Open Source Computer Vision Library https://opencv.org/
You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
 
 
 
 
 
 

1375 lines
56 KiB

// This file is part of OpenCV project.
// It is subject to the license terms in the LICENSE file found in the top-level directory
// of this distribution and at http://opencv.org/license.html.
#include <inttypes.h>
#include <opencv2/dnn/shape_utils.hpp>
#include "../precomp.hpp"
#include "../ie_ngraph.hpp"
#include "layers_common.hpp"
#include "cpu_kernels/fast_gemm.hpp"
namespace cv
{
namespace dnn
{
static bool IsTransposeReshapeForEinsum(const std::vector<size_t>& perm,
std::vector<int> input_dims,
MatShape& new_shape) {
// As long as the dims with values > 1 stay in the same order, it's a reshape.
// Example: Shape=(1,1,1024,4096) -> perm=(2,0,3,1).
size_t last_permuted_axis = 0;
for (size_t i = 0; i < perm.size(); ++i) {
if (input_dims[perm[i]] == 1)
continue;
if (perm[i] < last_permuted_axis)
return false;
last_permuted_axis = perm[i];
}
new_shape.assign(input_dims.begin(), input_dims.end());
for (size_t i = 0; i < perm.size(); ++i) {
new_shape[i] = input_dims[perm[i]];
}
return true;
}
static Mat Transpose(
const Mat& input,
const MatShape& input_shape_override,
const std::vector<size_t> permutation)
{
int input_rank = input_shape_override.size();
CV_Assert(input_rank == permutation.size());
bool reshape = input.dims != input_rank;
Mat input_reshaped;
if(reshape){
input_reshaped = input.reshape(1, input_shape_override.size(), input_shape_override.data());
}
MatShape outputDims;
outputDims.reserve(input_rank);
for (const auto& dim : permutation)
outputDims.emplace_back(input_shape_override[dim]);
Mat output;
MatShape order(permutation.begin(), permutation.end());
cv::transposeND((reshape ? input_reshaped : input), order, output);
return output;
}
bool IsTransposeRequired(size_t input_rank, const std::vector<size_t>& permutation) {
CV_Assert(input_rank == permutation.size());
// No transpose required for scalars
if (input_rank == 0){
return false;
}
// Weeds out cases where permutation is something like [0, 1, 2] for a 3D input and so on
bool transpose_required = false;
for (size_t i = 0; i < input_rank; ++i) {
if (permutation[i] != i) {
transpose_required = true;
break;
}
}
return transpose_required;
}
bool IsTransposeRequiredForDiagonal(int dim1, int dim2, int rank) {
// If the input is 2D, we don't need a transpose
if (rank == 2)
return false;
// If the two dims are the innermost dims, no transpose is required
if ((dim1 == rank - 1 && dim2 == rank - 2) ||
(dim1 == rank - 2 && dim2 == rank - 1))
return false;
// Transpose is required
return true;
}
template <typename T>
Mat DiagonalDataAssignment(Mat input) {
int rank = input.dims;
CV_Assert(rank >= 2);
CV_Assert(input.size[rank - 1] == input.size[rank - 2]);
MatShape original_dims = shape(input);
if (rank > 3){
//reshape to 3D mat
int collapsed_size = 1;
for (int i = 0; i < rank - 2; ++i) {
collapsed_size *= input.size[i];
}
std::vector<int> reshaped_dims = {collapsed_size, input.size[rank - 2], input.size[rank - 1]};
input = input.reshape(1, reshaped_dims);
}
// Compute total number of higher-dimensional slices
int total_slices = input.size[0];
original_dims[rank - 1] = 1; // Set the last dimension to 1, as we have extracted the diagonal
Mat output = Mat(original_dims, input.type());
int inner_stride = input.size[input.dims - 1];
auto inputPtr = input.ptr<T>();
auto outputPtr = output.ptr<T>();
for (int slice = 0; slice < total_slices; ++slice) {
for (int j = 0; j < inner_stride; ++j) {
// Direct memory access using raw pointers
outputPtr[slice * inner_stride + j] = inputPtr[slice * inner_stride * inner_stride + j * inner_stride + j];
}
}
return output;
}
/* Extract the diagonal elements from the last two dimensions of the tensor.
For instance, given an input_shape of [1, 2, 3, 3]:
The flexibility in this implementation allows one to choose which of the two
last dimensions retains its value, determined by the `preserve_innermost_dim_val` parameter.
When preserve_innermost_dim_val == true:
The resulting shape is [1, 2, 1, 3], indicating the diagonal has 3 elements,
and it keeps the dimension value of the innermost dimension.
When preserve_innermost_dim_val == false:
The resulting shape is [1, 2, 3, 1], indicating the diagonal also has 3 elements,
but it retains the dimension value of the penultimate dimension. */
Mat DiagonalInnermostDims(const Mat& input, bool preserve_innermost_dim_val) {
const MatShape input_dims = shape(input);
int rank = input_dims.size();
// This is an internal method and we already have finished all validations in the calling method.
// We proceed without duplicating all validations again here.
