cavis/libnd4j/include/helpers/cpu/jacobiSVD.cpp
Alex Black 1170827c18 Merge master to upstream (#7945)
* Shugeo strided slice zeros (#14)

* Modified strided_slice op to properly work with empty-like shapes.

* Fixed test for reduce_mean with empty-like input.

* [WIP] Last merge (#15)

* correct logsoftmax looss (#2)

* Small SameDiff listener fix (#4)

* Various fixes (#6)

* #7839 Fix for asXMatrix and tests

* #7866 EmbeddingSequenceLayer dtype fix + test

* #7856 SameDiff save/load stream methods

* #7859 RegressionEvaluation rank 4 fix + tests + axis configuration

* EvaluationBinary 3d/4d

* More evaluation 3d/4d tests

* #7847 Evaluation empty checks

* Small test ifx

* #7848 Fix median edge case

* Improve DL4J samediff layer tests

* [WIP] FastText wrapper implemented (#8)

* FastText implemented

* Some fixes

* Fix shapes for wordsNearest

* Validation of input vectors

* Fixes

* Fixed test

* Thread tagged

* Some tweaks

* setContextClassLoader for DeallocatorServiceThread

* Numpy format tests (#1)

* Various fixes (#11)

* #7852 SameDiff gather fix

* #7892 SameDiff placeholder to constant conversion

* #7890 validate input rank for MLN/CG init methods

* Fix broken permute shape calculation

* Permute and gather fixes

* Tests

* #7850 LogSumExp fix + test

* Handful of test fixes

* Empty arrays with non-scalar shapes (#10)

* minor rearrangements for lambdas

* empty tensors with non-scalar shapes

* numpy empty tensors with non-scalar shapes

* few more empty tweaks

* Small fixes

* conv3d signature update

* micro fix in batchnorm mkldnn

* Import fixes

* Fix

* MKL-DNN update

* Small fill fix

* fill with empty input + test

* Fixes

* Small error improvement

* Fix

* one special test

* couple of fixes for lstm

* Rewrite TFGraphMapper.getNDArrayFromTensor to be maintainable and less error prone

* Fixes

* FP16

* Unsigned

* BFloat16

* Fill op - empty tweaks

* - couple of fixes for empty arrays construction
- stack updated

* strided slice fix

* one transform test

* provide method for reducing shapeInfo in case of input array is empty

* Fixed reduceAlongDimensions to use empty input properly.

* couple of broadcast tests

* couple of tests broadcast tests + tweak to make them pass

* add check of non-empty to methods producing sub-arrays

* Fixed reshapeC with zeros in shape.

* complete empty check in reduce_... legacy ops

* Concat and cumsum/prod

* Tweak to empty shape inference on import

* add empty check to the rest of reduce legacy ops

* one more test

* correct typo in evalReduceShapeInfoEmpty

* Added tests for reduce_* ops to tests with zero shapes.

* few more tests for empty reductions

* Fixed strided_slice op with empty case and tests.

* one more empty reduction test

* Fixed strided_slice test.

* add empty check to NDArray::reshapei

* infOrMax

* empty min/max with infinity tests

* made unstack working correctly with empty arrays

* few IndexReduce tests + tweaks for empty shapes

* add test for empty concat

* few tests fixed

* Validation fix for reductions on empty shapes

* Reverse fix

* Reduction shape calc fixes

* SameDiff.generateOutputVariable: don't use shape function to determine number of outputs

* Range fix

* - NDArray constructor updated for scalars/empty arrays
- few tests fixed

* More fixes

* Empty creator fixes

* concat fix

* concat fix

* TF import tests: allow 'both all NaN' and 'both all inf' to pass

* Slice, zero fraction, and reshape fixes

* transpose, gather

* Zero fraction

* scalar cast fix

* Empty reduction axis support

* few more tests fixed

* Fixed input checks conforming with TF for concat op and tests.

