Brutex/cavis
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Shyrma transpose (#244)

Browse Source 
* - provide contiguous strides for ouput in transpose op

Signed-off-by: Yurii <iuriish@yahoo.com>

* - provide contiguous strides for output in permute op

Signed-off-by: Yurii <iuriish@yahoo.com>

* - take into account empty shapes properly in transpose/permute op

Signed-off-by: Yurii <iuriish@yahoo.com>
This commit is contained in:
Yurii Shyrma 2020-02-17 07:04:28 +02:00 committed by GitHub
parent 9e3c1b02b1
commit 011c272fde
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
10 changed files with 374 additions and 587 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

16
libnd4j/blas/NDArray.hpp
View File

@ -1976,7 +1976,7 @@ bool NDArray::permutei(const std::initializer_list<int>& dimensions) {
////////////////////////////////////////////////////////////////////////// //////////////////////////////////////////////////////////////////////////
bool NDArray::permutei(const std::vector<int>& dimensions) { bool NDArray::permutei(const std::vector<int>& dimensions) {
return permutei(dimensions.data(), dimensions.size()); return permutei(dimensions.data(), rankOf());
} }
////////////////////////////////////////////////////////////////////////// //////////////////////////////////////////////////////////////////////////
@ -1998,7 +1998,7 @@ bool NDArray::permutei(const std::vector<Nd4jLong>& dimensions) {
for (int e = 0; e < dimensions.size(); e++) for (int e = 0; e < dimensions.size(); e++)
ivec[e] = dimensions[e]; ivec[e] = dimensions[e];
return permutei(ivec.data(), ivec.size()); return permutei(ivec.data(), rankOf());
} }
////////////////////////////////////////////////////////////////////////// //////////////////////////////////////////////////////////////////////////
@ -2034,9 +2034,8 @@ NDArray NDArray::permute(const Nd4jLong* dimensions, const int rank) && {
////////////////////////////////////////////////////////////////////////// //////////////////////////////////////////////////////////////////////////
NDArray NDArray::permute(const std::vector<int>& dimensions) const &{ NDArray NDArray::permute(const std::vector<int>& dimensions) const &{
auto data = dimensions.data();
auto size = dimensions.size(); return permute(dimensions.data(), rankOf());
return permute(data, size);
} }
////////////////////////////////////////////////////////////////////////// //////////////////////////////////////////////////////////////////////////
@ -2048,7 +2047,8 @@ NDArray NDArray::permute(const std::vector<int>& dimensions) && {
////////////////////////////////////////////////////////////////////////// //////////////////////////////////////////////////////////////////////////
NDArray NDArray::permute(const std::vector<Nd4jLong>& dimensions) const & { NDArray NDArray::permute(const std::vector<Nd4jLong>& dimensions) const & {
return permute(dimensions.data(), dimensions.size());
return permute(dimensions.data(), rankOf());
} }
////////////////////////////////////////////////////////////////////////// //////////////////////////////////////////////////////////////////////////
@ -2111,12 +2111,12 @@ void NDArray::permute(const Nd4jLong *dimensions, const int rank, NDArray& targe
////////////////////////////////////////////////////////////////////////// //////////////////////////////////////////////////////////////////////////
void NDArray::permute(const std::vector<int>& dimensions, NDArray& target) const { void NDArray::permute(const std::vector<int>& dimensions, NDArray& target) const {
permute(dimensions.data(), dimensions.size(), target); permute(dimensions.data(), rankOf(), target);
} }
////////////////////////////////////////////////////////////////////////// //////////////////////////////////////////////////////////////////////////
void NDArray::permute(const std::vector<Nd4jLong>& dimensions, NDArray& target) const { void NDArray::permute(const std::vector<Nd4jLong>& dimensions, NDArray& target) const {
permute(dimensions.data(), dimensions.size(), target); permute(dimensions.data(), rankOf(), target);
} }
////////////////////////////////////////////////////////////////////////// //////////////////////////////////////////////////////////////////////////

6
libnd4j/include/helpers/ShapeUtils.h
View File

@ -50,11 +50,13 @@ namespace nd4j {
static std::vector<Nd4jLong> evalRepeatShape(int axis, const std::vector<int>& repeats, const NDArray& arr); static std::vector<Nd4jLong> evalRepeatShape(int axis, const std::vector<int>& repeats, const NDArray& arr);
// evaluate shapeInfo of permuted array // evaluate shapeInfo of permuted array
static Nd4jLong* evalPermShapeInfo(const int* dimensions, const int rank, const NDArray& arr, nd4j::memory::Workspace* workspace); // if setContigStrides = true, then set contiguous strides in output shapeInfo in accordance with arr order
static Nd4jLong* evalPermShapeInfo(const int* dimensions, const int rank, const NDArray& arr, nd4j::memory::Workspace* workspace, const bool setContigStrides = false);
static Nd4jLong* evalPermShapeInfo(const Nd4jLong* dimensions, const int rank, const NDArray& arr, nd4j::memory::Workspace* workspace); static Nd4jLong* evalPermShapeInfo(const Nd4jLong* dimensions, const int rank, const NDArray& arr, nd4j::memory::Workspace* workspace);
// evaluate shapeInfo of transposed array // evaluate shapeInfo of transposed array
static Nd4jLong* evalTranspShapeInfo(const NDArray& arr, nd4j::memory::Workspace* workspace); // if setContigStrides = true, then set contiguous strides in output shapeInfo in accordance with arr order
static Nd4jLong* evalTranspShapeInfo(const NDArray& arr, nd4j::memory::Workspace* workspace, const bool setContigStrides = false);
static bool copyVectorPart(std::vector<int>& target, std::vector<int>& source, int rank, int offset); static bool copyVectorPart(std::vector<int>& target, std::vector<int>& source, int rank, int offset);

