Files
deeplearning4j--deeplearning4j/libnd4j/include/helpers/impl/jacobiSVD.cpp
T
2026-07-13 12:47:05 +08:00

504 lines
15 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
******************************************************************************/
//
// Created by Yurii Shyrma on 11.01.2018
//
#include <helpers/MmulHelper.h>
#include <helpers/hhColPivQR.h>
#include <helpers/jacobiSVD.h>
#if NOT_EXCLUDED(svd)
namespace sd {
namespace ops {
namespace helpers {
//////////////////////////////////////////////////////////////////////////
template <typename T>
JacobiSVD<T>::JacobiSVD(NDArray& matrix, const bool calcU, const bool calcV, const bool fullUV)
: _m(matrix.dataType(), matrix.getContext(), true),
_s(matrix.dataType(), matrix.getContext(), true),
_u(matrix.dataType(), matrix.getContext(), true),
_v(matrix.dataType(), matrix.getContext(), true) {
if (matrix.rankOf() != 2 || matrix.isScalar())
THROW_EXCEPTION("ops::helpers::JacobiSVD constructor: input array must be 2D matrix !");
_rows = static_cast<int>(matrix.sizeAt(0));
_cols = static_cast<int>(matrix.sizeAt(1));
_diagSize = math::sd_min<int>(_rows, _cols);
_calcU = calcU;
_calcV = calcV;
_fullUV = fullUV;
std::vector<LongType> sShape = {_diagSize,1};
_s = NDArray(matrix.ordering(),sShape, matrix.dataType(), matrix.getContext());
if (_calcU) {
std::vector<LongType> rowsShape = {_rows,_rows};
std::vector<LongType> rowsShape2 = {_rows,_diagSize};
if (_fullUV)
_u = NDArray(matrix.ordering(), rowsShape, matrix.dataType(), matrix.getContext());
else
_u = NDArray(matrix.ordering(), rowsShape2, matrix.dataType(), matrix.getContext());
} else {
std::vector<LongType> rowsShape = {_rows, 1};
_u = NDArray(matrix.ordering(), rowsShape, matrix.dataType(), matrix.getContext());
}
if (_calcV) {
if (_fullUV) {
std::vector<LongType> colsShape = {_cols, _cols};
_v = NDArray(matrix.ordering(), colsShape, matrix.dataType(), matrix.getContext());
}
else {
std::vector<LongType> shape = {_cols, _diagSize};
_v = NDArray(matrix.ordering(),shape, matrix.dataType(), matrix.getContext());
}
} else {
std::vector<LongType> vShape = {_cols, 1};
_v = NDArray(matrix.ordering(), vShape, matrix.dataType(), matrix.getContext());
}
std::vector<LongType> mShape = {_diagSize, _diagSize};
_m = NDArray(matrix.ordering(), mShape, matrix.dataType(), matrix.getContext());
evalData(matrix);
}
//////////////////////////////////////////////////////////////////////////
template <typename T>
void JacobiSVD<T>::mulRotationOnLeft(const int i, const int j, NDArray& block, NDArray& rotation) {
if (i < j) {
if (j + 1 > block.sizeAt(0))
THROW_EXCEPTION(
"ops::helpers::JacobiSVD mulRotationOnLeft: second arguments is out of array row range !");
NDArray *tempPtr = block({i, j + 1, j - i, 0, 0, 0}, true, true);
NDArray temp = *tempPtr;
NDArray *tempAssignResult = mmul(rotation, temp);
temp.assign(tempAssignResult);
delete tempAssignResult;
delete tempPtr;
} else {
if (j + 1 > block.sizeAt(0) || i + 1 > block.sizeAt(0))
THROW_EXCEPTION(
"ops::helpers::JacobiSVD mulRotationOnLeft: some or both integer arguments are out of array row range !");
std::vector<LongType> tempShape = {2, block.sizeAt(1)};
NDArray temp(block.ordering(),tempShape, block.dataType(), block.getContext());
NDArray *row1Ptr = block({i, i + 1, 0, 0}, true);
NDArray row1 = *row1Ptr;
NDArray *row2Ptr = block({j, j + 1, 0, 0}, true);
NDArray row2 = *row2Ptr;
NDArray *rowTemp1Ptr = temp({0, 1, 0, 0}, true);
NDArray rowTemp1 = *rowTemp1Ptr;
NDArray *rowTemp2Ptr = temp({1, 2, 0, 0}, true);
