Merge pull request #25567 from Abdurrahheem:ash/01D-einsum-test

0/1D Einsum Layer Test #25567

This PR introduces 0/1D test cases for Einsum layer.

TODO:
- Add support for 0D tensors to Einsum layer

### Pull Request Readiness Checklist

See details at https://github.com/opencv/opencv/wiki/How_to_contribute#making-a-good-pull-request

- [x] I agree to contribute to the project under Apache 2 License.
- [x] To the best of my knowledge, the proposed patch is not based on a code under GPL or another license that is incompatible with OpenCV
- [x] The PR is proposed to the proper branch
- [x] There is a reference to the original bug report and related work
- [x] There is accuracy test, performance test and test data in opencv_extra repository, if applicable
      Patch to opencv_extra has the same branch name.
- [x] The feature is well documented and sample code can be built with the project CMake
pull/25409/head
Abduragim Shtanchaev 8 months ago committed by GitHub
parent b637e3a66e
commit 021e5184bc
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
  1. 18
      modules/dnn/src/layers/einsum_layer.cpp
  2. 75
      modules/dnn/test/test_layers_1d.cpp

@ -65,12 +65,12 @@ static Mat Transpose(
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){
if (input_rank == 0 || permutation.size() == 0){
return false;
}
CV_Assert(input_rank == permutation.size());
// Weeds out cases where permutation is something like [0, 1, 2] for a 3D input and so on
bool transpose_required = false;
@ -616,6 +616,10 @@ void LayerEinsumImpl::preProcessInputs(InputArrayOfArrays& inputs_arr)
// variable to hold processed version of the original input
MatShape input_dims = shape(input);
if (inputSubscriptIndices.empty()){
homogenizedInputDims.emplace_back(MatShape(numLetterIndices, 1));
continue;
}
const auto& currSubscriptIndices = inputSubscriptIndices[inputIter];
// There should be subscript index (subscript label) for each dim of the input
@ -870,6 +874,9 @@ 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();
if (lhs_eq_tokens.empty() || (lhs_eq_tokens.size() == 1 && lhs_eq_tokens[0].empty() && lhs_eq == ",") ) {
return;
}
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");
@ -1363,7 +1370,12 @@ Mat LayerEinsumImpl::batchwiseMatMul(
}
output = Mat(M, N, reshapedInput1.type());
fastGemm(false, false, 1.0, reshapedInput1, reshapedInput2, 0.0, output, opt);
if (shape(reshapedInput1).empty() && shape(reshapedInput2).empty())
{
output = reshapedInput1.mul(reshapedInput2); // fastGemm does not support 0D * 0D multiplication
} else {
fastGemm(false, false, 1.0, reshapedInput1, reshapedInput2, 0.0, output, opt);
}
output = output.reshape(1, {1, M, N});
}

@ -682,4 +682,79 @@ INSTANTIATE_TEST_CASE_P(/*nothing*/, Layer_Const_Test, testing::Values(
std::vector<int>({4, 1})
));
typedef testing::TestWithParam<tuple<std::vector<int>, std::string>> Layer_Einsum_Test;
TEST_P(Layer_Einsum_Test, Accuracy_01D)
{
auto tup = GetParam();
std::vector<int> input_shape = std::get<0>(tup);
std::string equation = std::get<1>(tup);
LayerParams lp;
lp.type = "Einsum";
lp.name = "EinsumLayer";
lp.set("equation", equation);
lp.set("inputSize", 2);
lp.set("outputSize", 1);
lp.set("inputShapes0", DictValue::arrayInt(&input_shape[0], input_shape.size()));
lp.set("inputShapes1", DictValue::arrayInt(&input_shape[0], input_shape.size()));
Ptr<Layer> layer = EinsumLayer::create(lp);
cv::Mat input1(input_shape.size(), input_shape.data(), CV_32F);
cv::Mat input2(input_shape.size(), input_shape.data(), CV_32F);
cv::randn(input1, 0.0, 1.0); cv::randn(input2, 0.0, 1.0);
std::vector<Mat> inputs = {input1, input2};
std::vector<Mat> outputs;
runLayer(layer, inputs, outputs);
ASSERT_EQ(1, outputs.size());
// create output_ref to compare with outputs
cv::Mat output_ref;
int size[] = {1};
if (equation == ",->"){
output_ref = input1.mul(input2);
}else if (equation == "i, i->i"){
output_ref = input1.mul(input2);
} else if (equation == "i, i->"){
output_ref = input1.mul(input2);
cv::Scalar sum = cv::sum(output_ref);
output_ref = cv::Mat(0, nullptr, CV_32F, sum[0]);
} else if (equation == "ij, ij->ij"){
output_ref = input1.mul(input2);
} else if (equation == "ij, ij->i"){
output_ref = input1.mul(input2);
if (input_shape[0] == 1){
cv::Scalar sum = cv::sum(output_ref);
output_ref = cv::Mat(1, size, CV_32F, sum[0]);
} else if (input_shape[1] == 1){
size[0] = input_shape[0];
output_ref = output_ref.reshape(1, 1, size);
} else {
cv::reduce(output_ref, output_ref, 1, cv::REDUCE_SUM, CV_32F);
size[0] = input_shape[0];
output_ref = output_ref.reshape(1, 1, size);
}
} else {
output_ref = cv::Mat();
}
ASSERT_EQ(shape(output_ref), shape(outputs[0]));
normAssert(output_ref, outputs[0]);
}
INSTANTIATE_TEST_CASE_P(/*nothing*/, Layer_Einsum_Test, testing::Values(
std::make_tuple(std::vector<int>({}), std::string(",->")),
std::make_tuple(std::vector<int>({1}), std::string("i, i->i")),
std::make_tuple(std::vector<int>({1}), std::string("i, i->")),
std::make_tuple(std::vector<int>({4}), std::string("i, i->i")),
std::make_tuple(std::vector<int>({4}), std::string("i, i->")),
std::make_tuple(std::vector<int>({1, 4}), std::string("ij, ij->ij")),
std::make_tuple(std::vector<int>({4, 1}), std::string("ij, ij->ij")),
std::make_tuple(std::vector<int>({1, 4}), std::string("ij, ij->i")),
std::make_tuple(std::vector<int>({4, 1}), std::string("ij, ij->i")),
std::make_tuple(std::vector<int>({4, 4}), std::string("ij, ij->i"))
));
}}

Loading…
Cancel
Save