* few tests fixed

* matmul scalar shape fix

* Fixed checkout for data type and scalarity with concat to allow non-empty scalars with vector concats.

* broadcast bool fix

* few more tests

* few more tests

* correct evalReduceShapeInfoEmpty

* argmax/argmin + tests

* one more empty edge case + one more test

* argmax/argmin/realdiv_bp tweaks

* empty reshape test + fix

* Helper fixes

* Small fixes

* Gather test fix

* Gather test fix

* Small fixes

* reduce scalar zero values

* scalar mean workaround

* Remove debug code

* along dim mean workaround

* one more test

* - equalsTo() tweak for empty arrays
- one more test

* broadcast tweaks

* [WIP] Fixing outstanding issues for NLP (#9)

* Avoid using not-inited objects

* Test fixed.

* Redundant method avoided for models like FastText

* KMeans++ implementation

* KMeans++ implementation

* Disable parallel execution

* KMeans++

* Tests

* Dev branch merge (#16)

* SameDiff: convertDataType and gradient check util improvements (#12)

* GradCheck util improvements

* StopGradient constructor + test

* SameDiff: Add datatype conversion

* Javadoc and add DataType.isNumerical()

* Small fix

* Fix SameDiff TF import test cases intermediate naming (workaround for bad default)

* TFGraphTestAllHelper: check intermediates in execution order

* Add missing debug listener

* [WIP] lstmBlock fix + other changes (#13)

- fixes lstmBlock issue
- changes NDArray method reshape(), permute(), transpose() by making them return instance instead of pointer
- CheckNumerics op
- fixes for ReduceBool IsInfOrNan & IsFinite

* Small test fix

* CheckNumerics op wrapper

* Fix some issues on master (#17)

* Fix DataVec test issue

* Fix issue with dl4j SameDiff output layer

* Dtype fix for lambda layers

* #7912 BertIterator dtype fix (use float32 not global default)

* [WIP] Next set of CUDA stuff (#7)

New CUDA implementations and improvements

* bad file

* Dev branch master merge (#23)

* SameDiff: convertDataType and gradient check util improvements (#12)

* GradCheck util improvements

* StopGradient constructor + test

* SameDiff: Add datatype conversion

* Javadoc and add DataType.isNumerical()

* Small fix

* Fix SameDiff TF import test cases intermediate naming (workaround for bad default)

* TFGraphTestAllHelper: check intermediates in execution order

* Add missing debug listener

* [WIP] lstmBlock fix + other changes (#13)

- fixes lstmBlock issue
- changes NDArray method reshape(), permute(), transpose() by making them return instance instead of pointer
- CheckNumerics op
- fixes for ReduceBool IsInfOrNan & IsFinite

* Small test fix

* CheckNumerics op wrapper

* Compatibility of deserialization (#18)

Signed-off-by: Alexander Stoyakin <alexander.stoyakin@gmail.com>

* SameDiff: add activation gradient checking support for debugging (#19)

* SameDiff gradient checker: first pass on activation gradient checks

* Fixes + tests for activation gradient checking

* Javadoc

* [WIP] Some nd4j data type corrections (#20)

* Adjust data type

* Set correct Data type.

* Size of proper data type.

* fix averaged cpu load (#22)

* SameDiff ops, TF import and fixes (#24)

* CheckNumerics tests + fixes + misc fixes

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Fake quant

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Fixes

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* FakeQuantWithMinMaxArgs

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* CheckNumerics fix

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Fix libnd4j ALL_INTS and ALL_FLOATS declaration (uint and bfloat types)

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Small fix

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Javadoc

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Exception tweak

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* fix

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Fix for out of scope stack allocated var use

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Ignores

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Ignore for known failing test (already logged issue)

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Merge upstream to fork (#25)

* Add thousand-separator commas to TotalParams (#7915)

* Add thousand-separator commas to TotalParams

The number of parameters can be quite large, and it would help the reading of the summary printout to have the TotalParams column & values at the bottom have thousand-separator-commas in them.

