296 lines
12 KiB
C++
296 lines
12 KiB
C++
/* ******************************************************************************
|
||
*
|
||
*
|
||
* 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.
|
||
*
|
||
* See the NOTICE file distributed with this work for additional
|
||
* information regarding copyright ownership.
|
||
* 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>
|
||
#include <execution/Threads.h>
|
||
#include <helpers/ConstantTadHelper.h>
|
||
|
||
namespace sd {
|
||
namespace ops {
|
||
namespace helpers {
|
||
|
||
/**
|
||
* 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) {
|
||
|
||
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);
|
||
|
||
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) {
|
||
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 {
|
||
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 ← e−λ, 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, typename Z>
|
||
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>();
|
||
Z* outputBuf = output->dataBuffer()->primaryAsT<Z>();
|
||
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 = Z(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
|
||
output->r<Z>(pos + e) = x;
|
||
}
|
||
}
|
||
}
|
||
|
||
|
||
void fillRandomPoisson(LaunchContext* context, graph::RandomGenerator& rng, NDArray* lambda, NDArray* output) {
|
||
BUILD_DOUBLE_SELECTOR(lambda->dataType(), output->dataType(), fillRandomPoisson_, (context, rng, lambda, output), FLOAT_TYPES, FLOAT_TYPES);
|
||
}
|
||
|
||
BUILD_DOUBLE_TEMPLATE(template void fillRandomPoisson_, (LaunchContext* context,
|
||
graph::RandomGenerator& rng, NDArray* lambda, NDArray* output), FLOAT_TYPES, 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++) {
|
||
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);
|
||
}
|
||
|
||
// 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) {
|
||
|
||
const Tx* x = input.bufferAsT<Tx>();
|
||
Tz* z = output.bufferAsT<Tz>();
|
||
|
||
Tx minVal = DataTypeUtils::min<Tx>();
|
||
Tx maxVal = 1.0;
|
||
|
||
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) {
|
||
|
||
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) {
|
||
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)));
|
||
Tx tValue = (xTad[nClass * xDimAstride] - unifornLog);
|
||
if (tValue > Max) {
|
||
Max = tValue;
|
||
arg = nClass;
|
||
}
|
||
}
|
||
}
|
||
}
|
||
};
|
||
|
||
samediff::Threads::parallel_for(func, 0, batchValue, 1, 0, numOfSamples, 1);
|
||
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);
|
||
}
|
||
|
||
}
|
||
}
|
||
}
|