// We have a minimalistic check here to make sure the innermost dims have the same dim value
// as the calling method may have done a transpose before calling this method
CV_CheckEQ(input.size[rank - 1], input.size[rank - 2],
"innermost dims should have the same dim value to parse the diagonal elements");
MatShape output_dims = input_dims; // Copy the original dims
if (preserve_innermost_dim_val) {
output_dims[rank - 2] = 1;
} else {
output_dims[rank - 1] = 1;
}
// TODO: hande different types
Mat output = DiagonalDataAssignment<float>(input);
if (output_dims != shape(output)){
CV_Error(Error::StsError, "Output shape does not match with calculated shape");
}
return output;
}
Mat Diagonal(const Mat& input, int dim1, int dim2)
{
const MatShape input_dims = shape(input);
int rank = input_dims.size();
if (!(rank >= 2 && dim1 != dim2 && input_dims[dim1] == input_dims[dim2])){
std::string input_dims_str = std::accumulate(std::next(input_dims.begin()), input_dims.end(), std::to_string(input_dims[0]),
[](const std::string& a, int b) {
return a + ' ' + std::to_string(b);
});
CV_Error(Error::StsError, cv::format("Cannot parse the diagonal elements along dims %d and %d for input shape %s",dim1, dim2, input_dims_str.c_str()));
}
int first_dim = std::min(dim1, dim2);
int second_dim = std::max(dim1, dim2);
Mat output;
bool preserve_innermost_dim_val = false;
bool is_transpose_required = IsTransposeRequiredForDiagonal(dim1, dim2, rank);
if (is_transpose_required)
{
std::vector<size_t> permutation(rank, 0);
int first_dim_axis = -1; // This is the axis eventually occupied by the first_dim
// If one of the diagonal dimensions is one of the 2 innermost dims, then leave it as such
// so as to avoid transpose overhead
if (first_dim == rank - 2) { // If rank - 2 is occupied by first_dim, keep it there
permutation[rank - 2] = first_dim;
first_dim_axis = rank - 2;
} else {
if (second_dim != rank - 2) { // If rank - 2 is not occupied by second_dim, then put first_dim there
permutation[rank - 2] = first_dim;
first_dim_axis = rank - 2;
} else { // If rank - 2 is occupied by second_dim, then put first_dim in rank - 1
permutation[rank - 1] = first_dim;
first_dim_axis = rank - 1;
preserve_innermost_dim_val = true; // We always want to preserve the dim value of the first_dim
}
}
// Put the second_dim in the dim not occupied by the first_dim
if (first_dim_axis != rank - 1) {
permutation[rank - 1] = second_dim;
} else {
permutation[rank - 2] = second_dim;
}
size_t iter = 0;
for (int i = 0; i < rank; ++i) {
if (i != first_dim && i != second_dim) {
permutation[iter++] = i;
}
}
// Permutate the input so that the dims from which we need the diagonal forms the innermost dims
Mat transposed = Transpose(input, input_dims, permutation);
// Parse the diagonal from the innermost dims
output = DiagonalInnermostDims(transposed, preserve_innermost_dim_val);
// Swap back the dimensions to the original axes ordering using a "reverse permutation"
// Find the "reverse" permutation
iter = 0;
std::vector<size_t> reverse_permutation(rank, 0);
for (const auto& perm : permutation) {
reverse_permutation[perm] = iter++;
}
// Permutate using the reverse permutation to get back the original axes ordering
// (Pass in CPU Transpose function here as this Diagonal method will only be used for CPU based diagonal parsing)
output = Transpose(output, shape(output), reverse_permutation);
} else {
// No transposing required
output = DiagonalInnermostDims(input, preserve_innermost_dim_val);
}
// Make copy of the output dims
MatShape output_dims = shape(output);
// Unsqueeze the reduced dim
auto iter = output_dims.begin() + second_dim;
output_dims.erase(iter);
output = output.reshape(1, output_dims);
return output;
}
/**
* Returns the index associated with the input character.
* - Returns a value between 0 and 25 for inputs in the range 'a' to 'z'.
* - Returns a value between 26 and 51 for inputs in the range 'A' to 'Z'.
* - Returns -1 for invalid input that is not in the range 'a' to 'z' or 'A' to 'Z' (the caller should handle the returned result accordingly).
*/
int letterToIndex(const char ch) {
if (ch >= 'a' && ch <= 'z') {
return static_cast<int>(ch) - 'a';
}
if (ch >= 'A' && ch <= 'Z') {
return static_cast<int>('z') + static_cast<int>(ch) - 'A';
}
// invalid character - return error value
return -1;
}
// Implementation of the Einsum layer is heavily influenced by Onnxruntime at the time of writing.
// Main logic is borrowed from onnxrutime:
// https://github.com/microsoft/onnxruntime/blob/eaea34f8e29df9fb21fab675a3a895084407f306/onnxruntime/core/providers/cpu/math/einsum_utils/einsum_compute_preprocessor.cc#L8
class LayerEinsumImpl CV_FINAL : public EinsumLayer
{
private:
Ptr<ReduceLayer> reduce;
public:
// Number of inputs and outputs of the layer
int numInputs;
// inputShapes;
std::vector<MatShape> einsumInpShapes;
// Preprocessed inputs
std::vector<Mat> preProcessedInputs;
// This is container for preporcessed inputs
std::vector<MatShape> homogenizedInputDims;
// Collect outpus dimentions
MatShape einsumOutDims; // vector to store output dimentions
// These hold equation subring, left hand side and right it of
String lhs_eq, rhs_eq, equation;
// Holds token from left hand side of the equation
std::vector<String> lhs_eq_tokens;
// Idicates if equation substring is defined in explit way such as "ij, jk->ik"
// as opposed to "ij->"
bool explicitEquation = false;
// Stores the subscript indices for each input in the equation
std::vector<std::vector<int>> inputSubscriptIndices;
// Keeps track of the input index of the last input that had the subscript label
// If the value is `-1`, it means the subscript label was never encountered or it appears in the output
std::vector<int> subscriptIndicesToLastInput;
// Holds the dimension value of the index corresponding to the subscript label
// `-1` indicates that the corresponding label was not encountered at all
std::vector<int> subscriptIndicesToDimValue;
// Index corresponding to each output dim corresponding to each subscript index
// A value of -1 means the corresponding subscript index is not found in the output
std::vector<int> subscriptIndicesToOutputIndices;
// Hold max number of alphabetic numbers
static const size_t numOfLetters = 52;
// Stores the count corresponding to each letter encountered
// A value of `0` indicates that the corresponding letter hasn't been seen at all
std::array<int, numOfLetters> letter2count;
// Hold the assigned index corresponding to the letter seen
// `-1` means the corresponding letter wasn't seen at all
std::array<int, numOfLetters> letter2index;
// Represents the count of unique subscript labels (subscript indices)
// Example 1: For the equation 'ij, jk -> ik', num_subscript_indices_ = 3 (i, j, k)
// Example 2: For the equation '...ij', 'jk' -> '...ik',
// num_subscript_indices_ = 3 (i, j, k) + number of dimensions specified by an ellipsis (across all inputs)
int numLetterIndices = 0;
// The number of dimensions that are encompassed by an "ellipsis" - "...".
