240 lines
11 KiB
C++
240 lines
11 KiB
C++
/*******************************************************************************
|
|
* Copyright (c) 2015-2018 Skymind, Inc.
|
|
*
|
|
* This program and the accompanying materials are made available under the
|
|
* terms of the Apache License, Version 2.0 which is available at
|
|
* https://www.apache.org/licenses/LICENSE-2.0.
|
|
*
|
|
* Unless required by applicable law or agreed to in writing, software
|
|
* distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
|
|
* WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the
|
|
* License for the specific language governing permissions and limitations
|
|
* under the License.
|
|
*
|
|
* SPDX-License-Identifier: Apache-2.0
|
|
******************************************************************************/
|
|
|
|
//
|
|
// @author George A. Shulinok <sgazeos@gmail.com>, created on 4/18/2019
|
|
//
|
|
|
|
#include <ops/declarable/helpers/BarnesHutTsne.h>
|
|
#include <execution/Threads.h>
|
|
|
|
namespace nd4j {
|
|
namespace ops {
|
|
namespace helpers {
|
|
|
|
Nd4jLong barnes_row_count(const NDArray* rowP, const NDArray* colP, Nd4jLong N, NDArray& rowCounts) {
|
|
|
|
int* pRowCounts = reinterpret_cast<int*>(rowCounts.buffer());
|
|
int const* pRows = reinterpret_cast<int const*>(rowP->getBuffer());
|
|
int const* pCols = reinterpret_cast<int const*>(colP->getBuffer());
|
|
for (int n = 0; n < N; n++) {
|
|
int begin = pRows[n];//->e<int>(n);
|
|
int end = pRows[n + 1];//rowP->e<int>(n + 1);
|
|
for (int i = begin; i < end; i++) {
|
|
bool present = false;
|
|
for (int m = pRows[pCols[i]]; m < pRows[pCols[i] + 1]; m++)
|
|
if (pCols[m] == n) {
|
|
present = true;
|
|
break;
|
|
}
|
|
|
|
++pRowCounts[n];
|
|
|
|
if (!present)
|
|
++pRowCounts[pCols[i]];
|
|
}
|
|
}
|
|
NDArray numElementsArr = rowCounts.sumNumber(); //reduceAlongDimension(reduce::Sum, {});
|
|
//rowCounts.printBuffer("Row counts");
|
|
auto numElements = numElementsArr.e<Nd4jLong>(0);
|
|
return numElements;
|
|
}
|
|
// static
|
|
// void printVector(std::vector<int> const& v) {
|
|
// for (auto x: v) {
|
|
// printf("%d ", x);
|
|
// }
|
|
// printf("\n");
|
|
// fflush(stdout);
|
|
// }
|
|
|
|
template <typename T>
|
|
static void barnes_symmetrize_(const NDArray* rowP, const NDArray* colP, const NDArray* valP, Nd4jLong N, NDArray* outputRows, NDArray* outputCols, NDArray* outputVals, NDArray* rowCounts) {
|
|
//auto N = rowP->lengthOf() - 1; /// 2 + rowP->lengthOf() % 2;
|
|
//auto numElements = output->lengthOf();
|
|
//std::vector<int> symRowP = rowCounts->asVectorT<int>();//NDArrayFactory::create<int>('c', {numElements});
|
|
//NDArray symValP = NDArrayFactory::create<double>('c', {numElements});
|
|
//symRowP.insert(symRowP.begin(),0);
|
|
//symRowP(1, {0}) = *rowCounts;
|
|
int const* pRows = reinterpret_cast<int const*>(rowP->getBuffer());
|
|
int* symRowP = reinterpret_cast<int*>(outputRows->buffer());
|
|
symRowP[0] = 0;
|
|
for (int n = 0; n < N; n++)
|
|
symRowP[n + 1] = symRowP[n] + rowCounts->e<int>(n);
|
|
// outputRows->printBuffer("output rows");
|
|
|
|
int* symColP = reinterpret_cast<int*>(outputCols->buffer());
|
|
// symRowP.p(n + 1, symRowP.e(n) + rowCounts.e(n))
|
|
// outputRows->printBuffer("SymRows are");
|
|
int const* pCols = reinterpret_cast<int const*>(colP->getBuffer());
|
|