NDArray rowTemp2 = *rowTemp2Ptr;
rowTemp1.assign(&row1);
rowTemp2.assign(&row2);
NDArray *tempAssignResult = mmul(rotation, temp);
temp.assign(tempAssignResult);
delete tempAssignResult;
row1.assign(&rowTemp1);
row2.assign(&rowTemp2);
delete row1Ptr;
delete row2Ptr;
delete rowTemp1Ptr;
delete rowTemp2Ptr;
}
}
//////////////////////////////////////////////////////////////////////////
template <typename T>
void JacobiSVD<T>::mulRotationOnRight(const int i, const int j, NDArray& block, NDArray& rotation) {
if (i < j) {
if (j + 1 > block.sizeAt(1))
THROW_EXCEPTION(
"ops::helpers::JacobiSVD mulRotationOnRight: second argument is out of array column range !");
NDArray *tempPtr = block({0, 0, 0, i, j + 1, j - i}, true, true);
NDArray temp = *tempPtr;
NDArray *tempAssignResult = mmul(temp, rotation);
temp.assign(tempAssignResult);
delete tempAssignResult;
delete tempPtr;
} else {
if (j + 1 > block.sizeAt(1) || i + 1 > block.sizeAt(1))
THROW_EXCEPTION(
"ops::helpers::JacobiSVD mulRotationOnRight: some or both integer arguments are out of array column range !");
std::vector<LongType> tempShape = {block.sizeAt(0), 2};
NDArray temp(block.ordering(), tempShape, block.dataType(), block.getContext());
NDArray *col1Ptr = block({0, 0, i, i + 1}, true);
NDArray col1 = *col1Ptr;
NDArray *col2Ptr = block({0, 0, j, j + 1}, true);
NDArray col2 = *col2Ptr;
NDArray *colTemp1Ptr = temp({0, 0, 0, 1}, true);
NDArray colTemp1 = *colTemp1Ptr;
NDArray *colTemp2Ptr = temp({0, 0, 1, 2}, true);
NDArray colTemp2 = *colTemp2Ptr;
colTemp1.assign(&col1);
colTemp2.assign(&col2);
NDArray *tempAssignResult = mmul(temp, rotation);
temp.assign(tempAssignResult);
delete tempAssignResult;
col1.assign(&colTemp1);
col2.assign(&colTemp2);
delete col1Ptr;
delete col2Ptr;
delete colTemp1Ptr;
delete colTemp2Ptr;
}
}
//////////////////////////////////////////////////////////////////////////
template <typename T>
bool JacobiSVD<T>::isBlock2x2NotDiag(NDArray& block, int p, int q, T& maxElem) {
std::vector<LongType> shape = {2, 2};
NDArray rotation(_m.ordering(), shape, _m.dataType(), _m.getContext());
T n = math::sd_sqrt<T, T>(block.t<T>(p, p) * block.t<T>(p, p) + block.t<T>(q, p) * block.t<T>(q, p));
const T almostZero = DataTypeUtils::min_positive<T>();
const T precision = DataTypeUtils::eps<T>();
if (n == (T)0.f) {
block.r<T>(p, p) = (T)0;
block.r<T>(q, p) = (T)0;
} else {
T v = block.t<T>(p, p) / n;
rotation.r<T>(0, 0) = rotation.r<T>(1, 1) = v;
v = block.t<T>(q, p) / n;
rotation.r<T>(0, 1) = v;
rotation.r<T>(1, 0) = -rotation.template t<T>(0, 1);
mulRotationOnLeft(p, q, block, rotation);
NDArray *rotT = rotation.transpose();
if (_calcU) mulRotationOnRight(p, q, _u, *rotT);
delete rotT;
}
maxElem =
math::sd_max<T>(maxElem, math::sd_max<T>(math::sd_abs<T,T>(block.t<T>(p, p)), math::sd_abs<T,T>(block.t<T>(q, q))));
T threshold = math::sd_max<T>(almostZero, precision * maxElem);
return math::sd_abs<T,T>(block.t<T>(p, q)) > threshold || math::sd_abs<T,T>(block.t<T>(q, p)) > threshold;
}
//////////////////////////////////////////////////////////////////////////
template <typename T>
bool JacobiSVD<T>::createJacobiRotation(const T& x, const T& y, const T& z, NDArray& rotation) {
T denom = (T)(2.f) * math::sd_abs<T,T>(y);
if (denom < DataTypeUtils::min_positive<T>()) {
rotation.r<T>(0, 0) = rotation.r<T>(1, 1) = (T)1.f;
rotation.r<T>(0, 1) = rotation.r<T>(1, 0) = (T)0.f;
return false;
} else {
T tau = (x - z) / denom;
T w = math::sd_sqrt<T, T>(tau * tau + (T)1.f);
T t;
if (tau > (T)0.)