size_t numOfEllipsisDims = 0;
// Backend for fastgemm
FastGemmOpt opt;
void parseEquation(String equation);
void processEquation(const std::vector<MatShape>& inputs);
void processBroadcastedDims();
void validateOutputSubscript();
void calculateOutputShape();
void preProcessInputs(InputArrayOfArrays& inputs);
Mat reduceSum(Mat& src, MatShape& reduceAxis);
Mat FinalizeOutput(const Mat& candidateOuput, const MatShape& ordered_subscript_indices_in_candidate);
Mat pairwiseOperandProcess(
const Mat& left,
const MatShape& leftShapeOverride,
const Mat& right,
const MatShape& rightShapeOverride,
const MatShape& reduceDims,
bool isFinalPair
);
Mat batchwiseMatMul(
const Mat& input1,
const MatShape& input1ShapeOverride,
const Mat& input2,
const MatShape& input2ShapeOverride
);
// constructor
LayerEinsumImpl(const LayerParams& params)
{
setParamsFrom(params);
equation = params.get<String>("equation");
int outputSize = params.get<int>("outputSize");
numInputs = params.get<int>("inputSize");
CV_CheckEQ(outputSize, 1, "Einsum layer should only have one output");
// get the input shapes from onnx importer
for (int i=0; i < numInputs; i++){
auto param = params.get("inputShapes" + cv::format("%d", i));
int inputDims = param.size();
std::vector<int> shape;
for (int i = 0; i < inputDims; ++i)
shape.emplace_back(param.get<int>(i));
einsumInpShapes.emplace_back(shape);
}
opt.init();
// Maintains a mapping between input indices and their corresponding subscript labels for each input
inputSubscriptIndices.reserve(numInputs);
// We allocate space for 10 values as a precaution,
// assuming that we won't encounter any input with a rank greater than 10.
// In such cases, the value of num_subscript_indices_ would be greater than 10.
subscriptIndicesToLastInput.reserve(10);
subscriptIndicesToDimValue.reserve(10);
// fill in vectors to avoid getting random numbers
letter2count.fill(0);
letter2index.fill(-1);
// parser equation and extract tokens from the equation
// save token to lhs_eq_tokens variable
parseEquation(equation); // TODO: return lhs_eq_tokens
// Start preprocessing related to equation parsing
// and dimention broadcasting
processEquation(einsumInpShapes);
processBroadcastedDims();
// calculate output shape
validateOutputSubscript();
calculateOutputShape();
}
virtual bool supportBackend(int backendId) CV_OVERRIDE {
return backendId == DNN_BACKEND_OPENCV ||
backendId == DNN_BACKEND_INFERENCE_ENGINE_NGRAPH;
}
// getMeoryShapes
bool getMemoryShapes(const std::vector<MatShape> &inputs,
const int requiredOutputs,
std::vector<MatShape> &outputs,
std::vector<MatShape> &internals) const CV_OVERRIDE
{
CV_UNUSED(internals);
// check if passed and parsed inputs match up in number and dimensions
CV_CheckEQ(static_cast<int>(inputs.size()), numInputs,
"Number of inputs in forward and inputs during graph constructions do not match");
for (int i = 0; i < numInputs; i++)
{
if (inputs[i] != einsumInpShapes[i])
CV_Error(Error::StsAssert, "Passed input shapes do not match with parsed input shapes!");
}
outputs.clear();
outputs.emplace_back(einsumOutDims);
return true;
} // getMemoryShape
// forward
void forward(InputArrayOfArrays inputs_arr,
OutputArrayOfArrays outputs_arr,
OutputArrayOfArrays internals_arr) CV_OVERRIDE
{
CV_TRACE_FUNCTION();
CV_TRACE_ARG_VALUE(name, "name", name.c_str());
if (inputs_arr.depth() == CV_16F)
{
forward_fallback(inputs_arr, outputs_arr, internals_arr);
return;
}
// homogenize inputs
preProcessInputs(inputs_arr);
std::vector<cv::Mat> rawInputs, outputs;
inputs_arr.getMatVector(rawInputs);
outputs_arr.getMatVector(outputs);
Mat result;
// Pre-process the first input so as to reduce any dims that only it has
{
MatShape reducedDims;
MatShape preservedDims;
MatShape preservedShape;
reducedDims.reserve(numLetterIndices); // num_subscript_labels is the upper bound. No harm in over-reserving.
preservedDims.reserve(numLetterIndices); // num_subscript_labels is the upper bound. No harm in over-reserving.
for (size_t i = 0; i < numLetterIndices; ++i) {
if (subscriptIndicesToLastInput[i] == 0) {
reducedDims.push_back(i);
} else {
preservedDims.push_back(i);
}
}
// Reduce the dims that are last seen in the first input alone
if (reducedDims.size() != 0)
{
result = reduceSum((!preProcessedInputs[0].empty() ? preProcessedInputs[0] : rawInputs[0]), reducedDims);
} else {
// Check if there is a pre-processed version of this input
// If so assign it to result
if (!preProcessedInputs[0].empty())
{
result = preProcessedInputs[0];
}
}
// Finalize the output at this stage if num_inputs == 1
if (numInputs == 1) {
// Finalize the output by applying any transpose required to get
// it to the required output ordering and move it to the op's output
result = FinalizeOutput(!result.empty() ? result : rawInputs[0], preservedDims);
}
}
// Process the operands in a pair-wise fashion
{
bool isFinalPair = false;
// Keep processing each input pair-wise
for (int input = 1; input < numInputs; ++input) {
MatShape reducedDims;
reducedDims.reserve(numLetterIndices); // num_subscript_labels is the upper bound. No harm in over-reserving by a small margin.