* Add thousand-separator commas to MultiLayerNetwork

Corresponding change to MultiLayerNetwork

Signed-off-by: Jxtps Jxtps <jxtps435@gmail.com>

* Update contributing and issue/PR templates (#7934)

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Fix link to AdaDelta paper (#7942)

Fix link to AdaDelta paper hosted on matthewzeiler.com

Signed-off-by: Jxtps

* Fixes, and ignores for known/logged failing issues (#7943)

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* SameDiff + DL4J/SameDiff: Multiple fixes (#28)

* #7919 HDF5 attribute buffer length fix

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* #7909 Arbiter constructor exception ux improvements

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* #7925 RNN output layer length checks

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* #7939 Add listener for validating inputs are not incorrectly modified

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* #7939 Integrate NonInplaceValidationListener into tests

* #7844 DL4J SameDiff fixes for variable minibatch size

* DL4J SameDiff fixes - ensure gradient for input placeholder is available

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Tweaks to ExternalErrorsFunction - use placeholders, make more robust

* Another fix

* More fixes

* More SameDiff/DL4J fixes

* Scope out scalar array creation in BaseScalarOp

* Remove debug code

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* [WIP] Final dev branch merge (#29)

* SameDiff: convertDataType and gradient check util improvements (#12)

* GradCheck util improvements

* StopGradient constructor + test

* SameDiff: Add datatype conversion

* Javadoc and add DataType.isNumerical()

* Small fix

* Fix SameDiff TF import test cases intermediate naming (workaround for bad default)

* TFGraphTestAllHelper: check intermediates in execution order

* Add missing debug listener

* [WIP] lstmBlock fix + other changes (#13)

- fixes lstmBlock issue
- changes NDArray method reshape(), permute(), transpose() by making them return instance instead of pointer
- CheckNumerics op
- fixes for ReduceBool IsInfOrNan & IsFinite

* Small test fix

* CheckNumerics op wrapper

* Compatibility of deserialization (#18)

Signed-off-by: Alexander Stoyakin <alexander.stoyakin@gmail.com>

* SameDiff: add activation gradient checking support for debugging (#19)

* SameDiff gradient checker: first pass on activation gradient checks

* Fixes + tests for activation gradient checking

* Javadoc

* [WIP] Some nd4j data type corrections (#20)

* Adjust data type

* Set correct Data type.

* Size of proper data type.

* fix averaged cpu load (#22)

* [WIP] Multiple dataset iterators (#27)

* Splitting dataset into arbitrary number

* Fixes

* Multiple split of iterator

* Test

* Test

* Some fixes

* signature change

* one more tweak

Signed-off-by: raver119 <raver119@gmail.com>

* one more test for sequential use of DataSetIteratorSplitter

Signed-off-by: raver119 <raver119@gmail.com>

* Fixes

* Fixes

* one more test for Alexander

Signed-off-by: raver119 <raver119@gmail.com>

* Some fixes

* Some fixes

* one more test for Alexander

Signed-off-by: raver119 <raver119@gmail.com>

* minor test fix

Signed-off-by: raver119 <raver119@gmail.com>

* Some fixes

* Some fixes

* couple of assertions tweaked

Signed-off-by: raver119 <raver119@gmail.com>

* MDS splitter test :/

Signed-off-by: raver119 <raver119@gmail.com>

* Minor refactoring

* Multi dataset

* Some fixes

* More tests

* Small number of test fixes/improvements (failures on CI) (#31)

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* [WIP] More CUDA stuff (#26)

* initial commit

Signed-off-by: raver119 <raver119@gmail.com>

* LRN BP CUDA

Signed-off-by: raver119 <raver119@gmail.com>

* less memory

Signed-off-by: raver119 <raver119@gmail.com>

* Fixed bug with crop_and_resize op helper.

