iuriish@yahoo.com> * - correct fails which appeared while testing on jenkins Signed-off-by: Yurii <iuriish@yahoo.com> * - trying to find out mistake in svd::deflation method Signed-off-by: Yurii <iuriish@yahoo.com> * Revert "- trying to find out mistake in svd::deflation method" This reverts commit 19d37baddbc509028e4bc67bc932fe7449becdb6. * Revert "- trying to find out mistake in svd::deflation method" This reverts commit 19d37baddbc509028e4bc67bc932fe7449becdb6. Signed-off-by: Yurii <iuriish@yahoo.com> * - change call semantic of r<> and t<> methods Signed-off-by: Yurii <iuriish@yahoo.com> * - ged rid of ambiguity in * operator overloads for windows buikd Signed-off-by: Yurii <iuriish@yahoo.com> * - get rid of ambiguity in * operator overloads for windows build 2 Signed-off-by: Yurii <iuriish@yahoo.com> * - get rid of ambiguity in * operator overloads for windows build 3 Signed-off-by: Yurii <iuriish@yahoo.com> * - resolve conflicts with master Signed-off-by: Yurii <iuriish@yahoo.com> * cmakelists updated Signed-off-by: raver119@gmail.com <raver119@gmail.com> * - minor fix in merge cpu helper - make use of reference getter Signed-off-by: Yurii <iuriish@yahoo.com> Co-authored-by: raver119 <raver119@gmail.com> Co-authored-by: shugeo <sgazeos@gmail.com>
2020-05-14 17:06:13 +02:00
Tx maxVal = 1.0;
Oleh multinomial (#163) * libnd4j: Multinomial op #8570 first raw step of multinomial random data generator implementation Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j: Multinomial op #8570 next step of multinomial random categories generator implementation on both cpu and cuda, need corrections and code clean up before review and testing * libnd4j: Multinomial op #8570 code clean up and fixed issues data selecting, moved from coords to tads * libnd4j: Multinomial op #8570 fixed cuda build add reference for math materials that was used for implementation * libnd4j: Multinomial op #8570 fixed several bugs, added several tests and improved cuda version. current implementation works, need testing of reproduction with the same seed * libnd4j: Multinomial op #8570 fixes and optimization after discussion in both cuda and cpu * libnd4j: Multinomial op #8570 add corrections after review, removed tads, replace 2D parallel loop by 3D Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j: Multinomial op fixed declaration and add tests need discussion * libnd4j: Multinomial op fix in test * libnd4j: Multinomial op corrected behavior to get reproducible results, fixed issue in uniform value getting, tests added, need cuda review and cuda testing Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j: Multinomial op fixed indexing on uniform calculation Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j: Multinomial op some corrections in max min declaration Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j: Multinomial op fixed index calculation, added rewind, corrected input declaration, added stats tests, both cuda and cpu. cuda need testing * libnd4j: Multinomial op fixed bugs on cuda nad cpu. need review Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j: Multinomial op corrected tests to handle different orders Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j: Multinomial op some improvements after code review Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j: Multinomial op more corrections after review Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j: Multinomial op fixed seed usage, update tests, fixed cuda based on comments, fixed bug of rewind, removed one behavior, minor corrections. Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j: Multinomial op minor corrections Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j: Multinomial op rise the bound of fluctuation for random cases Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j: Multinomial op modified operation inputs and update implementation and tests on both cpu and cuda * libnd4j: Multinomial op corrected data types according ops.proto Co-authored-by: raver119 <raver119@gmail.com>
2020-01-06 20:35:05 +01:00
auto dimA = (0 == dimC) ? 1 : 0;
const Nd4jLong batchValue = output.sizeAt(dimC);
const Nd4jLong numOfClassX = input.sizeAt(dimA);
const Nd4jLong zDimAstride = output.stridesOf()[dimA];
const Nd4jLong xDimAstride = input.stridesOf()[dimA];
const Nd4jLong zDimCstride = output.stridesOf()[dimC];
const Nd4jLong xDimCstride = input.stridesOf()[dimC];
auto func = PRAGMA_THREADS_FOR_2D{
for (auto nBatchIndex = start_x; nBatchIndex < stop_x; nBatchIndex += inc_x) {
for (auto nSampleIndexInBatch = start_y; nSampleIndexInBatch < stop_y; nSampleIndexInBatch += inc_y) {