for (int dim = 0; dim < numLetterIndices; ++dim)
{
if (subscriptIndicesToLastInput[dim] == input)
{
// This is the last input we are seeing this dimension (and it doesn't occur in the output), so reduce along the dimension
reducedDims.push_back(dim);
}
}
if (input == numInputs - 1)
isFinalPair = true;
// create temporary variable
MatShape tmpResult;
for (int i = 0; i < result.size.dims(); i++)
tmpResult.emplace_back(result.size[i]);
// Use either the preprocessed inputs (if it is available) or the corresponding raw inputs
result = pairwiseOperandProcess(!result.empty() ? result : rawInputs[0],
!result.empty() ? tmpResult : homogenizedInputDims[0],
!preProcessedInputs[input].empty() ? preProcessedInputs[input] : rawInputs[input],
homogenizedInputDims[input],
reducedDims,
isFinalPair);
}
}
// check of product of output dimentions and computed output dimentions match
size_t reqProd = std::accumulate(einsumOutDims.begin(), einsumOutDims.end(), 1, std::multiplies<int>());
MatShape realOutputDims = shape(result);
size_t realProd = std::accumulate(realOutputDims.begin(), realOutputDims.end(), 1, std::multiplies<int>());
CV_CheckEQ(reqProd, realProd, "Real output can not be shaped in to required output");
// reduce dimentions
result = result.reshape(1, einsumOutDims.size(), einsumOutDims.data());
result.copyTo(outputs[0]);
} // forward
#ifdef HAVE_DNN_NGRAPH
virtual Ptr<BackendNode> initNgraph(const std::vector<Ptr<BackendWrapper> >&,
const std::vector<Ptr<BackendNode> >& nodes) CV_OVERRIDE {
ov::OutputVector inputs(nodes.size());
for (size_t i = 0; i < nodes.size(); ++i) {
inputs[i] = nodes[i].dynamicCast<InfEngineNgraphNode>()->node;
}
auto einsum = std::make_shared<ov::op::v7::Einsum>(inputs, equation);
return new InfEngineNgraphNode(einsum);
}
#endif // HAVE_DNN_NGRAPH
}; // EinsumClass
Mat LayerEinsumImpl::reduceSum(Mat& src, MatShape& reduceAxis)
{
// initialize ReduceLayer
LayerParams lp;
lp.set("reduce", "SUM");
int num_axes = reduceAxis.size();
lp.set("axes", DictValue::arrayInt(&reduceAxis[0] , num_axes));
reduce = ReduceLayer::create(lp);
// Compute output shapes
std::vector<MatShape> inputShapes{shape(src)};
std::vector<MatShape> outputShapes, internalShapes;
reduce->getMemoryShapes(inputShapes, 1, outputShapes, internalShapes);
Mat output(outputShapes[0], CV_32F);
std::vector<Mat> inputs;
std::vector<Mat> outputs;
std::vector<Mat> internals;
inputs.emplace_back(src);
outputs.emplace_back(output);
reduce->forward(inputs, outputs, internals);
return outputs[0];
}
void LayerEinsumImpl::preProcessInputs(InputArrayOfArrays& inputs_arr)
{
std::vector<cv::Mat> inputs;
inputs_arr.getMatVector(inputs);
preProcessedInputs.reserve(inputs.size());
homogenizedInputDims.reserve(inputs.size());
int inputIter = 0;
for(const Mat& input : inputs)
{
Mat preprocessed;
// variable to hold processed version of the original input
MatShape input_dims = shape(input);
const auto& currSubscriptIndices = inputSubscriptIndices[inputIter];
// There should be subscript index (subscript label) for each dim of the input
CV_CheckEQ(input_dims.size(), currSubscriptIndices.size(),
"Rank of the input must match number of subscript labels corresponding to the input");
std::vector<int> subscriptIndicesToInputIndex(numLetterIndices, -1);
// this will hold input dims after reordering so that all inputs have
// same axes order
MatShape homogenizedInputDims_(numLetterIndices, 1);
int dimIndexInIreprocessedInput = 0;
int dimIndexInOriginalInput = 0;
for (const auto& subscriptIndex : currSubscriptIndices)
{
if(subscriptIndicesToInputIndex[subscriptIndex] == -1){
subscriptIndicesToInputIndex[subscriptIndex] = dimIndexInIreprocessedInput++;
homogenizedInputDims_[subscriptIndex] = input_dims[dimIndexInOriginalInput];
} else {
// Call diagonal
preprocessed = Diagonal(
!preprocessed.empty() ? preprocessed : inputs[inputIter],
subscriptIndicesToInputIndex[subscriptIndex],
dimIndexInIreprocessedInput);
}
++dimIndexInOriginalInput;
}
std::vector<size_t> permutation;
for(auto& d : subscriptIndicesToInputIndex)
{
if (d != -1)
permutation.emplace_back(d);
}
if (IsTransposeRequired(
!preprocessed.empty() ? preprocessed.size.dims() : inputs[inputIter].size.dims(),
permutation))
{
// call transpose
preprocessed = Transpose(
!preprocessed.empty() ? preprocessed : inputs[inputIter],
!preprocessed.empty() ? shape(preprocessed) : shape(inputs[inputIter]),
permutation);
}
if (!preprocessed.empty())
{
preprocessed = preprocessed.reshape(1, homogenizedInputDims_.size(), homogenizedInputDims_.data());
}
preProcessedInputs.emplace_back(preprocessed);
homogenizedInputDims.emplace_back(homogenizedInputDims_);
++inputIter;
}
}
void LayerEinsumImpl::parseEquation(String equation)
{
// remove white spaces in the copy
equation.erase(std::remove_if(equation.begin(), equation.end(), ::isspace), equation.end());
// check if '->' - the output subscript label is present in the equation;
std::size_t arrow_idx = equation.find("->");
if (arrow_idx != std::string::npos)
{
// split left and righ hand sides of the equation
lhs_eq = equation.substr(0, arrow_idx);
rhs_eq = equation.substr(arrow_idx + 2);
explicitEquation = true;
} else {
lhs_eq = equation;
}
// split lhs_eq by ',' - comma and put all created token - splits
// into lhs_eq_tokens vector
std::stringstream src(lhs_eq);
for (std::string token; std::getline(src, token, ',');) {
lhs_eq_tokens.emplace_back(token);
}
}
void LayerEinsumImpl::calculateOutputShape()
{
// Traverse through each of the subscript labels within the output subscript.
bool middleOfEllipsis = false;
int ellipsisCharCount = 0;
subscriptIndicesToOutputIndices.resize(numLetterIndices, -1);
std::array<int, numOfLetters> outputLetterToCount;
outputLetterToCount.fill(0);
int outputDimCounter = 0;
for (auto letter : rhs_eq)
{
if(letter == '.')