T const* pVals = reinterpret_cast<T const*>(valP->getBuffer());
|
|
T* pOutput = reinterpret_cast<T*>(outputVals->buffer());
|
|
//std::vector<int> rowCountsV = rowCounts->getBufferAsVector<int>();
|
|
std::vector<int> offset(N);// = NDArrayFactory::create<int>('c', {N});
|
|
|
|
//PRAGMA_OMP_PARALLEL_FOR_SIMD_ARGS(schedule(guided) shared(offset))
|
|
for (int n = 0; n < N; n++) {
|
|
int begin = pRows[n];
|
|
int bound = pRows[n + 1];
|
|
|
|
for (int i = begin; i < bound; i++) {
|
|
bool present = false;
|
|
int colPI = pCols[i];
|
|
int start = pRows[colPI];
|
|
int end = pRows[colPI + 1];
|
|
|
|
//PRAGMA_OMP_PARALLEL_FOR_ARGS(schedule(guided) firstprivate(offset))
|
|
for (int m = start; m < end; m++) {
|
|
if (pCols[m] == n) {
|
|
present = true;
|
|
if (n <= colPI) {
|
|
symColP[symRowP[n] + offset[n]] = colPI;
|
|
symColP[symRowP[colPI] + offset[colPI]] = n;
|
|
pOutput[symRowP[n] + offset[n]] = pVals[i] + pVals[m];
|
|
pOutput[symRowP[colPI] + offset[colPI]] = pVals[i] + pVals[m];
|
|
}
|
|
}
|
|
}
|
|
|
|
// If (colP[i], n) is not present, there is no addition involved
|
|
if (!present) {
|
|
//int colPI = pCols[i];
|
|
//if (n <= colPI) {
|
|
symColP[symRowP[n] + offset[n]] = colPI;
|
|
symColP[symRowP[pCols[i]] + offset[colPI]] = n;
|
|
pOutput[symRowP[n] + offset[n]] = pVals[i];
|
|
pOutput[symRowP[colPI] + offset[colPI]] = pVals[i];
|
|
//}
|
|
|
|
}
|
|
// Update offsets
|
|
if (!present || (present && n <= colPI)) {
|
|
++offset[n];
|
|
|
|
if (colPI != n)
|
|
++offset[colPI];
|
|
}
|
|
// printVector(offset);
|
|
}
|
|
}
|
|
}
|
|
void barnes_symmetrize(const NDArray* rowP, const NDArray* colP, const NDArray* valP, Nd4jLong N, NDArray* outputRows, NDArray* outputCols, NDArray* outputVals, NDArray* rowCounts) {
|
|
|
|
// Divide the result by two
|
|
BUILD_SINGLE_SELECTOR(valP->dataType(), barnes_symmetrize_, (rowP, colP, valP, N, outputRows, outputCols, outputVals, rowCounts), NUMERIC_TYPES);
|
|
|
|
*outputVals /= 2.0;
|
|
//output->assign(symValP);
|
|
}
|
|
BUILD_SINGLE_TEMPLATE(template void barnes_symmetrize_, (const NDArray* rowP, const NDArray* colP, const NDArray* valP, Nd4jLong N, NDArray* outputRows, NDArray* outputCols, NDArray* outputVals, NDArray* rowCounts), NUMERIC_TYPES);
|
|
|
|
template <typename T>
|
|
static void barnes_edge_forces_(const NDArray* rowP, NDArray const* colP, NDArray const* valP, int N, NDArray const* data, NDArray* output) {
|
|
T const* dataP = reinterpret_cast<T const*>(data->getBuffer());
|
|
T const* vals = reinterpret_cast<T const*>(valP->getBuffer());
|
|
T* outputP = reinterpret_cast<T*>(output->buffer());
|
|
int colCount = data->columns();
|
|
|
|
|
|
// auto shift = 0;
|
|
auto rowSize = sizeof(T) * colCount;
|
|
|
|
auto func = PRAGMA_THREADS_FOR {
|
|
for (auto n = start; n < stop; n += increment) {
|
|
int s = rowP->e<int>(n);
|
|
int end = rowP->e<int>(n + 1);
|
|
int shift = n * colCount;
|
|
for (int i = s; i < end; i++) {
|
|
T const *thisSlice = dataP + colP->e<int>(i) * colCount;
|
|
T res = 1;
|
|
|
|
for (int k = 0; k < colCount; k++) {
|
|
auto tempVal = dataP[shift + k] - thisSlice[k];//thisSlice[k];
|
|
res += tempVal * tempVal;
|
|
}
|
|
|
|
res = vals[i] / res;
|
|
for (int k = 0; k < colCount; k++)
|
|
outputP[shift + k] += ((dataP[shift + k] - thisSlice[k]) * res);
|
|
}
|
|
//shift += colCount;
|
|
}
|
|
};
|
|
|
|
