potok/opencv
1
0
Fork 0
mirror of https://github.com/opencv/opencv.git synced 2026-07-29 15:23:05 +04:00
Code Packages Releases Wiki Activity

Merge pull request #28859 from abhishek-gola:fuse_attention

Browse Source 
Added Attention fusion and thin gemm support #28859

Merge with: https://github.com/opencv/opencv_extra/pull/1350

Performance numbers after these optimizations:

For Device:  Intel(R) Core(TM) i9-14900KS, x86, 32 Cores, ubuntu 24.04,
| Model | `ENGINE_NEW (Before)` | `ENGINE_NEW (After)` | `ENGINE_ORT` |
| :--- | :--- | :--- | :--- | 
| **BERT** |26.3 ms| 9.15 ms| 9.13 ms|
| **ViT** | 79.65 ms| 63.23 ms| 32.3 ms|

### 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
This commit is contained in:
Abhishek Gola
2026-04-29 14:13:33 +05:30
committed by GitHub
parent faf34f95af
commit ae1f3a991c
10 changed files with 991 additions and 24 deletions
Download Patch File Download Diff File Expand all files Collapse all files
modules/dnn/test/test_model.cpp
+134
View File
@@ -948,6 +948,140 @@ TEST_P(Reproducibility_ResNet50_QDQ_ONNX, Accuracy)
INSTANTIATE_TEST_CASE_P(/**/, Reproducibility_ResNet50_QDQ_ONNX,
testing::ValuesIn(getAvailableTargets(DNN_BACKEND_OPENCV)));
typedef testing::TestWithParam<Target> Reproducibility_ViT_ONNX;
TEST_P(Reproducibility_ViT_ONNX, Accuracy)
{
Target targetId = GetParam();
applyTestTag(targetId == DNN_TARGET_CPU ? CV_TEST_TAG_MEMORY_1GB : CV_TEST_TAG_MEMORY_2GB);
ASSERT_TRUE(ocl::useOpenCL() || targetId == DNN_TARGET_CPU || targetId == DNN_TARGET_CPU_FP16);
auto engine_forced = static_cast<EngineType>(
cv::utils::getConfigurationParameterSizeT("OPENCV_FORCE_DNN_ENGINE", ENGINE_AUTO));
if (engine_forced == ENGINE_CLASSIC)
{
applyTestTag(CV_TEST_TAG_DNN_SKIP_PARSER);
return;
}
std::string modelname = _tf("vit_base_patch16_224_Opset16.onnx", false);
Net net = readNetFromONNX(modelname);
net.setPreferableBackend(DNN_BACKEND_OPENCV);
net.setPreferableTarget(targetId);
if (targetId == DNN_TARGET_CPU_FP16)
net.enableWinograd(false);
std::string imgname = _tf("sqcat.png");
Mat image = imread(imgname);
// ViT preprocessing: (pixel/255 - 0.5)/0.5 == pixel/127.5 - 1
// blobFromImage form: (pixel - 127.5) * (1/127.5)
Mat input = blobFromImage(image, 1.0 / 127.5, Size(224, 224),
Scalar(127.5, 127.5, 127.5),
true, true, CV_32F);
ASSERT_TRUE(!input.empty());
net.setInput(input);
Mat out = net.forward();
const int K = 5;
std::vector<std::pair<int, float> > res;
topK(out, res, K);
ASSERT_EQ(int(res.size()), K);
// Reference top-5 captured from the ONNX Runtime engine (OPENCV_FORCE_DNN_ENGINE=4).
std::vector<std::pair<int, float> > ref = {
{285, 7.683f}, {282, 7.182f}, {281, 6.894f}, {287, 3.623f}, {283, 3.287f}
};
const float eps = 0.5f;
std::vector<int> reflabels(K), reslabels(K);
for (int i = 0; i < K; i++) {
reflabels[i] = ref[i].first;
reslabels[i] = res[i].first;
}
ASSERT_EQ(reflabels, reslabels);
for (int i = 0; i < K; i++) {