{
middleOfEllipsis = true;
// Make sure there aren't more than 3 '.'s in the current subscript
if (++ellipsisCharCount > 3) {
CV_Error(Error::StsError, "Found a '.' not part of an ellipsis in the output subscript provided");
}
if (ellipsisCharCount == 3) { // Ellipsis is complete. Process it.
middleOfEllipsis = false;
for (size_t i = 0; i < numOfEllipsisDims; ++i) {
einsumOutDims.emplace_back(subscriptIndicesToDimValue[i]);
// The ellipsis is seen in the output and hence the corresponding dims are to not be reduced
subscriptIndicesToLastInput[i] = -1;
subscriptIndicesToOutputIndices[i] = outputDimCounter++;
}
}
} else {
CV_CheckEQ(middleOfEllipsis, false,
"Encountered '.' character that is not part of output subscript");
auto letterIndex = letterToIndex(letter);
CV_CheckNE(letterIndex, -1,
"The only permissible subscript labels are lowercase letters (a-z) and uppercase letters (A-Z).");
CV_CheckEQ(outputLetterToCount[letterIndex], 0,
"Output subscript constains repeated letters");
++outputLetterToCount[letterIndex];
auto mappedIndex = letter2index[letterIndex];
CV_CheckNE(mappedIndex, -1,
"Output subscript has letters that were not encountered in the inputs");
// Push output dimention
// Einsum layer only has one output vector
einsumOutDims.emplace_back(subscriptIndicesToDimValue[mappedIndex]);
// Reset the last input index for this subscript label
// given that it is seen in the output and hence can't be reduced
subscriptIndicesToLastInput[mappedIndex] = -1;
subscriptIndicesToOutputIndices[mappedIndex] = outputDimCounter++;
}
}
}
void LayerEinsumImpl::validateOutputSubscript()
{
// The explicit form requires no operation, as the output
// would have already been parsed during the input parsing process.
if(explicitEquation)
{
// Ensure that the provided explicit equation includes an ellipsis if the input contains ellipses.
if(numOfEllipsisDims > 0)
{
if(rhs_eq.find("...") == std::string::npos)
{
CV_Error(Error::StsError,
"Provided output subscript does not include ellipsis while Inputs subscrits constain ellipsis");
}
}
}
}
void LayerEinsumImpl::processBroadcastedDims()
{
// Only compute this function if ellipsis "..." was found in the equation
if (numOfEllipsisDims > 0)
{
// extend the number of subscript labels to include each ellipsis dim as
// theoretically each ellipsis dim does correspond to a "virtual" subscript label
numLetterIndices += numOfEllipsisDims;
// We are going to assign the broadcasted dims outermost subscript indices (i.e.) 0 -> numOfEllipsisDims - 1
// as most likely bradcasted dims will be batch dimensions (i.e.) outermost dimensions and hence we don't have to pay
// transposing while "homogenizing" the input
// Hence offset all subscript indices by numOfEllipsisDims
for (size_t i = 0; i < numOfLetters; ++i){
if (letter2count[i] != -1){
letter2index[i] += numOfEllipsisDims;
}
}
std::vector<int> tempIndex2LastInput(numLetterIndices, -1);
for (int i = 0; i < subscriptIndicesToLastInput.size(); ++i){
tempIndex2LastInput[i + numOfEllipsisDims] = subscriptIndicesToLastInput[i];
}
subscriptIndicesToLastInput = std::move(tempIndex2LastInput);
std::vector<int> tempIndexToDimValue(numLetterIndices, -1);
for (int i = 0; i < subscriptIndicesToDimValue.size(); ++i){
tempIndexToDimValue[i + numOfEllipsisDims] = subscriptIndicesToDimValue[i];
}
subscriptIndicesToDimValue = std::move(tempIndexToDimValue);
for (size_t i = 0; i < inputSubscriptIndices.size(); ++i)
{
auto& currentInputDimIndicesToSubscriptIndices = inputSubscriptIndices[i];
std::vector<int> tempCurrentInputDimIndicesToSubscriptIndices;
tempCurrentInputDimIndicesToSubscriptIndices.reserve(currentInputDimIndicesToSubscriptIndices.size());
// make sure it is correct
const auto& dims = einsumInpShapes[i];
auto rank = dims.size();
size_t dimIter = 0;
size_t numBroadcastedIndices = 0;
while (dimIter < currentInputDimIndicesToSubscriptIndices.size())
{
auto value = currentInputDimIndicesToSubscriptIndices[dimIter];
if (value == numOfLetters)
{ // This is a broadcasted dim
// Shouldn't hit this error - just a sanity check
CV_Assert(numBroadcastedIndices < numOfEllipsisDims);
tempCurrentInputDimIndicesToSubscriptIndices.push_back(static_cast<int>(numBroadcastedIndices));
subscriptIndicesToLastInput[numBroadcastedIndices] = i;
// This is the first time we are seeing this broadcasted dim
if (subscriptIndicesToDimValue[numBroadcastedIndices] == -1)
{
subscriptIndicesToDimValue[numBroadcastedIndices] = dims[dimIter];
} else { // We have seen this broadcasted dim before
// Check if the previous value is equal to the current value
if (subscriptIndicesToDimValue[numBroadcastedIndices] != dims[dimIter])
{
// If they are not equal, one of them needs to be 1
if (subscriptIndicesToDimValue[numBroadcastedIndices] == 1)
{
subscriptIndicesToDimValue[numBroadcastedIndices] = dims[dimIter];
} else {
CV_CheckEQ(dims[dimIter], 1, "The broadcasted dimensions of the inputs are incompatible");
}
}
}
++numBroadcastedIndices;
} else { // This is a regular dim - offset it by number of broadcasted dims
tempCurrentInputDimIndicesToSubscriptIndices.push_back(value + static_cast<int>(numOfEllipsisDims));
}
++dimIter;
}
// Shouldn't hit this error - just a sanity check
CV_Assert(dimIter == rank);
currentInputDimIndicesToSubscriptIndices = std::move(tempCurrentInputDimIndicesToSubscriptIndices);
}
}
}
void LayerEinsumImpl::processEquation(const std::vector<MatShape>& inputs)
{
// Check if number of tokens in equal to number of inputs.
// For install "ij, jk -> ik" needs to have 2 inputs tensors
int num_input_tensors = inputs.size();
CV_CheckEQ(static_cast<int>(lhs_eq_tokens.size()), num_input_tensors,
"Number of input tensors does not match the number of subscripts in the input equation");
int inputIdx = 0;
for (const auto& token : lhs_eq_tokens)
{
const MatShape shape = inputs[inputIdx];
size_t rank = shape.size();
size_t dim_count = 0;
std::vector<int> currTokenIndices;
currTokenIndices.reserve(rank);
// Variable to deal with "ellipsis" - '...' in the input
bool middleOfellipsis = false;
int ellipsisCharCount = 0;
for (auto letter : token)
{
if (letter == '.')