Shyrma sqrtm (#429) * - start working on implementation of sqrtm op Signed-off-by: Yurii <iuriish@yahoo.com> * - improving householder procedure Signed-off-by: Yurii <iuriish@yahoo.com> * - further polishing householder stuff Signed-off-by: Yurii <iuriish@yahoo.com> * - polishing hh pivoting qr procedure Signed-off-by: Yurii <iuriish@yahoo.com> * - polishing BiDiagonalUp procedure Signed-off-by: Yurii <iuriish@yahoo.com> * - polishing householder sequence class Signed-off-by: Yurii <iuriish@yahoo.com> * - polishing jacobi svd class Signed-off-by: Yurii <iuriish@yahoo.com> * - polishing svd stuff 1 Signed-off-by: Yurii <iuriish@yahoo.com> * - polishing svd stuff 2 Signed-off-by: Yurii <iuriish@yahoo.com> * - implementation and testing class which performs Hessenberg decomposition of square matrix Signed-off-by: Yurii <iuriish@yahoo.com> * - add static method to JacobiSVD class which makes the continuous Givens rotation generation algorithm Signed-off-by: Yurii <iuriish@yahoo.com> * - implementation and testing auxiliary methods of Schur decomp class Signed-off-by: Yurii <iuriish@yahoo.com> * some references here and there Signed-off-by: raver119 <raver119@gmail.com> * - trying figure out difference between eigen and our Schur alg Signed-off-by: Yurii <iuriish@yahoo.com> * - testing fixing bugs in Schur decomposition op Signed-off-by: Yurii <iuriish@yahoo.com> * - start to implement class which performs calculation of eigen values and vectors Signed-off-by: Yurii <iuriish@yahoo.com> * - add to EigenValsAndVecs method which calculates complex eigen vectors Signed-off-by: Yurii <iuriish@yahoo.com> * - testing and fixing bugs in EigenValsAndVecs class Signed-off-by: Yurii <iuriish@yahoo.com> * - implementation and testing triangularSolver class Signed-off-by: Yurii <iuriish@yahoo.com> * Added a 2D routine for triangular systems solve. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored triangularSolve2D routine and tests. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored another test for triangularSolve2D. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored test for triangularSolve for vector-bar case. Signed-off-by: shugeo <sgazeos@gmail.com> * Refactored triangularSolve2D routine and tests. Signed-off-by: shugeo <sgazeos@gmail.com> * - implementation of FullPivLU class Signed-off-by: Yurii <iuriish@yahoo.com> * - fix bugs in FullPivLU::solve method Signed-off-by: Yurii <iuriish@yahoo.com> * - correct permutation vector in FullPivLU::solve Signed-off-by: Yurii <iuriish@yahoo.com> * - correct include headers Signed-off-by: Yurii <iuriish@yahoo.com> * - implementation of Sqrtm class Signed-off-by: Yurii <iuriish@yahoo.com> * - testing and fixing bugs in Sqrtm class Signed-off-by: Yurii <iuriish@yahoo.com> * - include sqrtm classes to cuda folder, investigate in what places synchronization doesn't work Signed-off-by: Yurii <iuriish@yahoo.com> * Added implementation for cuda triangularSolve2D and also refactored triangularSolve2D for cpu. Signed-off-by: shugeo <sgazeos@gmail.com> * Eliminated waste implementations. Signed-off-by: shugeo <sgazeos@gmail.com> * - make offset calculation faster in t<> methods Signed-off-by: Yurii <iuriish@yahoo.com> * - rename refference T& NDArray::t<> method Signed-off-by: Yurii <iuriish@yahoo.com> * - further work on cuda sqrtm Signed-off-by: Yurii <iuriish@yahoo.com> * - provide correct synchronization to device in Sqrtm class Signed-off-by: Yurii <iuriish@yahoo.com> * - add tests for sqrtm op Signed-off-by: Yurii <iuriish@yahoo.com> * - correct fails which appeared while testing on jenkins Signed-off-by: Yurii <iuriish@yahoo.com> * - trying to find out mistake in svd::deflation method Signed-off-by: Yurii <iuriish@yahoo.com> * Revert "- trying to find out mistake in svd::deflation method" This reverts commit 19d37baddbc509028e4bc67bc932fe7449becdb6. * Revert "- trying to find out mistake in svd::deflation method" This reverts commit 19d37baddbc509028e4bc67bc932fe7449becdb6. Signed-off-by: Yurii <iuriish@yahoo.com> * - change call semantic of r<> and t<> methods Signed-off-by: Yurii <iuriish@yahoo.com> * - ged rid of ambiguity in * operator overloads for windows buikd Signed-off-by: Yurii <iuriish@yahoo.com> * - get rid of ambiguity in * operator overloads for windows build 2 Signed-off-by: Yurii <iuriish@yahoo.com> * - get rid of ambiguity in * operator overloads for windows build 3 Signed-off-by: Yurii <iuriish@yahoo.com> * - resolve conflicts with master Signed-off-by: Yurii <iuriish@yahoo.com> * cmakelists updated Signed-off-by: raver119@gmail.com <raver119@gmail.com> * - minor fix in merge cpu helper - make use of reference getter Signed-off-by: Yurii <iuriish@yahoo.com> Co-authored-by: raver119 <raver119@gmail.com> Co-authored-by: shugeo <sgazeos@gmail.com>
2020-05-14 17:06:13 +02:00
Oleh multinomial (#163) * libnd4j: Multinomial op #8570 first raw step of multinomial random data generator implementation Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j: Multinomial op #8570 next step of multinomial random categories generator implementation on both cpu and cuda, need corrections and code clean up before review and testing * libnd4j: Multinomial op #8570 code clean up and fixed issues data selecting, moved from coords to tads * libnd4j: Multinomial op #8570 fixed cuda build add reference for math materials that was used for implementation * libnd4j: Multinomial op #8570 fixed several bugs, added several tests and improved cuda version. current implementation works, need testing of reproduction with the same seed * libnd4j: Multinomial op #8570 fixes and optimization after discussion in both cuda and cpu * libnd4j: Multinomial op #8570 add corrections after review, removed tads, replace 2D parallel loop by 3D Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j: Multinomial op fixed declaration and add tests need discussion * libnd4j: Multinomial op fix in test * libnd4j: Multinomial op corrected behavior to get reproducible results, fixed issue in uniform value getting, tests added, need cuda review and cuda testing Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j: Multinomial op fixed