* get rid of unnecessary index-calculation dunction

Signed-off-by: Yurii <yurii@skymind.io>

* Fixed sort with nth_element cuda-based helper.

* Refactored nth_element.

* Refactored nth_element op and tests.

* Modified usage of dim array with sortTad routine.

* Refactored main routine of helper for non_max_image_suppression op.

* non_max_image_suppression op helper with cuda kernel implementation. Initial revision.

* fix vol2col cuda kernel

* meh

Signed-off-by: raver119 <raver119@gmail.com>

* topK concept

Signed-off-by: raver119 <raver119@gmail.com>

* unsorted topK with scanWitdh of 1

Signed-off-by: raver119 <raver119@gmail.com>

* correct vol2col tests

* sorted/unsorted topK

Signed-off-by: raver119 <raver119@gmail.com>

* implementation and fixing col2im/col2vol

* Corrected usage flags with input/output with reverse op.

* dup is const now

Signed-off-by: raver119 <raver119@gmail.com>

* percentile op

Signed-off-by: raver119 <raver119@gmail.com>

* group tests for mapool2d

Signed-off-by: Yurii <yurii@skymind.io>

* special test for george

Signed-off-by: raver119 <raver119@gmail.com>

* less threads for sortTad

Signed-off-by: raver119 <raver119@gmail.com>

* provide conv2d for cuda

Signed-off-by: Yurii <yurii@skymind.io>

* remove auther in sort tad kernel code

Signed-off-by: Yurii <yurii@skymind.io>

* provide depthwise_conv2d for cuda

Signed-off-by: Yurii <yurii@skymind.io>

* - max_pooling_with_argmax
- null check for special use

Signed-off-by: raver119 <raver119@gmail.com>

* dts cuda

Signed-off-by: raver119 <raver119@gmail.com>

* provide sconv2d for cuda

Signed-off-by: Yurii <yurii@skymind.io>

* std cuda

Signed-off-by: raver119 <raver119@gmail.com>

* Refactored non_max_suppression op to conform TF implementation.

* Improved suppression helper.

* provide pooling3d for cuda

Signed-off-by: Yurii <yurii@skymind.io>

* minor lstm rearrangements

Signed-off-by: raver119 <raver119@gmail.com>

* more of minor lstm rearrangements

Signed-off-by: raver119 <raver119@gmail.com>

* (bi)dynamic_rnn

Signed-off-by: raver119 <raver119@gmail.com>

* templates init order

Signed-off-by: raver119 <raver119@gmail.com>

* Refactored non_max_suppression op.

* Added cuda kernel for non_max_suppression.

* CPU sort by key/value

Signed-off-by: raver119 <raver119@gmail.com>

* CPU sort TAD by key/value

Signed-off-by: raver119 <raver119@gmail.com>

* CPU sort TAD by key/value tests

Signed-off-by: raver119 <raver119@gmail.com>

* Eliminate compiler error with cuda implementation.

* - repaired gradCheck in cuda
- provide conv2d_bp for cuda

Signed-off-by: Yurii <yurii@skymind.io>

* missed signature

Signed-off-by: raver119 <raver119@gmail.com>

* provide depthwise_conv2d_bp for cuda

Signed-off-by: Yurii <yurii@skymind.io>

* Implementation of lup helper with cuda kernel. Initial commit.

* further work on backprops for convolutions

Signed-off-by: Yurii <yurii@skymind.io>

* CUDA linear sort by key/val

Signed-off-by: raver119 <raver119@gmail.com>

* CUDA tad sort by key/val

Signed-off-by: raver119 <raver119@gmail.com>

* start providing of backprop for pooling2d/3d

Signed-off-by: Yurii <yurii@skymind.io>

* Added atomicAdd for bool datatype.