{
middleOfellipsis = true;
// there should not be more than 3 '.'s in the current subscript
if (++ellipsisCharCount > 3)
{
CV_Error(Error::StsError, cv::format("Found a '.' not part of an ellipsis in input: %d", inputIdx));
}
// We have seen all 3 '.'s. We can safely process the ellipsis now.
if (ellipsisCharCount == 3)
{
middleOfellipsis = false;
// Example for the following line of code
// Subscript "...ij" for an input of rank 6
// numOfEllipsisDims = 6 - 5 + 3 = 4
int currentNumOfEllipsisDims = static_cast<int>(rank) - token.length() + 3;
CV_CheckGE(currentNumOfEllipsisDims, 0,
"Einsum subscripts string contains too many subscript labels when compared to the rank of the input");
// Theoretically, currentNumOfEllipsisDims could be 0
// Example: For an input of rank 2 paired with a subscript "...ij"
if (currentNumOfEllipsisDims != 0)
{
// We have seen a ellipsis before - make sure ranks align as per the ONNX spec -
// "Ellipsis must indicate a fixed number of dimensions."
if (numOfEllipsisDims != 0){
CV_CheckEQ(numOfEllipsisDims, static_cast<size_t>(currentNumOfEllipsisDims),
"Ellipsis must indicate a fixed number of dimensions across all inputs");
} else {
numOfEllipsisDims = static_cast<size_t>(currentNumOfEllipsisDims);
}
// We reserve 'numOfLetters' for broadcasted dims as we only allow 'a' - 'z'
// and 'A' - 'Z' (0 - 51) for non-broadcasted dims.
// We will assign appropriate indices (based on number of dimensions the ellipsis corresponds to)
// during broadcasting related post-processing.
for (size_t i = 0; i < numOfEllipsisDims; ++i){
currTokenIndices.push_back(numOfLetters);
}
// Offset 'dim_count' by number of dimensions the ellipsis corresponds to
dim_count += numOfEllipsisDims;
}
}
} else {
if (middleOfellipsis){
CV_Error(Error::StsAssert,
cv::format(
"Encountered '.' character that is not part of an ellipsis in the input: [%d]",
inputIdx));
}
int letterIdx = letterToIndex(letter);
CV_CheckNE(letterIdx, -1,
"The only permissible subscript labels are lowercase letters (a-z) and uppercase letters (A-Z).");
int dimValue = shape[dim_count];
// The subscript label was not found in the global subscript label array
// Therefore, it is added to both the local and global subscript arrays
if(letter2count[letterIdx] == 0){
letter2index[letterIdx] = numLetterIndices++;
subscriptIndicesToDimValue.push_back(dimValue);
subscriptIndicesToLastInput.push_back(inputIdx);
} else {
// This letter has been seen in at least one other operand's subscript
// It must be equal unless one of them is a 1 (Numpy allows this)
auto mappedIndx = letter2index[letterIdx];
subscriptIndicesToLastInput[mappedIndx] = inputIdx;
if (subscriptIndicesToDimValue[mappedIndx] != dimValue) {
if (dimValue != 1) {
CV_Error(Error::StsError, cv::format("Einsum operands can not be broadcasted."
"Check input shapes/equation passed."
"Input shape of operand [%d]", inputIdx) +
cv::format(" is incompatible in the dimention [%zu].", static_cast<size_t>(dim_count)));
}
}
}
++letter2count[letterIdx];
currTokenIndices.push_back(letter2index[letterIdx]);
CV_CheckLE(++dim_count, rank,
"The Einsum subscripts string has an excessive number of subscript labels compared to the rank of the input.");
}
}
// When no broadcasting is requested, the number of subscript labels (dim_counter) should match the input's rank.
CV_Assert(!(numOfEllipsisDims == 0 && dim_count != rank)
&& "The Einsum subscripts string does not contain required amount of subscript labels and no ellipsis is provided in the input.");
inputSubscriptIndices.emplace_back(std::move(currTokenIndices));
++inputIdx;
}
}
Mat LayerEinsumImpl::FinalizeOutput(
const Mat& candidateOutput,
const MatShape& ordered_subscript_indices_in_candidate)
{
const std::vector<int>& subscript_indices_to_output_indices = subscriptIndicesToOutputIndices;
const auto output_dims = einsumOutDims;
MatShape output_shape = output_dims;
const auto output_rank = output_dims.size();
// CV_CheckEQ((int) candidateOutput.dims, (int) output_shape.size(),
// "Einsum op: The candidate output cannot be reshaped into the op's output");
const MatShape candidate_output_dims = MatShape(candidateOutput.size.p, candidateOutput.size.p + candidateOutput.dims);
const int candidate_output_rank = candidate_output_dims.size();
// This vector holds the shape of the candidate_output after removing the dims that have
// been reduced in the final output
MatShape candidate_output_shape_without_reduced_dims;
candidate_output_shape_without_reduced_dims.reserve(candidate_output_rank); // reserve upper bound
// Identify the permutation required by the op's output
std::vector<size_t> output_permutation;
output_permutation.resize(output_rank, 0);
size_t output_iter = 0;
for (size_t iter = 0, end = ordered_subscript_indices_in_candidate.size(); iter < end; ++iter)
{
auto output_index = subscript_indices_to_output_indices[ordered_subscript_indices_in_candidate[iter]];
// If output_index is -1, then this dimension does not show up in the op's output and has been reduced along the way
if (output_index != -1)
{
output_permutation[output_index] = output_iter++;
candidate_output_shape_without_reduced_dims.push_back(candidate_output_dims[iter]);
} else {
// This dim doesn't show up in the op's output and hence we check if the dim has been reduced in the candidate output
CV_CheckEQ(candidate_output_dims[iter], 1,
"Not all dimensions to be reduced have been reduced in the candidate output. Candidate output dims: "); //%d", candidateOutput.size));