45
libnd4j/include/helpers/impl/ShapeUtils.cpp
View File

@ -313,32 +313,37 @@ std::vector<Nd4jLong> ShapeUtils::evalRepeatShape(int axis, const std::vector<in
return outShape; return outShape;
} }
////////////////////////////////////////////////////////////////////////// //////////////////////////////////////////////////////////////////////////
// evaluate shapeInfo of permuted array // evaluate shapeInfo of permuted array
Nd4jLong* ShapeUtils::evalPermShapeInfo(const int* dimensions, const int rank, const NDArray& arr, nd4j::memory::Workspace* workspace) { Nd4jLong* ShapeUtils::evalPermShapeInfo(const int* dimensions, const int rank, const NDArray& arr, nd4j::memory::Workspace* workspace, const bool setContigStrides) {
if (!arr.nonNull()) if (!arr.nonNull())
throw std::runtime_error("ShapeUtils::evalPermShapeInfo static method: wrong arguments: array is nullptr!"); throw std::runtime_error("ShapeUtils::evalPermShapeInfo static method: wrong arguments: array is nullptr!");
if (rank != arr.rankOf()) if (rank != arr.rankOf())
throw std::runtime_error("ShapeUtils::evalPermShapeInfo static method: wrong arguments: rank is not suitable!"); throw std::runtime_error("ShapeUtils::evalPermShapeInfo static method: wrong arguments: rank is not suitable!");
auto shapeInfoLength = shape::shapeInfoLength(rank); auto shapeInfoLength = shape::shapeInfoLength(rank);
// allocate memory for new array - shapeInfo
Nd4jLong *shapeInfoNew = nullptr; // allocate memory for new array - shapeInfo
ALLOCATE(shapeInfoNew, workspace, shapeInfoLength, Nd4jLong); Nd4jLong *shapeInfoNew = nullptr;
// copy arr _shapeInfo into new array ALLOCATE(shapeInfoNew, workspace, shapeInfoLength, Nd4jLong);
memcpy(shapeInfoNew, arr.getShapeInfo(), shape::shapeInfoByteLength(rank));
// perform buffer permutation
shape::doPermuteShapeInfo(shapeInfoNew, dimensions);
ShapeDescriptor descriptor(shapeInfoNew); // copy arr _shapeInfo into new array
RELEASE(shapeInfoNew, workspace); memcpy(shapeInfoNew, arr.getShapeInfo(), shape::shapeInfoByteLength(rank));
return ConstantShapeHelper::getInstance()->bufferForShapeInfo(descriptor).primaryAsT<Nd4jLong>();
}
// perform buffer permutation
shape::doPermuteShapeInfo(shapeInfoNew, dimensions, arr.lengthOf());
if(setContigStrides)
shape::updateStrides(shapeInfoNew, arr.ordering());
ShapeDescriptor descriptor(shapeInfoNew);
RELEASE(shapeInfoNew, workspace);
return ConstantShapeHelper::getInstance()->bufferForShapeInfo(descriptor).primaryAsT<Nd4jLong>();
}
////////////////////////////////////////////////////////////////////////// //////////////////////////////////////////////////////////////////////////
// evaluate shapeInfo of permuted array // evaluate shapeInfo of permuted array
@ -350,14 +355,14 @@ std::vector<Nd4jLong> ShapeUtils::evalRepeatShape(int axis, const std::vector<in
////////////////////////////////////////////////////////////////////////// //////////////////////////////////////////////////////////////////////////
// evaluate shapeInfo of transposed array // evaluate shapeInfo of transposed array
Nd4jLong* ShapeUtils::evalTranspShapeInfo(const NDArray& arr, nd4j::memory::Workspace* workspace) { Nd4jLong* ShapeUtils::evalTranspShapeInfo(const NDArray& arr, nd4j::memory::Workspace* workspace, const bool setContigStrides) {
int rank = arr.rankOf(); int rank = arr.rankOf();
std::vector<int> dimensions(rank); std::vector<int> dimensions(rank);
for (int i = 0; i < rank; ++i) for (int i = 0; i < rank; ++i)
dimensions[i] = rank - 1 - i; dimensions[i] = rank - 1 - i;
return evalPermShapeInfo(dimensions.data(), dimensions.size(), arr, workspace); return evalPermShapeInfo(dimensions.data(), dimensions.size(), arr, workspace, setContigStrides);
} }
////////////////////////////////////////////////////////////////////////// //////////////////////////////////////////////////////////////////////////

117
libnd4j/include/ops/declarable/generic/shape/permute.cpp
View File

@ -15,7 +15,8 @@
******************************************************************************/ ******************************************************************************/
// //
// Created by raver119 on 29/10/17. // @author raver119@gmail.com
// @author Yurii Shyrma (iuriish@yahoo.com)
// //
#include <op_boilerplate.h> #include <op_boilerplate.h>
@ -29,80 +30,52 @@ namespace nd4j {
////////////////////////////////////////////////////////////////////////// //////////////////////////////////////////////////////////////////////////
// here iArgs is int vector of ordered set of dimensions to be permuted // here iArgs is int vector of ordered set of dimensions to be permuted
CUSTOM_OP_IMPL(permute, 1, 1, true, 0, -2) { CUSTOM_OP_IMPL(permute, 1, 1, true, 0, -2) {
auto x = INPUT_VARIABLE(0);
bool replace = false; auto x = INPUT_VARIABLE(0);
auto z = OUTPUT_VARIABLE(0);
auto origArgs = block.width() > 1 ? INPUT_VARIABLE(1)->asVectorT<int>() : *block.getIArguments(); if (x->isEmpty()) {
std::vector<int> arguments({}); REQUIRE_TRUE(z->isEmpty(), 0, "PERMUTE OP: when input is empty, output must also be empty");
if(origArgs.size() > 0){ return Status::OK(); //No op
for (int e = 0; e < origArgs.size(); e++) {
int ax = origArgs[e];
if (ax < 0)
ax += x->rankOf();
arguments.emplace_back(ax);
}
replace = true;
} else {
for (int e = x->rankOf() - 1; e >= 0; e--)
arguments.emplace_back(e);
}
// 0D edge case
if (x->rankOf() == 0) {
REQUIRE_TRUE(arguments.size() == 1, 0, "Permute: only one axis is allowed for scalar");
auto output = OUTPUT_VARIABLE(0);
if (!block.isInplace())
output->assign(x);
return Status::OK();
}
if(block.isInplace()) { // in-place
x->permutei(arguments);
STORE_RESULT(x);
} else {
auto output = OUTPUT_VARIABLE(0);
auto result = x->permute(arguments);
output->assign(result);
STORE_RESULT(output);
}
return Status::OK();
}
DECLARE_TYPES(permute) {
getOpDescriptor()
->setAllowedInputTypes(0, nd4j::DataType::ANY)
->setAllowedInputTypes(1, {ALL_INTS})
->setSameMode(true);
}
DECLARE_SHAPE_FN(permute) {
auto shapeList = SHAPELIST();
auto arguments = block.width() > 1 ? INPUT_VARIABLE(1)->asVectorT<int>() : *block.getIArguments();
if (shape::rank(inputShape->at(0)) == 0) {
shapeList->push_back(ConstantShapeHelper::getInstance()->scalarShapeInfo(ArrayOptions::dataType(inputShape->at(0))));
} else if (inputShape->size() == 1 && !arguments.empty()) {
shapeList->push_back(ShapeUtils::evalPermShapeInfo(arguments.data(), arguments.size(), *INPUT_VARIABLE(0), block.workspace()));
} else {
if(arguments.size() == 0){
//Reverse dimensions
int rank = shape::rank(inputShape->at(0));
for (int e = rank - 1; e >= 0; e--)
arguments.emplace_back(e);
}
shapeList->push_back(ShapeUtils::evalPermShapeInfo(arguments.data(), arguments.size(), *INPUT_VARIABLE(0), block.workspace()));
}
return shapeList;
}
} }
if (block.width() == 1 && block.getIArguments()->size() == 0) {
z->assign(x->transpose());
return Status::OK();
}
std::vector<int> permutationVector = block.width() > 1 ? INPUT_VARIABLE(1)->asVectorT<int>() : *block.getIArguments();
z->assign(x->permute(permutationVector));
return Status::OK();
}
//////////////////////////////////////////////////////////////////////////
DECLARE_TYPES(permute) {
getOpDescriptor()
->setAllowedInputTypes(0, nd4j::DataType::ANY)
->setAllowedInputTypes(1, {ALL_INTS})
->setSameMode(true);
}
//////////////////////////////////////////////////////////////////////////
DECLARE_SHAPE_FN(permute) {
auto x = INPUT_VARIABLE(0);
if (block.width() == 1 && block.getIArguments()->size() == 0)
return SHAPELIST(ShapeUtils::evalTranspShapeInfo(*x, block.workspace(), true));
std::vector<int> permutationVector = block.width() > 1 ? INPUT_VARIABLE(1)->asVectorT<int>() : *block.getIArguments();
auto outputShapeInfo = ShapeUtils::evalPermShapeInfo(permutationVector.data(), x->rankOf(), *x, block.workspace(), true);
return SHAPELIST(outputShapeInfo);
}
}
} }
#endif #endif