* dynamic partition concept

Signed-off-by: raver119 <raver119@gmail.com>

* dynamic partition concept

Signed-off-by: raver119 <raver119@gmail.com>

* dynamic partition scalar CUDA

Signed-off-by: raver119 <raver119@gmail.com>

* important comment

Signed-off-by: raver119 <raver119@gmail.com>

* fix pooling2d/3d backprop helpers

Signed-off-by: Yurii <yurii@skymind.io>

* Added non-linear test with dynamic_partition.

* Improved test for dynamic_partition.

* dynamic_partition TAD concept

Signed-off-by: raver119 <raver119@gmail.com>

* - dynamic_partition TAD CUDA impl
- dynamic_partition TAD CPU fix

Signed-off-by: raver119 <raver119@gmail.com>

* - rewrite cpu code for usampling2d/3d
- write cuda code for usampling2d/3d

Signed-off-by: Yurii <yurii@skymind.io>

* dynamic_stitch CUDA vector case

Signed-off-by: raver119 <raver119@gmail.com>

* dynamic_stitch CUDA TAD case concept

Signed-off-by: raver119 <raver119@gmail.com>

* dynamic_stitch CUDA TAD case impl

Signed-off-by: raver119 <raver119@gmail.com>

* Added tests for dynamic_stitch 3D-4D cases.

* minor tests tweaks

Signed-off-by: raver119 <raver119@gmail.com>

* Fixed type check for dynamic stitch.

* min/max bp

Signed-off-by: raver119 <raver119@gmail.com>

* rewrite code for upsampling2d/3d cpu

Signed-off-by: Yurii <yurii@skymind.io>

* reduce min/max/norm_max bp

Signed-off-by: raver119 <raver119@gmail.com>

* lup implementation. Additional enhancements.

* provide code for upsamling2d/3d backprop

Signed-off-by: Yurii <yurii@skymind.io>

* weightedCrossEntropyWithLogits

Signed-off-by: raver119 <raver119@gmail.com>

* Fixed template math atomicMul for 64bit ints.

* Refactored dynamic_partition_bp op.

* inverseBroadcast fix

Signed-off-by: raver119 <raver119@gmail.com>

* DynamicPartitionBP test datatype fixed.