}
}
// Transpose to the required final output order
// (Identify no-op transposes and prevent triggering the transpose)
if (IsTransposeRequired(candidate_output_shape_without_reduced_dims.size(), output_permutation))
{
auto candidate_output_transposed = Transpose(
candidateOutput,
candidate_output_shape_without_reduced_dims,
output_permutation);
return candidate_output_transposed;
}
return candidateOutput;
}
Mat LayerEinsumImpl::pairwiseOperandProcess(
const Mat& left,
const MatShape& leftShapeOverride,
const Mat& right,
const MatShape& rightShapeOverride,
const MatShape& reduceDims,
bool isFinalPair
)
{
size_t matDimSize = left.total();
size_t overrideDimSize = total(leftShapeOverride);
CV_CheckEQ(matDimSize, overrideDimSize, "Override dims are not compatible with left tensor shape");
matDimSize = right.total();
overrideDimSize = total(rightShapeOverride);
CV_CheckEQ(matDimSize, overrideDimSize, "Override dims are not compatible with right tensor shape");
// Make copy as this may be overridden downstream
const auto& leftDims = leftShapeOverride;
const auto& rightDims = rightShapeOverride;
int leftRank = static_cast<int>(leftDims.size());
int rightRank = static_cast<int>(rightDims.size());
Mat currentLeft;
Mat currentRight;
CV_CheckEQ(leftRank, rightRank, "Raks of pair-wise operands must be equal");
// Following vectors hold:
// lro: dim indices that are present in left, right, and reduce_dims
// lo: dim indices that are present in left and reduce_dims
// ro: dim indices that are present in right and reduce_dims
std::vector<size_t> lro;
lro.reserve(5); // Reserve an arbitrary amount of space for this vector (not bound to see a tensor of rank > kTensorShapeSmallBufferElementsSize)
std::vector<size_t> lo;
lo.reserve(5); // Reserve an arbitrary amount of space for this vector (not bound to see a tensor of rank > kTensorShapeSmallBufferElementsSize)
std::vector<size_t> ro;
ro.reserve(5); // Reserve an arbitrary amount of space for this vector (not bound to see a tensor of rank > kTensorShapeSmallBufferElementsSize)
// Maintain sizes to create reshaped "views"
int lro_size = 1;
int lo_size = 1;
int ro_size = 1;
int reduced_size = 1;
size_t reduceDimsIter = 0;
size_t reduceDimsSize = reduceDims.size();
for (int i = 0; i < leftRank; ++i)
{
int leftDim = leftDims[i];
int rightDim = rightDims[i];
bool hasLeftDim = leftDim > 1; // non-trivial dimension (dim_value != 1)
bool hasRightDim = rightDim > 1; // non-trivial dimension (dim_value != 1)
if (reduceDimsIter < reduceDimsSize && reduceDims[reduceDimsIter] == i)
{
// This dimension is to be reduced after this pair-wise operation
++reduceDimsIter;
if (hasLeftDim && hasRightDim){
// Both inputs have non-trivial dim values along this dimension
// Both the left and right operands have non-trivial dimension value along this axis
CV_CheckEQ(leftDim, rightDim, "Einsum op: Input dimensions must be equal along an axis to be reduced across all inputs");
reduced_size *= leftDim;
} else if (hasLeftDim){
// if the dim to be reduced is only in one of left and right, we can reduce right away
Mat tensorToReduce = !currentLeft.empty() ? currentLeft : left;
MatShape shapeToReduce = !currentLeft.empty() ? shape(currentLeft) : leftDims;
currentLeft = reduceSum(tensorToReduce, shapeToReduce);
} else if (hasRightDim){
Mat tensorToReduce = !currentRight.empty() ? currentRight : right;
MatShape shapeToReduce = !currentRight.empty() ? shape(currentRight) : rightDims;
currentLeft = reduceSum(tensorToReduce, shapeToReduce);
}
} else {
// This dimension is not reduced (i.e.) it appears in the output after processing these 2 operands
// Both the left and right operands have non-trivial dimension value along this axis
// They must be equal
if (hasLeftDim && hasRightDim){
CV_CheckEQ(leftDim, rightDim, "Input shapes do not align");
lro.push_back(i);
lro_size *= leftDim;
} else if (hasLeftDim) {
// The left operand has non-trivial dimension value
lo.push_back(i);
lo_size *= leftDim;
} else {
// The right operand may or may not have non-trivial dim value
// If it has trivial dim value (1),
// it will just form a trailing dimension for the right operand
ro.push_back(i);
ro_size *= rightDim;
}
}
}
// Permutate the left operand so that the axes order go like this: [lro, lo, reduce_dims, ro]
MatShape reshaped_dims;
std::vector<size_t> left_permutation;
left_permutation.reserve(lro.size() + lo.size() + reduceDims.size() + ro.size());
left_permutation.insert(left_permutation.end(), lro.begin(), lro.end());
left_permutation.insert(left_permutation.end(), lo.begin(), lo.end());
// left_permutation.insert(left_permutation.end(), reduce_dims.begin(), reduce_dims.end());
for (auto& a : reduceDims)
{
left_permutation.push_back(a);
}
left_permutation.insert(left_permutation.end(), ro.begin(), ro.end());
if (IsTransposeRequired(!currentLeft.empty() ? currentLeft.dims : leftDims.size(),
left_permutation))
{
if (!currentLeft.empty() && IsTransposeReshapeForEinsum(left_permutation,
shape(currentLeft),
reshaped_dims))
{
// This can be done because curent_* tensors (if they exist) and output tensors are
// intermediate tensors and cannot be input tensors to the Einsum node itself
// (which are immutable).
currentLeft = currentLeft.reshape(1, reshaped_dims.size(), reshaped_dims.data());
} else {
// Covered by ExplicitEinsumAsTensorContraction, DiagonalWithMatmul, ...