indexing on uniform calculation Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j: Multinomial op some corrections in max min declaration Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j: Multinomial op fixed index calculation, added rewind, corrected input declaration, added stats tests, both cuda and cpu. cuda need testing * libnd4j: Multinomial op fixed bugs on cuda nad cpu. need review Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j: Multinomial op corrected tests to handle different orders Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j: Multinomial op some improvements after code review Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j: Multinomial op more corrections after review Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j: Multinomial op fixed seed usage, update tests, fixed cuda based on comments, fixed bug of rewind, removed one behavior, minor corrections. Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j: Multinomial op minor corrections Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j: Multinomial op rise the bound of fluctuation for random cases Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j: Multinomial op modified operation inputs and update implementation and tests on both cpu and cuda * libnd4j: Multinomial op corrected data types according ops.proto Co-authored-by: raver119 <raver119@gmail.com>
2020-01-06 20:35:05 +01:00
const Tx* xTad = x + (nBatchIndex * xDimCstride);
Tz* zTad = z + (nBatchIndex * zDimCstride);
Tz& arg = zTad[nSampleIndexInBatch * zDimAstride];
Tx Max = -minVal;
auto nSamplesPerBatch = nBatchIndex * numOfClassX * numOfSamples;
auto nClassesPerSample = nSampleIndexInBatch * numOfClassX;
for (Nd4jLong nClass = 0; nClass < numOfClassX; nClass += 1) {
Oleh multinomial (#163) * libnd4j: Multinomial op #8570 first raw step of multinomial random data generator implementation Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j: Multinomial op #8570 next step of multinomial random categories generator implementation on both cpu and cuda, need corrections and code clean up before review and testing * libnd4j: Multinomial op #8570 code clean up and fixed issues data selecting, moved from coords to tads * libnd4j: Multinomial op #8570 fixed cuda build add reference for math materials that was used for implementation * libnd4j: Multinomial op #8570 fixed several bugs, added several tests and improved cuda version. current implementation works, need testing of reproduction with the same seed * libnd4j: Multinomial op #8570 fixes and optimization after discussion in both cuda and cpu * libnd4j: Multinomial op #8570 add corrections after review, removed tads, replace 2D parallel loop by 3D Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j: Multinomial op fixed declaration and add tests need discussion * libnd4j: Multinomial op fix in test * libnd4j: Multinomial op corrected behavior to get reproducible results, fixed issue in uniform value getting, tests added, need cuda review and cuda testing Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j: Multinomial op fixed indexing on uniform calculation Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j: Multinomial op some corrections in max min declaration Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j: Multinomial op fixed index calculation, added rewind, corrected input declaration, added stats tests, both cuda and cpu. cuda need testing * libnd4j: Multinomial op fixed bugs on cuda nad cpu. need review Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j: Multinomial op corrected tests to handle different orders Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j: Multinomial op some improvements after code review Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j: Multinomial op more corrections after review Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j: Multinomial op fixed seed usage, update tests, fixed cuda based on comments, fixed bug of rewind, removed one behavior, minor corrections. Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j: Multinomial op minor corrections Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j: Multinomial op rise the bound of fluctuation for random cases Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j: Multinomial op modified operation inputs and update implementation and tests on both cpu and cuda * libnd4j: Multinomial op corrected data types according ops.proto Co-authored-by: raver119 <raver119@gmail.com>
2020-01-06 20:35:05 +01:00
auto nIndex = nSamplesPerBatch + nClassesPerSample + nClass;
auto unifornLog = sd::math::nd4j_log<Tx, Tx>(-sd::math::nd4j_log<Tx, Tx>(rng.relativeT<Tx>(nIndex, minVal, maxVal)));