* - nd4j_atomicMul Windows fix
- cpu/NDArrayLambda.hpp excluded from CUDA

Signed-off-by: raver119 <raver119@gmail.com>
2019-06-27 18:37:04 +03:00

430 lines
13 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
******************************************************************************/
//
// Created by Yurii Shyrma on 11.01.2018
//
#include <jacobiSVD.h>
#include <hhColPivQR.h>
#include <NDArrayFactory.h>
namespace nd4j {
namespace ops {
namespace helpers {
//////////////////////////////////////////////////////////////////////////
template <typename T>
JacobiSVD<T>::JacobiSVD(const NDArray& matrix, const bool calcU, const bool calcV, const bool fullUV) {
if(matrix.rankOf() != 2 || matrix.isScalar())
throw std::runtime_error("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::nd4j_min<int>(_rows, _cols);
_calcU = calcU;
_calcV = calcV;
_fullUV = fullUV;
_s = NDArrayFactory::create(matrix.ordering(), {_diagSize, 1}, matrix.dataType(), matrix.getContext());
if(_calcU) {
if(_fullUV)
_u = NDArrayFactory::create(matrix.ordering(), {_rows, _rows}, matrix.dataType(), matrix.getContext());
else
_u = NDArrayFactory::create(matrix.ordering(), {_rows, _diagSize}, matrix.dataType(), matrix.getContext());
}
else
_u = NDArrayFactory::create(matrix.ordering(), {_rows, 1}, matrix.dataType(), matrix.getContext());
if(_calcV) {
if(_fullUV)
_v = NDArrayFactory::create(matrix.ordering(), {_cols, _cols}, matrix.dataType(), matrix.getContext());
else
_v = NDArrayFactory::create(matrix.ordering(), {_cols, _diagSize}, matrix.dataType(), matrix.getContext());
}
else
_v = NDArrayFactory::create(matrix.ordering(), {_cols, 1}, matrix.dataType(), matrix.getContext());
_m = NDArrayFactory::create(matrix.ordering(), {_diagSize, _diagSize}, matrix.dataType(), matrix.getContext());
evalData(matrix);
}
//////////////////////////////////////////////////////////////////////////
template <typename T>
void JacobiSVD<T>::mulRotationOnLeft(const int i, const int j, NDArray& block, const NDArray& rotation) {
if(i < j) {
if(j+1 > block.sizeAt(0))
throw std::runtime_error("ops::helpers::JacobiSVD mulRotationOnLeft: second arguments is out of array row range !");
auto pTemp = block({i,j+1,j-i, 0,0,0}, true, true);
auto temp = pTemp;
pTemp.assign(mmul(rotation, temp));
}
else {
if(j+1 > block.sizeAt(0) || i+1 > block.sizeAt(0))
throw std::runtime_error("ops::helpers::JacobiSVD mulRotationOnLeft: some or both integer arguments are out of array row range !");
auto temp = NDArrayFactory::create(block.ordering(), {2, block.sizeAt(1)}, block.dataType(), block.getContext());
auto row1 = block({i,i+1, 0,0}, true);
auto row2 = block({j,j+1, 0,0}, true);
auto rowTemp1 = temp({0,1, 0,0}, true);
auto rowTemp2 = temp({1,2, 0,0}, true);
rowTemp1.assign(row1);
rowTemp2.assign(row2);
temp.assign(mmul(rotation, temp));
row1.assign(rowTemp1);
row2.assign(rowTemp2);
}
}
//////////////////////////////////////////////////////////////////////////
template <typename T>
void JacobiSVD<T>::mulRotationOnRight(const int i, const int j, NDArray& block, const NDArray& rotation) {
if(i < j) {
if(j+1 > block.sizeAt(1))
throw std::runtime_error("ops::helpers::JacobiSVD mulRotationOnRight: second argument is out of array column range !");
auto pTemp = block({0,0,0, i,j+1,j-i}, true, true);
auto temp = pTemp;