446
libnd4j/include/ops/declarable/generic/shape/reshape.cpp
View File

@ -24,254 +24,240 @@
#include <ops/declarable/CustomOperations.h> #include <ops/declarable/CustomOperations.h>
namespace nd4j { namespace nd4j {
namespace ops { namespace ops {
//////////////////////////////////////////////////////////////////////////
// here iArgs is a vector with (optional) negative of order as first element:
// ({-order, dim1, dim2, dim3, ...})
CUSTOM_OP_IMPL(reshape, 1, 1, false, 0, -2) {
auto x = INPUT_VARIABLE(0);
if (block.width() == 1) { //////////////////////////////////////////////////////////////////////////
auto arguments = block.getIArguments(); // here iArgs is a vector with (optional) negative of order as first element:
int argsSize = arguments->size(); // ({-order, dim1, dim2, dim3, ...})
CUSTOM_OP_IMPL(reshape, 1, 1, false, 0, -2) {
//Special case: empty.reshape(<other empty shape>) -> return empty auto x = INPUT_VARIABLE(0);
if (x->isEmpty()) { auto z = OUTPUT_VARIABLE(0);
REQUIRE_TRUE(OUTPUT_VARIABLE(0)->isEmpty(), 0, "Reshape: when input is empty, output must also be empty");
return ND4J_STATUS_OK; //No op //Special case: empty.reshape(<other empty shape>) -> return empty
if (x->isEmpty()) {
REQUIRE_TRUE(z->isEmpty(), 0, "Reshape: when input is empty, output must also be empty");
return Status::OK(); //No op
}
if (block.width() == 1) {
auto arguments = block.getIArguments();
int argsSize = arguments->size();
int e = 1;
char order = (char) -(*arguments)[0];
if (order != 'c' && order != 'f') {
order = 'c'; //x->ordering();
e = 0;
}
REQUIRE_TRUE(argsSize - e >= 1, 0, "Reshape arguments should have at least 1 dimension");
std::vector<Nd4jLong> shapeNew;
int e2 = e;
for (; e < (int) arguments->size(); e++) {
if (arguments->at(e) == -1){
Nd4jLong shapeLength = 1;
for(; e2 < e; e2++){
shapeLength *= arguments->at(e2);
} }
for(e2 = e + 1; e2 < arguments->size(); e2++){
int e = 1; shapeLength *= arguments->at(e2);
char order = (char) -(*arguments)[0];
if (order != 'c' && order != 'f') {
order = 'c'; //x->ordering();
e = 0;
}
REQUIRE_TRUE(argsSize - e >= 1, 0, "Reshape arguments should have at least 1 dimension");
std::vector<Nd4jLong> shapeNew;
int e2 = e;
for (; e < (int) arguments->size(); e++) {
if (arguments->at(e) == -1){
Nd4jLong shapeLength = 1;
for(; e2 < e; e2++){
shapeLength *= arguments->at(e2);
}
for(e2 = e + 1; e2 < arguments->size(); e2++){
shapeLength *= arguments->at(e2);
}
Nd4jLong realShape = x->lengthOf() / shapeLength;
shapeNew.push_back(realShape);
}
else{
shapeNew.push_back(arguments->at(e));
}
}
auto len = shape::prodLong(shapeNew.data(), shapeNew.size());
REQUIRE_TRUE(len == x->lengthOf(), 0, "Reshape: lengths before and after reshape should match, but got %i vs %i", x->lengthOf(), len);
if (Environment::getInstance()->isDebugAndVerbose()) {
nd4j_printv("Reshape: new shape", shapeNew);
}
if (block.isInplace()) {
if (x->reshapei(order, shapeNew)) {
STORE_RESULT(*x);
return ND4J_STATUS_OK;
}
} else {
auto ret = OUTPUT_VARIABLE(0);
auto xr = x->reshape(order, shapeNew);
ret->assign(xr);
STORE_RESULT(*ret);
return Status::OK();
}
} else if (block.width() == 2) {
auto s = INPUT_VARIABLE(1);
//Special case: empty.reshape(-1) -> return empty
if (x->isEmpty()) {
//REQUIRE_TRUE(s->lengthOf() == 1 && s->e<Nd4jLong>(0) == -1, 0, "Reshape: when input is empty, shape must be [-1]");
REQUIRE_TRUE(OUTPUT_VARIABLE(0)->isEmpty(), 0, "Reshape: when input is empty, output must also be empty");
return Status::OK(); //No op
}
char order = 'c';
if (block.numI() > 0)
order = (char) -INT_ARG(0);
std::vector<Nd4jLong> shapeNew(s->lengthOf());
for (int e = 0; e < (int) s->lengthOf(); e++) {
auto dim = s->e<Nd4jLong >(e);
if (dim == -1){
Nd4jLong shapeLength = 1;
for(int e2 = 0; e2 < e; e2++){
shapeLength *= s->e<Nd4jLong>(e2);
}
for(int e2 = e + 1; e2 < (int) s->lengthOf(); e2++){
REQUIRE_TRUE(s->e<Nd4jLong>(e2) != -1, 0, "Reshape : Only one unknown dimension (-1) is allowed.");
shapeLength *= s->e<Nd4jLong>(e2);
}
Nd4jLong realShape = x->lengthOf() / shapeLength;
shapeNew[e] = realShape;
}
else{
shapeNew[e] = dim;
}
}