t = (T)1.f / (tau + w);
else
t = (T)1.f / (tau - w);
T sign = t > (T)0. ? (T)1.f : (T)-1.f;
T cos = (T)1.f / math::sd_sqrt<T, T>(t * t + (T)1.f);
T sin = -sign * (y / math::sd_abs<T,T>(y)) * math::sd_abs<T,T>(t) * cos;
rotation.r<T>(0, 1) = sin;
rotation.r<T>(1, 0) = -sin;
rotation.r<T>(0, 0) = rotation.r<T>(1, 1) = cos;
return true;
}
}
//////////////////////////////////////////////////////////////////////////
template <typename T>
void JacobiSVD<T>::createJacobiRotationGivens(const T& p, const T& q, NDArray& rotation) {
T cos, sin;
if (q == (T)0) {
cos = p < (T)0 ? (T)-1 : (T)1;
sin = (T)0;
} else if (p == (T)0) {
cos = (T)0;
sin = q < (T)0 ? (T)1 : (T)-1;
} else if (math::sd_abs<T,T>(p) > math::sd_abs<T,T>(q)) {
T t = q / p;
T u = math::sd_sqrt<T, T>((T)1 + t * t);
if (p < (T)0) u = -u;
cos = (T)1 / u;
sin = -t * cos;
} else {
T t = p / q;
T u = math::sd_sqrt<T, T>((T)1 + t * t);
if (q < (T)0) u = -u;
sin = -(T)1 / u;
cos = -t * sin;
}
rotation.r<T>(0, 1) = sin;
rotation.r<T>(1, 0) = -sin;
rotation.r<T>(0, 0) = rotation.r<T>(1, 1) = cos;
}
//////////////////////////////////////////////////////////////////////////
template <typename T>
void JacobiSVD<T>::svd2x2(NDArray& block, int p, int q, NDArray& left, NDArray& right) {
std::vector<LongType> shape = {2, 2};
NDArray m(block.ordering(), shape, block.dataType(), block.getContext());
m.r<T>(0, 0) = block.t<T>(p, p);
m.r<T>(0, 1) = block.t<T>(p, q);
m.r<T>(1, 0) = block.t<T>(q, p);
m.r<T>(1, 1) = block.t<T>(q, q);
NDArray rotation(block.ordering(),shape, block.dataType(), block.getContext());
T t = m.t<T>(0, 0) + m.t<T>(1, 1);
T d = m.t<T>(1, 0) - m.t<T>(0, 1);
if (math::sd_abs<T,T>(d) < DataTypeUtils::min<T>()) {
rotation.r<T>(0, 0) = rotation.r<T>(1, 1) = (T)1;
rotation.r<T>(0, 1) = rotation.r<T>(1, 0) = (T)0;
} else {
T u = t / d;
T tmp = math::sd_sqrt<T, T>((T)1.f + u * u);
rotation.r<T>(0, 0) = rotation.r<T>(1, 1) = u / tmp;
rotation.r<T>(0, 1) = (T)1.f / tmp;
rotation.r<T>(1, 0) = -rotation.t<T>(0, 1);
}
NDArray *mAssignResult = mmul(rotation, m);
m.assign(mAssignResult);
delete mAssignResult;
createJacobiRotation(m.t<T>(0, 0), m.t<T>(0, 1), m.t<T>(1, 1), right);
NDArray *rightT = right.transpose();
NDArray *leftAssignResult = mmul(rotation, *rightT);
left.assign(leftAssignResult);
delete leftAssignResult;
delete rightT;
}
//////////////////////////////////////////////////////////////////////////
template <typename T>
void JacobiSVD<T>::evalData(NDArray& matrix) {
const T precision = (T)2.f * DataTypeUtils::eps<T>();
const T almostZero = DataTypeUtils::min_positive<T>();
auto* scaleResult = matrix.reduceNumber(reduce::AMax);
T scale = scaleResult->template t<T>(0);
delete scaleResult;
if (scale < (T)1.f) scale = (T)1.f;
if (_rows > _cols) {
NDArray *scaled = matrix / scale;
HHcolPivQR qr(*scaled);
delete scaled;
NDArray qrRef = *qr._qr;
NDArray *mAssignPtr = qrRef({0, _cols, 0, _cols});
NDArray mAssign = *mAssignPtr;
_m.assign(&mAssign);
delete mAssignPtr;
_m.fillAsTriangular<T>(0., 0, 0, _m, 'l',false);
HHsequence hhSeg(qr._qr, qr._coeffs, 'u');
if (_fullUV)
hhSeg.applyTo(&_u);
else if (_calcU) {
_u.setIdentity();
hhSeg.mulLeft(&_u);
}
if (_calcV) _v.assign(qr._permut);
} else if (_rows < _cols) {
NDArray *matrixT = matrix.transpose();
NDArray *scaled = (*matrixT) / scale;
HHcolPivQR qr(*scaled);
delete scaled;
NDArray qrRef = *qr._qr;
NDArray *mAssignPtr = qrRef({0, _rows, 0, _rows});
NDArray mAssign = *mAssignPtr;
_m.assign(&mAssign);
delete mAssignPtr;
_m.fillAsTriangular<T>(0., 0, 0, _m, 'l',false);
_m.transposei();
HHsequence hhSeg(qr._qr, qr._coeffs, 'u'); // type = 'u' is not mistake here !