pTemp.assign(mmul(temp, rotation));
}
else {
if(j+1 > block.sizeAt(1) || i+1 > block.sizeAt(1))
throw std::runtime_error("ops::helpers::JacobiSVD mulRotationOnRight: some or both integer arguments are out of array column range !");
auto temp = NDArrayFactory::create(block.ordering(), {block.sizeAt(0), 2}, block.dataType(), block.getContext());
auto col1 = block({0,0, i,i+1}, true);
auto col2 = block({0,0, j,j+1}, true);
auto colTemp1 = temp({0,0, 0,1}, true);
auto colTemp2 = temp({0,0, 1,2}, true);
colTemp1.assign(col1);
colTemp2.assign(col2);
temp.assign(mmul(temp, rotation));
col1.assign(colTemp1);
col2.assign(colTemp2);
}
}
//////////////////////////////////////////////////////////////////////////
template <typename T>
bool JacobiSVD<T>::isBlock2x2NotDiag(NDArray& block, int p, int q, T& maxElem) {
auto rotation = NDArrayFactory::create(_m.ordering(), {2, 2}, _m.dataType(), _m.getContext());
T n = math::nd4j_sqrt<T,T>(block.e<T>(p,p) * block.e<T>(p,p) + block.e<T>(q,p) * block.e<T>(q,p));
const T almostZero = DataTypeUtils::min<T>();
const T precision = DataTypeUtils::eps<T>();
if(n == (T)0.f) {
block.p(p, p, 0.f);
block.p(q, p, 0.f);
} else {
T v = block.e<T>(p, p) / n;
rotation.p(0, 0, v);
rotation.p(1,1, v);
v = block.e<T>(q,p) / n;
rotation.p(0, 1, v);
rotation.p(1,0, -rotation.template e<T>(0, 1));
mulRotationOnLeft(p, q, block, rotation);
if(_calcU) {
auto temp2 = rotation.transpose();
mulRotationOnRight(p, q, _u, temp2);
}
}
maxElem = math::nd4j_max<T>(maxElem, math::nd4j_max<T>(math::nd4j_abs<T>(block.e<T>(p,p)), math::nd4j_abs<T>(block.e<T>(q,q))));
T threshold = math::nd4j_max<T>(almostZero, precision * maxElem);
const bool condition1 = math::nd4j_abs<T>(block.e<T>(p,q)) > threshold;
const bool condition2 = math::nd4j_abs<T>(block.e<T>(q,p)) > threshold;
return condition1 || condition2;
}
//////////////////////////////////////////////////////////////////////////
template <typename T>
bool JacobiSVD<T>::createJacobiRotation(const T& x, const T& y, const T& z, NDArray& rotation) {
T denom = 2.* math::nd4j_abs<T>(y);
if(denom < DataTypeUtils::min<T>()) {
rotation.p(0,0, 1.f);
rotation.p(1,1, 1.f);
rotation.p(0,1, 0.f);
rotation.p(1,0, 0.f);
return false;
}
else {
T tau = (x-z)/denom;
T w = math::nd4j_sqrt<T,T>(tau*tau + 1.);
T t;
if(tau > (T)0.)
t = 1. / (tau + w);
else
t = 1. / (tau - w);
T sign = t > (T)0. ? 1. : -1.;
T n = 1. / math::nd4j_sqrt<T,T>(t*t + 1.f);
rotation.p(0,0, n);
rotation.p(1,1, n);
rotation.p(0,1, -sign * (y / math::nd4j_abs<T>(y)) * math::nd4j_abs<T>(t) * n);
rotation.p(1,0, -rotation.e<T>(0,1));
return true;
}
}
//////////////////////////////////////////////////////////////////////////
template <typename T>
void JacobiSVD<T>::svd2x2(const NDArray& block, int p, int q, NDArray& left, NDArray& right) {
auto m = NDArrayFactory::create(block.ordering(), {2, 2}, block.dataType(), block.getContext());
m.p<T>(0,0, block.e<T>(p,p));
m.p<T>(0,1, block.e<T>(p,q));
m.p<T>(1,0, block.e<T>(q,p));
m.p<T>(1,1, block.e<T>(q,q));
auto rotation = NDArrayFactory::create(block.ordering(), {2, 2}, block.dataType(), block.getContext());
T t = m.e<T>(0,0) + m.e<T>(1,1);
T d = m.e<T>(1,0) - m.e<T>(0,1);
if(math::nd4j_abs<T>(d) < DataTypeUtils::min<T>()) {
rotation.p(0,0, 1.f);
rotation.p(1,1, 1.f);
rotation.p(0,1, 0.f);
rotation.p(1,0, 0.f);
}
else {
T u = t / d;