if (Environment::getInstance()->isDebugAndVerbose()) {
nd4j_printv("Reshape: new shape", shapeNew);
}
if (block.isInplace()) {
if (x->reshapei(order, shapeNew)) {
STORE_RESULT(*x);
return Status::OK();
}
} else {
auto ret = OUTPUT_VARIABLE(0);
if (s->isEmpty()) {
// just a scalar
ret->assign(x);
} else {
auto xr = x->reshape(order, shapeNew);
ret->assign(xr);
}
return Status::OK();
} }
Nd4jLong realShape = x->lengthOf() / shapeLength;
shapeNew.push_back(realShape);
}
else{
shapeNew.push_back(arguments->at(e));
} }
return ND4J_STATUS_BAD_INPUT;
} }
auto len = shape::prodLong(shapeNew.data(), shapeNew.size());
REQUIRE_TRUE(len == x->lengthOf(), 0, "Reshape: lengths before and after reshape should match, but got %i vs %i", x->lengthOf(), len);
DECLARE_TYPES(reshape) { if (Environment::getInstance()->isDebugAndVerbose()) {
getOpDescriptor() nd4j_printv("Reshape: new shape", shapeNew);
->setAllowedInputTypes(0, nd4j::DataType::ANY)
->setAllowedInputTypes(1, {ALL_INTS})
->setSameMode(true);
} }
DECLARE_SHAPE_FN(reshape) { auto xr = x->reshape(order, shapeNew);
auto inp = inputShape->at(0); z->assign(xr);
STORE_RESULT(*z);
// we can launch op using Int arguments return Status::OK();
if (inputShape->size() == 1) {
REQUIRE_TRUE(block.numI() > 0, 0, "Reshape: new shape should be provided as NDArray or int arguments, but nothing was defined");
std::vector<int> *arguments = block.getIArguments();
int e = 1; } else if (block.width() == 2) {
char order = (char) -(*arguments)[0];
if (order != 'c' && order != 'f') { auto s = INPUT_VARIABLE(1);
order = shape::order(inp);
e = 0; char order = 'c';
if (block.numI() > 0)
order = (char) -INT_ARG(0);
std::vector<Nd4jLong> shapeNew(s->lengthOf());
for (int e = 0; e < (int) s->lengthOf(); e++) {
auto dim = s->e<Nd4jLong >(e);
if (dim == -1){
Nd4jLong shapeLength = 1;
for(int e2 = 0; e2 < e; e2++){
shapeLength *= s->e<Nd4jLong>(e2);
} }
for(int e2 = e + 1; e2 < (int) s->lengthOf(); e2++){
std::vector<Nd4jLong> shapeNew; REQUIRE_TRUE(s->e<Nd4jLong>(e2) != -1, 0, "Reshape : Only one unknown dimension (-1) is allowed.");
shapeLength *= s->e<Nd4jLong>(e2);
int e2 = e;
for (; e < (int) arguments->size(); e++) {
if ((int) arguments->at(e) == -1){
Nd4jLong shapeLength = 1;
for(; e2 < e; e2 ++){
shapeLength *= arguments->at(e2);
}
for(e2 = e + 1; e2 < arguments->size(); e2++){
REQUIRE_TRUE(arguments->at(e2) != -1, 0, "Reshape : Only one unknown dimension (-1) is allowed.");
shapeLength *= arguments->at(e2);
}
if(shapeLength == 0){
//Edge case for empty:
shapeNew.push_back(0);
} else {
//Standard case
Nd4jLong realShape = shape::length(inp) / shapeLength;
shapeNew.push_back(realShape);
}
}
else{
shapeNew.push_back(arguments->at(e));
}
} }
Nd4jLong realShape = x->lengthOf() / shapeLength;
return SHAPELIST(ConstantShapeHelper::getInstance()->createShapeInfo(ShapeDescriptor(ArrayOptions::dataType(inp), order, shapeNew))); shapeNew[e] = realShape;
} else { }
// or, with second input "as shape" else{
auto x = INPUT_VARIABLE(0); shapeNew[e] = dim;
auto y = INPUT_VARIABLE(1);
// special case here
if (y->isEmpty()) {
REQUIRE_TRUE(x->lengthOf() == 1, 0, "Reshape: new length doesn't match existing array");
return SHAPELIST(ConstantShapeHelper::getInstance()->scalarShapeInfo(ArrayOptions::dataType(inp)));
}
//Special case: empty.reshape(-1) -> return empty
if (x->isEmpty()) {
//REQUIRE_TRUE(y->lengthOf() == 1 && y->e<Nd4jLong>(0) == -1, 0, "Reshape: when input is empty, shape must be [-1]");
auto shapeOf = y->getBufferAsVector<Nd4jLong>();
Nd4jLong prod = 1;
bool hasNegs = false;
for (auto v:shapeOf) {
if (v < 0) {
hasNegs = true;
v = 0;
}
prod *= v;
}
REQUIRE_TRUE(prod == 0, 0, "Reshape: in case of empty arrays reshape must return empty array as well");
// if there are -1s - we turn them into zeros
if (hasNegs) {
for (int e = 0; e < shapeOf.size(); e++)
if (shapeOf[e] < 0)
shapeOf[e] = 0;
}
auto newShape = ShapeBuilders::createShapeInfo(ArrayOptions::dataType(inp), shape::order(inp), y->lengthOf(), shapeOf.data());
return SHAPELIST(CONSTANT(newShape));
}
std::vector<Nd4jLong> shapeNew(y->lengthOf());
for (int e = 0; e < (int) y->lengthOf(); e++) {