EXPECT_NEAR(ref[i].second, res[i].second, eps);
}
}
INSTANTIATE_TEST_CASE_P(/**/, Reproducibility_ViT_ONNX,
testing::ValuesIn(getAvailableTargets(DNN_BACKEND_OPENCV)));
typedef testing::TestWithParam<Target> Reproducibility_BERT_ONNX;
TEST_P(Reproducibility_BERT_ONNX, Accuracy)
{
Target targetId = GetParam();
applyTestTag(targetId == DNN_TARGET_CPU ? CV_TEST_TAG_MEMORY_1GB : CV_TEST_TAG_MEMORY_2GB);
ASSERT_TRUE(ocl::useOpenCL() || targetId == DNN_TARGET_CPU || targetId == DNN_TARGET_CPU_FP16);
auto engine_forced = static_cast<EngineType>(
cv::utils::getConfigurationParameterSizeT("OPENCV_FORCE_DNN_ENGINE", ENGINE_AUTO));
if (engine_forced == ENGINE_CLASSIC)
{
applyTestTag(CV_TEST_TAG_DNN_SKIP_PARSER);
return;
}
std::string modelname = _tf("onnx/models/bert.onnx", false);
Net net = readNetFromONNX(modelname);
net.setPreferableBackend(DNN_BACKEND_OPENCV);
net.setPreferableTarget(targetId);
// Tokenized with bert-base-uncased: "The [MASK] sat on the mat."
// [CLS]=101 the=1996 [MASK]=103 sat=2938 on=2006 the=1996 mat=13523 .=1012 [SEP]=102
const int seq_len = 9;
int64_t input_ids_data[seq_len] = {101, 1996, 103, 2938, 2006, 1996, 13523, 1012, 102};
int64_t attention_mask_data[seq_len] = {1, 1, 1, 1, 1, 1, 1, 1, 1};
int64_t token_type_ids_data[seq_len] = {0, 0, 0, 0, 0, 0, 0, 0, 0};
int shape[2] = {1, seq_len};
Mat input_ids(2, shape, CV_64S, input_ids_data);
Mat attention_mask(2, shape, CV_64S, attention_mask_data);
Mat token_type_ids(2, shape, CV_64S, token_type_ids_data);
net.setInput(input_ids, "input_ids");
net.setInput(attention_mask, "attention_mask");
net.setInput(token_type_ids, "token_type_ids");
Mat out = net.forward();
// Output shape: [1, 9, 30522] (batch, seq_len, vocab_size)
ASSERT_EQ(out.dims, 3);
ASSERT_EQ(out.size[0], 1);
ASSERT_EQ(out.size[1], seq_len);
ASSERT_EQ(out.size[2], 30522);
const int vocab = 30522;
const float* mask_logits = out.ptr<float>() + 2 * vocab;
const int K = 5;
std::vector<std::pair<int, float> > res;
std::vector<int> idx(vocab);
for (int i = 0; i < vocab; i++) idx[i] = i;
std::partial_sort(idx.begin(), idx.begin() + K, idx.end(),
[&](int a, int b) { return mask_logits[a] > mask_logits[b]; });
for (int k = 0; k < K; k++)
res.emplace_back(idx[k], mask_logits[idx[k]]);
std::vector<std::pair<int, float> > ref = {
{2611, 8.222f}, {2158, 8.196f}, {3899, 8.021f}, {2879, 7.972f}, {2450, 7.393f}
};
const float eps = 0.5f;
std::vector<int> reflabels(K), reslabels(K);
for (int i = 0; i < K; i++) {
reflabels[i] = ref[i].first;
reslabels[i] = res[i].first;
}
ASSERT_EQ(reflabels, reslabels);
for (int i = 0; i < K; i++) {
EXPECT_NEAR(ref[i].second, res[i].second, eps);
}
}
INSTANTIATE_TEST_CASE_P(/**/, Reproducibility_BERT_ONNX,
testing::ValuesIn(getAvailableTargets(DNN_BACKEND_OPENCV)));
typedef testing::TestWithParam<Target> Reproducibility_MobileNetSSD_ONNX;
TEST_P(Reproducibility_MobileNetSSD_ONNX, Accuracy)
{