samediff::Threads::parallel_tad(func, 0, N);
|
|
}
|
|
|
|
void barnes_edge_forces(const NDArray* rowP, NDArray const* colP, NDArray const* valP, int N, NDArray* output, NDArray const& data) {
|
|
// Loop over all edges in the graph
|
|
BUILD_SINGLE_SELECTOR(output->dataType(), barnes_edge_forces_, (rowP, colP, valP, N, &data, output), FLOAT_TYPES);
|
|
}
|
|
BUILD_SINGLE_TEMPLATE(template void barnes_edge_forces_, (const NDArray* rowP, NDArray const* colP, NDArray const* valP, int N, NDArray const* data, NDArray* output), FLOAT_TYPES);
|
|
|
|
template <typename T>
|
|
static void barnes_gains_(NDArray* input, NDArray* gradX, NDArray* epsilon, NDArray* output) {
|
|
// gains = gains.add(.2).muli(sign(yGrads)).neq(sign(yIncs)).castTo(Nd4j.defaultFloatingPointType())
|
|
// .addi(gains.mul(0.8).muli(sign(yGrads)).neq(sign(yIncs)));
|
|
auto gainsInternal = LAMBDA_TTT(x, grad, eps) {
|
|
// return T((x + 2.) * nd4j::math::nd4j_sign<T,T>(grad) != nd4j::math::nd4j_sign<T,T>(eps)) + T(x * 0.8 * nd4j::math::nd4j_sign<T,T>(grad) != nd4j::math::nd4j_sign<T,T>(eps));
|
|
//return T((x + 2.) * nd4j::math::nd4j_sign<T,T>(grad) == nd4j::math::nd4j_sign<T,T>(eps)) + T(x * 0.8 * nd4j::math::nd4j_sign<T,T>(grad) == nd4j::math::nd4j_sign<T,T>(eps));
|
|
T res = nd4j::math::nd4j_sign<T,T>(grad) != nd4j::math::nd4j_sign<T,T>(eps) ? x + T(.2) : x * T(.8);
|
|
if(res < .01) res = .01;
|
|
return res;
|
|
};
|
|
|
|
input->applyTriplewiseLambda<T>(gradX, epsilon, gainsInternal, output);
|
|
}
|
|
|
|
void barnes_gains(NDArray* input, NDArray* gradX, NDArray* epsilon, NDArray* output) {
|
|
// gains = gains.add(.2).muli(sign(yGrads)).neq(sign(yIncs)).castTo(Nd4j.defaultFloatingPointType())
|
|
// .addi(gains.mul(0.8).muli(sign(yGrads)).neq(sign(yIncs)));
|
|
BUILD_SINGLE_SELECTOR(input->dataType(), barnes_gains_, (input, gradX, epsilon, output), NUMERIC_TYPES);
|
|
// auto signGradX = *gradX;
|
|
// auto signEpsilon = *epsilon;
|
|
// gradX->applyTransform(transform::Sign, &signGradX, nullptr);
|
|
// epsilon->applyTransform(transform::Sign, &signEpsilon, nullptr);
|
|
// auto leftPart = (*input + 2.) * signGradX;
|
|
// auto leftPartBool = NDArrayFactory::create<bool>(leftPart.ordering(), leftPart.getShapeAsVector());
|
|
//
|
|
// leftPart.applyPairwiseTransform(pairwise::NotEqualTo, &signEpsilon, &leftPartBool, nullptr);
|
|
// auto rightPart = *input * 0.8 * signGradX;
|
|
// auto rightPartBool = NDArrayFactory::create<bool>(rightPart.ordering(), rightPart.getShapeAsVector());
|
|
// rightPart.applyPairwiseTransform(pairwise::NotEqualTo, &signEpsilon, &rightPartBool, nullptr);
|
|
// leftPart.assign(leftPartBool);
|
|
// rightPart.assign(rightPartBool);
|
|
// leftPart.applyPairwiseTransform(pairwise::Add, &rightPart, output, nullptr);
|
|
|
|
}
|
|
BUILD_SINGLE_TEMPLATE(template void barnes_gains_, (NDArray* input, NDArray* gradX, NDArray* epsilon, NDArray* output), NUMERIC_TYPES);
|
|
|
|
bool cell_contains(NDArray* corner, NDArray* width, NDArray* point, Nd4jLong dimension) {
|
|
auto cornerMinusWidth = *corner - *width;
|
|
auto cornerPlusWidth = *corner + *width;
|
|
|
|
for (Nd4jLong i = 0; i < dimension; i++) {
|
|
if (cornerMinusWidth.e<double>(i) > point->e<double>(i))
|
|
return false;
|
|
if (cornerPlusWidth.e<double>(i) < point->e<double>(i))
|
|
return false;
|
|
}
|
|
|
|
return true;
|
|
}
|
|
}
|
|
}
|
|
}
|
|
|