auto dim = y->e<Nd4jLong>(e);
if (dim == -1){
Nd4jLong shapeLength = 1;
for(int e2 = 0; e2 < e; e2++){
shapeLength *= y->e<Nd4jLong>(e2);
}
for(int e2 = e + 1; e2 < (int)y->lengthOf(); e2++){
REQUIRE_TRUE(y->e<Nd4jLong>(e2) != -1, 0, "Reshape : Only one unknown dimension (-1) is allowed.");
shapeLength *= y->e<Nd4jLong>(e2);
}
if(shapeLength == 0){
//Edge case for empty:
shapeNew[e] = 0;
} else {
Nd4jLong realShape = shape::length(inp) / shapeLength;
shapeNew[e] = realShape;
}
}else {
shapeNew[e] = dim;
}
}
return SHAPELIST(ConstantShapeHelper::getInstance()->createShapeInfo(ArrayOptions::dataType(inp), 'c', shapeNew));
} }
} }
if (Environment::getInstance()->isDebugAndVerbose()) {
nd4j_printv("Reshape: new shape", shapeNew);
}
if (s->isEmpty()) {
// just a scalar
z->assign(x);
} else {
auto xr = x->reshape(order, shapeNew);
z->assign(xr);
}
return Status::OK();
} }
return ND4J_STATUS_BAD_INPUT;
}
DECLARE_TYPES(reshape) {
getOpDescriptor()
->setAllowedInputTypes(0, nd4j::DataType::ANY)
->setAllowedInputTypes(1, {ALL_INTS})
->setSameMode(true);
}
DECLARE_SHAPE_FN(reshape) {
auto inp = inputShape->at(0);
// we can launch op using Int arguments
if (inputShape->size() == 1) {
REQUIRE_TRUE(block.numI() > 0, 0, "Reshape: new shape should be provided as NDArray or int arguments, but nothing was defined");
std::vector<int> *arguments = block.getIArguments();
int e = 1;
char order = (char) -(*arguments)[0];
if (order != 'c' && order != 'f') {
order = shape::order(inp);
e = 0;
}
std::vector<Nd4jLong> shapeNew;
int e2 = e;
for (; e < (int) arguments->size(); e++) {
if ((int) arguments->at(e) == -1){
Nd4jLong shapeLength = 1;
for(; e2 < e; e2 ++){
shapeLength *= arguments->at(e2);
}
for(e2 = e + 1; e2 < arguments->size(); e2++){
REQUIRE_TRUE(arguments->at(e2) != -1, 0, "Reshape : Only one unknown dimension (-1) is allowed.");
shapeLength *= arguments->at(e2);
}
if(shapeLength == 0){
//Edge case for empty:
shapeNew.push_back(0);
} else {
//Standard case
Nd4jLong realShape = shape::length(inp) / shapeLength;
shapeNew.push_back(realShape);
}
}
else{
shapeNew.push_back(arguments->at(e));
}
}
return SHAPELIST(ConstantShapeHelper::getInstance()->createShapeInfo(ShapeDescriptor(ArrayOptions::dataType(inp), order, shapeNew)));
} else {
// or, with second input "as shape"
auto x = INPUT_VARIABLE(0);
auto y = INPUT_VARIABLE(1);
// special case here
if (y->isEmpty()) {
REQUIRE_TRUE(x->lengthOf() == 1, 0, "Reshape: new length doesn't match existing array");
return SHAPELIST(ConstantShapeHelper::getInstance()->scalarShapeInfo(ArrayOptions::dataType(inp)));
}
//Special case: empty.reshape(-1) -> return empty
if (x->isEmpty()) {
//REQUIRE_TRUE(y->lengthOf() == 1 && y->e<Nd4jLong>(0) == -1, 0, "Reshape: when input is empty, shape must be [-1]");
auto shapeOf = y->getBufferAsVector<Nd4jLong>();
Nd4jLong prod = 1;
bool hasNegs = false;
for (auto v:shapeOf) {
if (v < 0) {
hasNegs = true;
v = 0;
}
prod *= v;
}
REQUIRE_TRUE(prod == 0, 0, "Reshape: in case of empty arrays reshape must return empty array as well");
// if there are -1s - we turn them into zeros
if (hasNegs) {
for (int e = 0; e < shapeOf.size(); e++)
if (shapeOf[e] < 0)
shapeOf[e] = 0;
}
auto newShape = ShapeBuilders::createShapeInfo(ArrayOptions::dataType(inp), shape::order(inp), y->lengthOf(), shapeOf.data());
return SHAPELIST(CONSTANT(newShape));
}
std::vector<Nd4jLong> shapeNew(y->lengthOf());
for (int e = 0; e < (int) y->lengthOf(); e++) {
auto dim = y->e<Nd4jLong>(e);
if (dim == -1){
Nd4jLong shapeLength = 1;
for(int e2 = 0; e2 < e; e2++){
shapeLength *= y->e<Nd4jLong>(e2);
}
for(int e2 = e + 1; e2 < (int)y->lengthOf(); e2++){
REQUIRE_TRUE(y->e<Nd4jLong>(e2) != -1, 0, "Reshape : Only one unknown dimension (-1) is allowed.");
shapeLength *= y->e<Nd4jLong>(e2);
}
if(shapeLength == 0){
//Edge case for empty:
shapeNew[e] = 0;
} else {
Nd4jLong realShape = shape::length(inp) / shapeLength;
shapeNew[e] = realShape;
}
}else {
shapeNew[e] = dim;
}
}
return SHAPELIST(ConstantShapeHelper::getInstance()->createShapeInfo(ArrayOptions::dataType(inp), 'c', shapeNew));
}
}
}
} }
#endif #endif