currentLeft = Transpose(!currentLeft.empty() ? currentLeft: left,
!currentLeft.empty() ? shape(currentLeft) : leftDims,
left_permutation);
}
}
// Permutate the right operand so that the axes order go like this: [lro, reduce_dims, ro, lo]
std::vector<size_t> right_permutation;
right_permutation.reserve(lro.size() + lo.size() + reduceDims.size() + ro.size());
right_permutation.insert(right_permutation.end(), lro.begin(), lro.end());
// right_permutation.insert(right_permutation.end(), reduce_dims.begin(), reduce_dims.end());
for (auto& a : reduceDims) {
right_permutation.push_back(a);
}
right_permutation.insert(right_permutation.end(), ro.begin(), ro.end());
right_permutation.insert(right_permutation.end(), lo.begin(), lo.end());
if (IsTransposeRequired(!currentRight.empty() ? currentRight.dims: rightDims.size(),
right_permutation))
{
if (!currentRight.empty() && IsTransposeReshapeForEinsum(right_permutation,
shape(currentRight),
reshaped_dims))
{
currentRight = currentRight.reshape(1, reshaped_dims.size(), reshaped_dims.data());
} else {
currentRight = Transpose(!currentRight.empty() ? currentRight : right,
!currentRight.empty() ? shape(currentRight) : rightDims,
right_permutation);
}
}
// Calculate output size
// Output shape will be determined by rules of MatMul:
// because we are multiplying two tensors of shapes [lro, lo, reduce_dims] , [lro, reduce_dims, ro]
// [dim_value of `lro` dims,
// dim_value of `lo` dims,
// `1` for each of the `reduce_dims`,
// dim_value of `ro` dims]
MatShape outputDims;
outputDims.reserve(lro.size() + lo.size() + reduceDims.size() + ro.size());
for (size_t i = 0; i < lro.size(); ++i)
{
outputDims.emplace_back(leftDims[lro[i]]);
}
for (size_t i = 0; i < lo.size(); ++i)
{
outputDims.emplace_back(leftDims[lo[i]]);
}
for (size_t i = 0; i < reduceDims.size(); ++i)
{
outputDims.emplace_back(1); // reduced dimensions will have a value 1 in it
}
for (size_t i = 0; i < ro.size(); ++i) {
outputDims.emplace_back(rightDims[ro[i]]);
}
MatShape currentSubscriptOrder;
// Calculate output permutation
// After the MatMul op, the because the two operands have been permutated,
// the output is permutated as well with respect to the original ordering of the axes.
// The permutated order will be the dims in: [lro, lo, reduced_dims, ro]
// Hence invert the permutation by a permutation that puts the axes in the same ordering
std::vector<size_t> outputPermutation;
if (!isFinalPair) { // If this is not the final pair, we need to permutate the result to match the pre-fixed order for the next iteration
outputPermutation.resize(lro.size() + lo.size() + reduceDims.size() + ro.size(), 0);
size_t iter = 0;
for (size_t i = 0; i < lro.size(); ++i)
{
outputPermutation[lro[i]] = iter++;
}
for (size_t i = 0; i < lo.size(); ++i)
{
outputPermutation[lo[i]] = iter++;
}
for (size_t i = 0; i < reduceDims.size(); ++i)
{
outputPermutation[reduceDims[i]] = iter++;
}
for (size_t i = 0; i < ro.size(); ++i)
{
outputPermutation[ro[i]] = iter++;
}
} else {
currentSubscriptOrder.reserve(lro.size() + lo.size() + reduceDims.size() + ro.size());
currentSubscriptOrder.insert(currentSubscriptOrder.end(), lro.begin(), lro.end());
currentSubscriptOrder.insert(currentSubscriptOrder.end(), lo.begin(), lo.end());
currentSubscriptOrder.insert(currentSubscriptOrder.end(), reduceDims.begin(), reduceDims.end());
currentSubscriptOrder.insert(currentSubscriptOrder.end(), ro.begin(), ro.end());
}
Mat output = batchwiseMatMul(
!currentLeft.empty() ? currentLeft : left,
MatShape({static_cast<int>(lro_size), static_cast<int>(lo_size), static_cast<int>(reduced_size)}),
!currentRight.empty() ? currentRight : right,
MatShape({static_cast<int>(lro_size), static_cast<int>(reduced_size), static_cast<int>(ro_size)})
);
//reshape
output = output.reshape(1, outputDims.size(), outputDims.data());
if (!isFinalPair)
{ // This is not the final pair - so bring the axes order to what the inputs conformed to
if (IsTransposeRequired(outputDims.size(), outputPermutation))
{
if (IsTransposeReshapeForEinsum(outputPermutation,
outputDims,
reshaped_dims))
{
// See note following the previous call of function IsTransposeReshapeForEinsum.
// Covered by ExplicitEinsumAsTensorContractionReshapeFinal.
output = output.reshape(1, reshaped_dims.size(), reshaped_dims.data());
}
else {
output = Transpose(
output,
outputDims,
outputPermutation);
}
}
} else { // This is the final pair - Transpose directly to the output ordering required and copy the contents to the op's output
// not sure if this finalize shape is needed at all
output = FinalizeOutput(output, currentSubscriptOrder);
}
return output;
};
Mat LayerEinsumImpl::batchwiseMatMul(
const Mat& input1,
const MatShape& input1ShapeOverride,
const Mat& input2,
const MatShape& input2ShapeOverride)
{
// Sanity checks before the actual MatMul
CV_CheckType(input1.type(), input2.type(), "Data types of the inputs must match for MatMul");
CV_CheckEQ(input1ShapeOverride.size(), (size_t) 3, "Only 1 batch dimension is allowed for MatMul");
CV_CheckEQ(input2ShapeOverride.size(), (size_t) 3, "Only 1 batch dimension is allowed for MatMul");
CV_CheckEQ((size_t) input1ShapeOverride[0], (size_t) input2ShapeOverride[0], "Batch dimension should match for MatMul;");
CV_CheckEQ((size_t) input1ShapeOverride[2], (size_t) input2ShapeOverride[1], "Incompatible matrix dimensions for matMul");
int batches = input1ShapeOverride[0];
int M = input1ShapeOverride[1];
int K = input1ShapeOverride[2];
int N = input2ShapeOverride[2];
Mat reshapedInput1 = input1;
Mat reshapedInput2 = input2;
Mat output;
if (batches > 1)
{
// create tmpout with type like input1
output = Mat({batches, M, N}, input1.type());
reshapedInput2 = reshapedInput2.reshape(1, input2ShapeOverride);
reshapedInput1 = reshapedInput1.reshape(1, input1ShapeOverride);
fastGemmBatch(false, false, 1.0, reshapedInput1, reshapedInput2, 0.0, output, opt);
} else {
// input1 should of size MxK
if (input1.dims > 2 || input1.size[0] != M || input1.size[1] != K)
{
int shape[] = {M, K};
reshapedInput1 = input1.reshape(1, 2, shape);
}
// input2 should be of size KxN
if (input2.dims > 2 || input2.size[0] != K || input2.size[1] != N)
{
int shape2[] = {K, N};
reshapedInput2 = input2.reshape(1, 2, shape2);
}
output = Mat(M, N, reshapedInput1.type());
fastGemm(false, false, 1.0, reshapedInput1, reshapedInput2, 0.0, output, opt);
output = output.reshape(1, {1, M, N});
}
return output;
};
Ptr<EinsumLayer> EinsumLayer::create(const LayerParams& params)
{
return makePtr<LayerEinsumImpl>(params);
}
}} // namespace cv::dnn