potok/opencv
1
0
Fork 0
mirror of https://github.com/opencv/opencv.git synced 2026-07-30 15:53:03 +04:00
Code Packages Releases Wiki Activity

Merge pull request #28000 from dkurt:d.kurtaev:reset_winograd_impl

Browse Source 
### Pull Request Readiness Checklist

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

resolves https://github.com/opencv/opencv/issues/27580

- [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:
Dmitry Kurtaev
2025-11-13 14:19:46 +08:00
committed by GitHub
parent 78adab13e9
commit 5c02d32cca
2 changed files with 51 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files
modules/dnn/src/layers/convolution_layer.cpp
+8 -3
View File
@@ -253,6 +253,7 @@ public:
Ptr<ActivationLayer> activ;
Ptr<FastConv> fastConvImpl;
bool canUseWinograd = false;
#ifdef HAVE_OPENCL
Ptr<OCL4DNNConvSpatial<float> > convolutionOp;
@@ -448,6 +449,13 @@ public:
#ifdef HAVE_OPENCL
convolutionOp.release();
#endif
// Winograd only works when input h and w >= 12.
canUseWinograd = useWinograd && inputs[0].dims == 4 && inputs[0].size[2] >= 12 && inputs[0].size[3] >= 12;
if (fastConvImpl && (fastConvImpl->conv_type == CONV_TYPE_WINOGRAD3X3) ^ canUseWinograd)
{
fastConvImpl.reset();
}
}
bool setActivation(const Ptr<ActivationLayer>& layer) CV_OVERRIDE
@@ -1292,9 +1300,6 @@ public:
int K = outputs[0].size[1];
int C = inputs[0].size[1];
// Winograd only works when input h and w >= 12.
bool canUseWinograd = useWinograd && conv_dim == CONV_2D && inputs[0].size[2] >= 12 && inputs[0].size[3] >= 12;
CV_Assert(outputs[0].size[1] % ngroups == 0);
fastConvImpl = initFastConv(weightsMat, &biasvec[0], ngroups, K, C, kernel_size, strides,
dilations, pads_begin, pads_end, conv_dim,
modules/dnn/test/test_layers.cpp
+43
View File
@@ -2788,4 +2788,47 @@ INSTANTIATE_TEST_CASE_P(TestLayerFusion, ConvolutionActivationEltwiseFusion, Com
TestLayerFusion::dnnBackendsAndTargetsForFusionTests()
));
TEST(ConvolutionWinograd, Accuracy)
{
Mat weights({2, 1, 3, 3}, CV_32F);
randn(weights, 0, 1);
// Check convolution can switch between implementations on changed shape.
auto getNet = [&]() {
Net net;
LayerParams lp;
lp.name = "conv";
lp.type = "Convolution";
lp.set("kernel_size", 3);
lp.set("num_output", 2);
lp.set("pad", 0);
lp.set("stride", 1);
lp.set("bias_term", false);
lp.blobs.push_back(weights);
net.addLayerToPrev(lp.name, lp.type, lp);
return net;
};
Mat inpSmall({1, 1, 5, 5}, CV_32F);
Mat inpLarge({1, 1, 64, 64}, CV_32F);
randn(inpSmall, 0, 1);
randn(inpLarge, 0, 1);
Net net1 = getNet();
Net net2 = getNet();
net1.setInput(inpSmall);
net2.setInput(inpLarge);
Mat refSmall = net1.forward();
Mat refLarge = net2.forward();
net1.setInput(inpLarge);
net2.setInput(inpSmall);
Mat outLarge = net1.forward();
Mat outSmall = net2.forward();
normAssert(outSmall, refSmall, "Small input after large", 0.0, 0.0);
normAssert(outLarge, refLarge, "Large input after small", 0.0, 0.0);
}
}} // namespace