18
libnd4j/include/ops/declarable/generic/shape/reshape_as.cpp
View File

@ -34,12 +34,10 @@ namespace nd4j {
auto y = INPUT_VARIABLE(1); auto y = INPUT_VARIABLE(1);
auto z = OUTPUT_VARIABLE(0); auto z = OUTPUT_VARIABLE(0);
std::vector<Nd4jLong> shapeNew(y->shapeOf(), y->shapeOf() + y->rankOf());
char order = y->ordering();
if (x->reshapei(order, shapeNew)) { if (x->reshapei(y->ordering(), y->getShapeAsVector())) {
*z = *x;
STORE_RESULT(*z); z->assign(x);
return Status::OK(); return Status::OK();
} }
@ -49,14 +47,8 @@ namespace nd4j {
DECLARE_SHAPE_FN(reshapeas) { DECLARE_SHAPE_FN(reshapeas) {
auto inputShapeInfo = inputShape->at(1); return SHAPELIST(ShapeBuilders::copyShapeInfo(INPUT_VARIABLE(1)->getShapeInfo(), false, block.workspace()));
int shapeInfoLength = inputShapeInfo[0]*2 + 4; }
Nd4jLong* outputShapeInfo(nullptr);
COPY_SHAPE(inputShapeInfo, outputShapeInfo);
return SHAPELIST(CONSTANT(outputShapeInfo));
}
DECLARE_TYPES(reshapeas) { DECLARE_TYPES(reshapeas) {
getOpDescriptor() getOpDescriptor()

136
libnd4j/include/ops/declarable/generic/shape/transpose.cpp
View File

@ -15,7 +15,8 @@
******************************************************************************/ ******************************************************************************/
// //
// Created by raver119 on 29/10/17. // @author raver119@gmail.com
// @author Yurii Shyrma (iuriish@yahoo.com)
// //
#include <op_boilerplate.h> #include <op_boilerplate.h>
@ -25,113 +26,52 @@
#include <helpers/ShapeUtils.h> #include <helpers/ShapeUtils.h>
namespace nd4j { namespace nd4j {
namespace ops { namespace ops {
////////////////////////////////////////////////////////////////////////// //////////////////////////////////////////////////////////////////////////
CUSTOM_OP_IMPL(transpose, 1, 1, false, 0, 0) { CUSTOM_OP_IMPL(transpose, 1, 1, false, 0, 0) {
auto x = INPUT_VARIABLE(0);
if (block.width() == 1) {
if (block.isInplace()) {
x->transposei();
STORE_RESULT(*x);
} else {
auto output = OUTPUT_VARIABLE(0);
auto t = x->transpose();
output->assign(t);
STORE_RESULT(*output);
}
} else {
// this is tf-mode transpose, that's nd4j permute
bool replace = false;
std::vector<int> arguments(*block.getIArguments());
auto w = block.width(); auto x = INPUT_VARIABLE(0);
auto a = arguments.size(); auto z = OUTPUT_VARIABLE(0);
if (w == 2 && a == 0) { //Special case: empty.reshape(<other empty shape>) -> return empty
auto axis = INPUT_VARIABLE(1); if (x->isEmpty()) {
for (int e = 0; e < axis->lengthOf(); e++) { REQUIRE_TRUE(z->isEmpty(), 0, "TRANSPOSE OP: when input is empty, output must also be empty");
auto ax = axis->e<int>(e); return Status::OK(); //No op
if (ax < 0) }
ax += x->rankOf();
arguments.emplace_back(ax); if (block.width() == 1 && block.getIArguments()->size() == 0) {
} z->assign(x->transpose());
replace = true;
} else if (a == 0) {
for (int e = x->rankOf() - 1; e >= 0; e--)
arguments.emplace_back(e);
}
// 0D edge case
if (x->rankOf() == 0) {
REQUIRE_TRUE(arguments.size() == 1, 0, "Permute: only one axis is allowed for scalar");
auto output = OUTPUT_VARIABLE(0);
if (!block.isInplace())
output->assign(x);
return Status::OK();
}
if(block.isInplace()) { // in-place
x->permutei(arguments);
STORE_RESULT(x);
} else {
auto input = x->permute(arguments);
auto output = OUTPUT_VARIABLE(0);
output->assign(input);
}
}
return Status::OK(); return Status::OK();
} }
DECLARE_TYPES(transpose) { std::vector<int> permutationVector = block.width() > 1 ? INPUT_VARIABLE(1)->asVectorT<int>() : *block.getIArguments();
getOpDescriptor()
->setAllowedInputTypes(nd4j::DataType::ANY)
->setSameMode(true);
}
DECLARE_SHAPE_FN(transpose) { z->assign(x->permute(permutationVector));
if (block.width() == 1) {
auto outputShapeInfo = ShapeUtils::evalTranspShapeInfo(*INPUT_VARIABLE(0), block.workspace());
return SHAPELIST(outputShapeInfo);
} else {
// this is basically permute mode
auto shapeList = SHAPELIST();
auto arguments = block.getIArguments();
if (shape::rank(inputShape->at(0)) == 0) {
Nd4jLong *newshape;
ALLOCATE(newshape, block.getWorkspace(), shape::shapeInfoLength(inputShape->at(0)), Nd4jLong);
newshape[0] = 0;
newshape[1] = 0;
newshape[2] = 1;
newshape[3] = 99;
ArrayOptions::copyDataType(newshape, inputShape->at(0));
shapeList->push_back(newshape);
} else if (arguments->size() > 0 || inputShape->size() > 1) {
auto axis = arguments->size() > 0 ? *arguments : (INPUT_VARIABLE(1))->template asVectorT<int>();
auto outputShapeInfo = ShapeUtils::evalPermShapeInfo(axis.data(), axis.size(), *INPUT_VARIABLE(0), block.workspace());
shapeList->push_back(outputShapeInfo);
} else if (inputShape->size() == 2) {
// dead end
auto axis = INPUT_VARIABLE(1);
auto axisV = axis->template asVectorT<Nd4jLong>();
auto newshape = ShapeUtils::evalPermShapeInfo(axisV.data(), axisV.size(), *INPUT_VARIABLE(0), block.workspace());
shapeList->push_back(newshape);
} else {
int rank = shape::rank(inputShape->at(0));
for (int e = rank - 1; e >= 0; e--)
arguments->emplace_back(e);
auto outputShapeInfo = ShapeUtils::evalPermShapeInfo(arguments->data(), arguments->size(), *INPUT_VARIABLE(0), block.workspace()); return Status::OK();
shapeList->push_back(outputShapeInfo); }
}
DECLARE_TYPES(transpose) {
getOpDescriptor()
->setAllowedInputTypes(nd4j::DataType::ANY)
->setSameMode(true);
}
DECLARE_SHAPE_FN(transpose) {
auto x = INPUT_VARIABLE(0);
if (block.width() == 1 && block.getIArguments()->size() == 0)
return SHAPELIST(ShapeUtils::evalTranspShapeInfo(*x, block.workspace(), true));
std::vector<int> permutationVector = block.width() > 1 ? INPUT_VARIABLE(1)->asVectorT<int>() : *block.getIArguments();
auto outputShapeInfo = ShapeUtils::evalPermShapeInfo(permutationVector.data(), x->rankOf(), *x, block.workspace(), true);
return SHAPELIST(outputShapeInfo);
}
return shapeList;
}
}
} }
} }