T tmp = math::nd4j_sqrt<T,T>(1. + u*u);
rotation.p(0,0, u / tmp);
rotation.p(1,1, u / tmp);
rotation.p(0,1, 1.f / tmp);
rotation.p(1,0, -rotation.e<T>(0,1));
}
m.assign(mmul(rotation, m));
auto _x = m.e<T>(0,0);
auto _y = m.e<T>(0,1);
auto _z = m.e<T>(1,1);
createJacobiRotation(_x, _y, _z, right);
m.p<T>(0, 0, _x);
m.p<T>(0, 1, _y);
m.p<T>(1, 1, _z);
auto temp = right.transpose();
left.assign(mmul(rotation, temp));
}
//////////////////////////////////////////////////////////////////////////
template <typename T>
void JacobiSVD<T>::evalData(const NDArray& matrix) {
const T precision = (T)2.f * DataTypeUtils::eps<T>();
const T almostZero = DataTypeUtils::min<T>();
T scale = matrix.reduceNumber(reduce::AMax).e<T>(0);
if(scale== (T)0.f)
scale = (T)1.f;
if(_rows > _cols) {
HHcolPivQR qr(matrix / scale);
_m.assign(qr._qr({0,_cols, 0,_cols}));
_m.fillAsTriangular<T>(0., 0, 0, 'l');
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) {
auto matrixT = matrix.transpose();
HHcolPivQR qr(matrixT / scale);
_m.assign(qr._qr({0,_rows, 0,_rows}));
_m.fillAsTriangular<T>(0., 0, 0, 'l');
_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);
}
else {
_m.assign(matrix({0,_diagSize, 0,_diagSize}) / scale);
if(_calcU)
_u.setIdentity();
if(_calcV)
_v.setIdentity();
}
T maxDiagElem = 0.;
for(int i = 0; i < _diagSize; ++i) {
T current = math::nd4j_abs<T>(_m.e<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::nd4j_max<T>(almostZero, precision * maxDiagElem);
if(math::nd4j_abs<T>(_m.e<T>(p,q)) > threshold || math::nd4j_abs<T>(_m.e<T>(q,p)) > threshold){
stop = false;
// if(isBlock2x2NotDiag(_m, p, q, maxDiagElem))
{
auto rotLeft = NDArrayFactory::create(_m.ordering(), {2, 2}, _m.dataType(), _m.getContext());
auto rotRight = NDArrayFactory::create(_m.ordering(), {2, 2}, _m.dataType(), _m.getContext());
svd2x2(_m, p, q, rotLeft, rotRight);
mulRotationOnLeft(p, q, _m, rotLeft);
if(_calcU) {
auto temp = rotLeft.transpose();
mulRotationOnRight(p, q, _u, temp);
}
mulRotationOnRight(p, q, _m, rotRight);
if(_calcV)
mulRotationOnRight(p, q, _v, rotRight);
maxDiagElem = math::nd4j_max<T>(maxDiagElem, math::nd4j_max<T>(math::nd4j_abs<T>(_m.e<T>(p,p)), math::nd4j_abs<T>(_m.e<T>(q,q))));
}
}
}
}
}
for(int i = 0; i < _diagSize; ++i) {
_s.p(i, math::nd4j_abs<T>(_m.e<T>(i,i)));
if(_calcU && _m.e<T>(i,i) < (T)0.) {
auto temp = _u({0,0, i,i+1}, true);
temp.applyTransform(transform::Neg, &temp, nullptr);
}
}
_s *= scale;
for(int i = 0; i < _diagSize; i++) {
int pos = (_s({i,-1, 0,0}).indexReduceNumber(indexreduce::IndexMax, nullptr)).template e<int>(0);
T maxSingVal = _s({i,-1, 0,0}).reduceNumber(reduce::Max).template e<T>(0);
if(maxSingVal == (T)0.)
break;
if(pos) {
pos += i;
T _e0 = _s.e<T>(i);
T _e1 = _s.e<T>(pos);
_s.p(pos, _e0);
_s.p(i, _e1);
//math::nd4j_swap<T>(_s(i), _s(pos));
if(_calcU) {
auto temp1 = _u({0,0, pos,pos+1}, true);
auto temp2 = _u({0,0, i,i+1}, true);
auto temp3 = temp1;
temp1.assign(temp2);
temp2.assign(temp3);
}
if(_calcV) {
auto temp1 = _v({0,0, pos, pos+1}, true);
auto temp2 = _v({0,0, i, i+1}, true);
auto temp3 = temp1;
temp1.assign(temp2);
temp2.assign(temp3);
}
}
}
}
template class ND4J_EXPORT JacobiSVD<float>;
template class ND4J_EXPORT JacobiSVD<float16>;
template class ND4J_EXPORT JacobiSVD<bfloat16>;
template class ND4J_EXPORT JacobiSVD<double>;
}
}
}