Oleh multinomial (#163) * libnd4j: Multinomial op #8570 first raw step of multinomial random data generator implementation Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j: Multinomial op #8570 next step of multinomial random categories generator implementation on both cpu and cuda, need corrections and code clean up before review and testing * libnd4j: Multinomial op #8570 code clean up and fixed issues data selecting, moved from coords to tads * libnd4j: Multinomial op #8570 fixed cuda build add reference for math materials that was used for implementation * libnd4j: Multinomial op #8570 fixed several bugs, added several tests and improved cuda version. current implementation works, need testing of reproduction with the same seed * libnd4j: Multinomial op #8570 fixes and optimization after discussion in both cuda and cpu * libnd4j: Multinomial op #8570 add corrections after review, removed tads, replace 2D parallel loop by 3D Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j: Multinomial op fixed declaration and add tests need discussion * libnd4j: Multinomial op fix in test * libnd4j: Multinomial op corrected behavior to get reproducible results, fixed issue in uniform value getting, tests added, need cuda review and cuda testing Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j: Multinomial op fixed indexing on uniform calculation Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j: Multinomial op some corrections in max min declaration Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j: Multinomial op fixed index calculation, added rewind, corrected input declaration, added stats tests, both cuda and cpu. cuda need testing * libnd4j: Multinomial op fixed bugs on cuda nad cpu. need review Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j: Multinomial op corrected tests to handle different orders Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j: Multinomial op some improvements after code review Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j: Multinomial op more corrections after review Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j: Multinomial op fixed seed usage, update tests, fixed cuda based on comments, fixed bug of rewind, removed one behavior, minor corrections. Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j: Multinomial op minor corrections Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j: Multinomial op rise the bound of fluctuation for random cases Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j: Multinomial op modified operation inputs and update implementation and tests on both cpu and cuda * libnd4j: Multinomial op corrected data types according ops.proto Co-authored-by: raver119 <raver119@gmail.com>
2020-01-06 20:35:05 +01:00
Tx tValue = (xTad[nClass * xDimAstride] - unifornLog);
if (tValue > Max) {
Max = tValue;
arg = nClass;
}
}
}
}
};
samediff::Threads::parallel_for(func, 0, batchValue, 1, 0, numOfSamples, 1);
Oleh multinomial (#163) * libnd4j: Multinomial op #8570 first raw step of multinomial random data generator implementation Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j: Multinomial op #8570 next step of multinomial random categories generator implementation on both cpu and cuda, need corrections and code clean up before review and testing * libnd4j: Multinomial op #8570 code clean up and fixed issues data selecting, moved from coords to tads * libnd4j: Multinomial op #8570 fixed cuda build add reference for math materials that was used for implementation * libnd4j: Multinomial op #8570 fixed several bugs, added several tests and improved cuda version. current implementation works, need testing of reproduction with the same seed * libnd4j: Multinomial op #8570 fixes and optimization after discussion in both cuda and cpu * libnd4j: Multinomial op #8570 add corrections after review, removed tads, replace 2D parallel loop by 3D Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j: Multinomial op fixed declaration and add tests need discussion * libnd4j: Multinomial op fix in test * libnd4j: Multinomial op corrected behavior to get reproducible results, fixed issue in uniform value getting, tests added, need cuda review and cuda testing Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j: Multinomial op fixed indexing on uniform calculation Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j: Multinomial op some corrections in max min declaration Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j: Multinomial op fixed index calculation, added rewind, corrected input declaration, added stats tests, both cuda and cpu. cuda need testing * libnd4j: Multinomial op fixed bugs on cuda nad cpu. need review Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j: Multinomial op corrected tests to handle different orders Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j: Multinomial op some improvements after code review Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j: Multinomial op more corrections after review Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j: Multinomial op fixed seed usage, update tests, fixed cuda based on comments, fixed bug of rewind, removed one behavior, minor corrections. Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j: Multinomial op minor corrections Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j: Multinomial op rise the bound of fluctuation for random cases Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j: Multinomial op modified operation inputs and update implementation and tests on both cpu and cuda * libnd4j: Multinomial op corrected data types according ops.proto Co-authored-by: raver119 <raver119@gmail.com>
2020-01-06 20:35:05 +01:00
rng.rewindH(output.lengthOf()*numOfClassX);
return;
}
void fillRandomMultiNomial(LaunchContext* context, graph::RandomGenerator& rng, NDArray& input, NDArray& output, const Nd4jLong numOfSamples, const int dimC) {
BUILD_DOUBLE_SELECTOR(input.dataType(), output.dataType(), fillRandomMultiNomial_, (context, rng, input, output, numOfSamples, dimC), FLOAT_TYPES, INDEXING_TYPES);
}
}
}
2020-06-08 12:14:22 +02:00
}