109
libnd4j/tests_cpu/layers_tests/DeclarableOpsTests1.cpp
View File

@ -1882,36 +1882,6 @@ TEST_F(DeclarableOpsTests1, TestGemv1) {
#endif #endif
//////////////////////////////////////////////////////////////////////
TEST_F(DeclarableOpsTests1, Reshape1) {
const std::vector<Nd4jLong> xShape = {5,4,3};
const std::vector<Nd4jLong> yShape = {3,5,4};
auto x = NDArrayFactory::create_<float>('f', xShape);
auto y = NDArrayFactory::create_<float>('f', yShape);
auto variableSpace = new VariableSpace();
variableSpace->putVariable(-1, x);
auto block = new Context(1, variableSpace, true);
block->fillInputs({-1});
std::vector<int>* arguments = block->getIArguments();
arguments->push_back(-y->ordering());
arguments->push_back(3);
arguments->push_back(5);
arguments->push_back(4);
nd4j::ops::reshape reshape;
reshape.execute(block);
ASSERT_TRUE(x->isSameShape(y));
delete y;
delete block;
delete variableSpace;
}
////////////////////////////////////////////////////////////////////// //////////////////////////////////////////////////////////////////////
TEST_F(DeclarableOpsTests1, Reshape2) { TEST_F(DeclarableOpsTests1, Reshape2) {
const std::vector<Nd4jLong> xShape = {5,4,3}; const std::vector<Nd4jLong> xShape = {5,4,3};
@ -2022,37 +1992,8 @@ TEST_F(DeclarableOpsTests1, Reshape7){
////////////////////////////////////////////////////////////////////// //////////////////////////////////////////////////////////////////////
TEST_F(DeclarableOpsTests1, Transpose1) { TEST_F(DeclarableOpsTests1, Transpose1) {
auto x = NDArrayFactory::create_<float>('c', {3,5,2}); auto x = NDArrayFactory::create_<float>('c', {3,5,2});
auto exp = NDArrayFactory::create_<float>('f', {2,5,3}); auto exp = NDArrayFactory::create_<float>('c', {2,5,3});
auto variableSpace = new VariableSpace();
variableSpace->putVariable(-1, x);
auto block = new Context(1, variableSpace, true); // in-place
block->fillInputs({-1});
nd4j::ops::transpose transpose;
Nd4jStatus status = transpose.execute(block);
ASSERT_EQ(ND4J_STATUS_OK, status);
// ASSERT_TRUE(x.isSameShapeStrict(exp));
for (int e = 0; e < x->rankOf() * 2 + 2; e++) {
ASSERT_EQ(x->getShapeInfo()[e], exp->getShapeInfo()[e]);
}
// ASSERT_EQ(x.getShapeInfo()[x.rankOf() * 2 + 2],-exp.getShapeInfo()[x.rankOf() * 2 + 2]);
ASSERT_EQ(x->getShapeInfo()[x->rankOf() * 2 + 3], exp->getShapeInfo()[x->rankOf() * 2 + 3]);
delete exp;
delete block;
delete variableSpace;
}
//////////////////////////////////////////////////////////////////////
TEST_F(DeclarableOpsTests1, Transpose2) {
auto x = NDArrayFactory::create_<float>('c', {3,5,2});
auto exp = NDArrayFactory::create_<float>('f', {2,5,3});
auto variableSpace = new VariableSpace(); auto variableSpace = new VariableSpace();
variableSpace->putVariable(-1, x); variableSpace->putVariable(-1, x);
@ -2066,12 +2007,10 @@ TEST_F(DeclarableOpsTests1, Transpose2) {
ASSERT_EQ(ND4J_STATUS_OK, status); ASSERT_EQ(ND4J_STATUS_OK, status);
auto result = variableSpace->getVariable(block->getNodeId())->getNDArray(); auto result = variableSpace->getVariable(block->getNodeId())->getNDArray();
// ASSERT_TRUE(result->isSameShapeStrict(exp));
for (int e = 0; e < result->rankOf() * 2 + 2; e++) { ASSERT_TRUE(exp->isSameShape(result));
ASSERT_EQ(result->getShapeInfo()[e], exp->getShapeInfo()[e]); ASSERT_TRUE(exp->dataType() == result->dataType());
} ASSERT_TRUE(exp->ordering() == result->ordering());
//ASSERT_EQ(result->getShapeInfo()[x.rankOf() * 2 + 2],-exp.getShapeInfo()[x.rankOf() * 2 + 2]);
ASSERT_EQ(result->getShapeInfo()[x->rankOf() * 2 + 3], exp->getShapeInfo()[x->rankOf() * 2 + 3]);
delete exp; delete exp;
delete block; delete block;
@ -2079,44 +2018,12 @@ TEST_F(DeclarableOpsTests1, Transpose2) {
} }
//////////////////////////////////////////////////////////////////////
// in-place
TEST_F(DeclarableOpsTests1, Permute1) {
Nd4jLong shapeX[] = {3, 5, 10, 15, 150, 15, 1, 0, 1, 99};
Nd4jLong shapeExp[] = {3, 15, 5, 10, 1, 150, 15, 0, 0, 99};
const std::vector<int> perm = {2, 0, 1};
ArrayOptions::setDataType(shapeX, nd4j::DataType::FLOAT32);
ArrayOptions::setDataType(shapeExp, nd4j::DataType::FLOAT32);
auto x = new NDArray(shapeX,true);
auto exp = new NDArray(shapeExp,true);
auto variableSpace = new VariableSpace();
variableSpace->putVariable(-1, x);
auto block = new Context(1, variableSpace, true); // in-place
block->fillInputs({-1});
std::vector<int>* arguments = block->getIArguments();
*arguments = perm; // set dimensions to be permuted
nd4j::ops::permute permute;
Nd4jStatus status = permute.execute(block);
ASSERT_EQ(ND4J_STATUS_OK, status);
ASSERT_TRUE(x->isSameShapeStrict(*exp));
delete exp;
delete block;
delete variableSpace;
}
////////////////////////////////////////////////////////////////////// //////////////////////////////////////////////////////////////////////
// not-in-place // not-in-place
TEST_F(DeclarableOpsTests1, Permute2) { TEST_F(DeclarableOpsTests1, Permute1) {
Nd4jLong shapeX[] = {3, 5, 10, 15, 150, 15, 1, 0, 1, 99}; Nd4jLong shapeX[] = {3, 5,10,15, 150,15,1, 0,1,99};
Nd4jLong shapeExp[] = {3, 15, 5, 10, 1, 150, 15, 0, 0, 99}; Nd4jLong shapeExp[] = {3, 15,5,10, 50,10,1, 0,1,99};
const std::vector<int> perm = {2, 0, 1}; const std::vector<int> perm = {2, 0, 1};
ArrayOptions::setDataType(shapeX, nd4j::DataType::FLOAT32); ArrayOptions::setDataType(shapeX, nd4j::DataType::FLOAT32);