if (_fullUV)
hhSeg.applyTo(&_v);
else if (_calcV) {
_v.setIdentity();
hhSeg.mulLeft(&_v);
}
if (_calcU) _u.assign(qr._permut);
delete matrixT;
} else {
NDArray *mAssignPtr = matrix({0, _diagSize, 0, _diagSize});
NDArray *mAssignDiv = (*mAssignPtr) / scale;
_m.assign(mAssignDiv);
delete mAssignDiv;
delete mAssignPtr;
if (_calcU) _u.setIdentity();
if (_calcV) _v.setIdentity();
}
T maxDiagElem = static_cast<T>(0.);
for (int i = 0; i < _diagSize; ++i) {
T current = math::sd_abs<T,T>(_m.t<T>(i, i));
if (maxDiagElem < current) maxDiagElem = current;
}
bool stop = false;
while (!stop) {
stop = true;
for (int p = 1; p < _diagSize; ++p) {
for (int q = 0; q < p; ++q) {
T threshold = math::sd_max<T>(almostZero, precision * maxDiagElem);
if (math::sd_abs<T,T>(_m.t<T>(p, q)) > threshold || math::sd_abs<T,T>(_m.t<T>(q, p)) > threshold) {
stop = false;
std::vector<LongType> shape = {2, 2};
NDArray rotLeft(_m.ordering(), shape, _m.dataType(), _m.getContext());
NDArray rotRight(_m.ordering(), shape, _m.dataType(), _m.getContext());
svd2x2(_m, p, q, rotLeft, rotRight);
mulRotationOnLeft(p, q, _m, rotLeft);
NDArray *rotLeftTranspose = rotLeft.transpose();
if (_calcU) mulRotationOnRight(p, q, _u, *rotLeftTranspose);
mulRotationOnRight(p, q, _m, rotRight);
if (_calcV) mulRotationOnRight(p, q, _v, rotRight);
maxDiagElem = math::sd_max<T>(
maxDiagElem, math::sd_max<T>(math::sd_abs<T,T>(_m.t<T>(p, p)), math::sd_abs<T,T>(_m.t<T>(q, q))));
delete rotLeftTranspose;
}
}
}
}
for (int i = 0; i < _diagSize; ++i) {
_s.r<T>(i) = math::sd_abs<T,T>(_m.t<T>(i, i));
if (_calcU && _m.t<T>(i, i) < (T)0.) {
NDArray *tempPtr = _u({0, 0, i, i + 1}, true);
NDArray temp = *tempPtr;
temp.applyTransform(transform::Neg, &temp, nullptr);
delete tempPtr;
}
}
_s *= scale;
for (int i = 0; i < _diagSize; i++) {
NDArray *sSlicePtr = _s({i, -1, 0, 0});
NDArray sSlice = *sSlicePtr;
NDArray *indexNum = sSlice.indexReduceNumber(indexreduce::IndexMax, nullptr);
int pos = indexNum->template e<int>(0);
auto* maxResult = sSlice.reduceNumber(reduce::Max);
T maxSingVal = maxResult->template t<T>(0);
delete maxResult;
delete sSlicePtr;
delete indexNum;
if (maxSingVal == (T)0.) break;
if (pos) {
pos += i;
math::sd_swap<T>(_s.r<T>(i), _s.r<T>(pos));
if (_calcU) {
NDArray *temp1Ptr = _u({0, 0, pos, pos + 1}, true);
NDArray temp1 = *temp1Ptr;
NDArray *temp2Ptr = _u({0, 0, i, i + 1}, true);
NDArray temp2 = *temp2Ptr;
temp1.swapUnsafe(temp2);
delete temp1Ptr;
delete temp2Ptr;
}
if (_calcV) {
NDArray *temp1Ptr = _v({0, 0, pos, pos + 1}, true);
NDArray temp1 = *temp1Ptr;
NDArray *temp2Ptr = _v({0, 0, i, i + 1}, true);
NDArray temp2 = *temp2Ptr;
temp1.swapUnsafe(temp2);
delete temp1Ptr;
delete temp2Ptr;
}
}
}
}
BUILD_SINGLE_TEMPLATE( class JacobiSVD, , SD_FLOAT_TYPES);
} // namespace helpers
} // namespace ops
} // namespace sd
#endif