17
libnd4j/tests_cpu/layers_tests/EmptyTests.cpp
View File

@ -161,23 +161,6 @@ TEST_F(EmptyTests, Test_Reshape_1) {
delete result; delete result;
} }
TEST_F(EmptyTests, Test_Reshape_2) {
auto vector = NDArrayFactory::create<float>('c', {1}, {119.0f});
auto exp = NDArrayFactory::create<float>(119.0f);
auto empty = NDArrayFactory::empty_<Nd4jLong>();
nd4j::ops::reshape op;
auto result = op.evaluate({&vector, empty}, {}, {}, {}, {}, true);
ASSERT_EQ(Status::OK(), result->status());
ASSERT_EQ(exp, *result->at(0));
ASSERT_EQ(exp, vector);
delete empty;
delete result;
}
TEST_F(EmptyTests, Test_Reshape_3) { TEST_F(EmptyTests, Test_Reshape_3) {
auto x = NDArrayFactory::create<float>('c', {1, 0, 0, 2}); auto x = NDArrayFactory::create<float>('c', {1, 0, 0, 2});
auto y = NDArrayFactory::create<int>('c', {2}, {10, 0}); auto y = NDArrayFactory::create<int>('c', {2}, {10, 0});

43
libnd4j/tests_cpu/layers_tests/PlaygroundTests.cpp
View File

@ -65,7 +65,7 @@ TEST_F(PlaygroundTests, test_avx) {
nd4j_printf("Optimal level: %i; Binary level: %i;\n", ::optimalLevel(), ::binaryLevel()); nd4j_printf("Optimal level: %i; Binary level: %i;\n", ::optimalLevel(), ::binaryLevel());
} }
/*
TEST_F(PlaygroundTests, test_bert_1) { TEST_F(PlaygroundTests, test_bert_1) {
// this test will run ONLY if this model exists // this test will run ONLY if this model exists
if (nd4j::graph::getFileSize("/home/raver119/Downloads/Bert_minimal_model/bert_minimal_model.fb") < 0) if (nd4j::graph::getFileSize("/home/raver119/Downloads/Bert_minimal_model/bert_minimal_model.fb") < 0)
@ -86,15 +86,15 @@ TEST_F(PlaygroundTests, test_bert_1) {
graph->getVariableSpace()->putVariable(86,0, u); graph->getVariableSpace()->putVariable(86,0, u);
graph->getVariableSpace()->putVariable(87,0, v); graph->getVariableSpace()->putVariable(87,0, v);
/*
// validating graph now
auto status = GraphExecutioner::execute(graph);
ASSERT_EQ(Status::OK(), status);
ASSERT_TRUE(graph->getVariableSpace()->hasVariable(198));
auto array = graph->getVariableSpace()->getVariable(198)->getNDArray(); // validating graph now
ASSERT_EQ(z, *array); // auto status = GraphExecutioner::execute(graph);
*/ // ASSERT_EQ(Status::OK(), status);
// ASSERT_TRUE(graph->getVariableSpace()->hasVariable(198));
// auto array = graph->getVariableSpace()->getVariable(198)->getNDArray();
// ASSERT_EQ(z, *array);
nd4j::Environment::getInstance()->setProfiling(true); nd4j::Environment::getInstance()->setProfiling(true);
auto profile = GraphProfilingHelper::profile(graph, 1); auto profile = GraphProfilingHelper::profile(graph, 1);
@ -104,28 +104,27 @@ TEST_F(PlaygroundTests, test_bert_1) {
nd4j::Environment::getInstance()->setProfiling(false); nd4j::Environment::getInstance()->setProfiling(false);
delete profile; delete profile;
/*
std::vector<Nd4jLong> values;
for (int e = 0; e < 1; e++) { // std::vector<Nd4jLong> values;
auto timeStart = std::chrono::system_clock::now();
GraphExecutioner::execute(graph); // for (int e = 0; e < 1; e++) {
// auto timeStart = std::chrono::system_clock::now();
auto timeEnd = std::chrono::system_clock::now(); // GraphExecutioner::execute(graph);
auto outerTime = std::chrono::duration_cast<std::chrono::microseconds>(timeEnd - timeStart).count();
values.emplace_back(outerTime);
}
std::sort(values.begin(), values.end()); // auto timeEnd = std::chrono::system_clock::now();
// auto outerTime = std::chrono::duration_cast<std::chrono::microseconds>(timeEnd - timeStart).count();
// values.emplace_back(outerTime);
// }
nd4j_printf("Time: %lld us;\n", values[values.size() / 2]); // std::sort(values.begin(), values.end());
*/
// nd4j_printf("Time: %lld us;\n", values[values.size() / 2]);
delete graph; delete graph;
} }
/*
TEST_F(PlaygroundTests, test_broadcast_1) { TEST_F(PlaygroundTests, test_broadcast_1) {
int pool = 10; int pool = 10;
std::vector<NDArray*> aX(pool); std::vector<NDArray*> aX(pool);