mirror/opencv
1
0
Fork 0
mirror of https://github.com/opencv/opencv.git synced 2025-01-18 22:44:02 +08:00
Code Issues Actions Packages Projects Releases Wiki Activity

Merge pull request #17284 from dkurt:dnn_bn_fusion

Browse Source
This commit is contained in:
Alexander Alekhin 2020-05-14 06:57:36 +00:00
commit bcf96d637e
2 changed files with 66 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

9
modules/dnn/src/layers/batch_norm_layer.cpp
View File

@ -94,6 +94,15 @@ public:
dstWeightsData[i] = w;
dstBiasData[i] = (hasBias ? biasData[i] : 0.0f) - w * meanData[i] * varMeanScale;
}
// We will use blobs to store origin weights and bias to restore them in case of reinitialization.
weights_.copyTo(blobs[0].reshape(1, 1));
bias_.copyTo(blobs[1].reshape(1, 1));
}
virtual void finalize(InputArrayOfArrays, OutputArrayOfArrays) CV_OVERRIDE
{
blobs[0].reshape(1, 1).copyTo(weights_);
blobs[1].reshape(1, 1).copyTo(bias_);
}
void getScaleShift(Mat& scale, Mat& shift) const CV_OVERRIDE

57
modules/dnn/test/test_layers.cpp
View File

@ -1780,4 +1780,61 @@ TEST_P(Layer_Test_Slice, variable_input_shape)
INSTANTIATE_TEST_CASE_P(/**/, Layer_Test_Slice, dnnBackendsAndTargets());
typedef testing::TestWithParam<tuple<Backend, Target> > Layer_Test_BatchNorm;
TEST_P(Layer_Test_BatchNorm, fusion)
{
// This tests reinitializes network by forwarding different batch size input.
// We check BatchNorm layer weights restoring after fusion.
int backendId = get<0>(GetParam());
int targetId = get<1>(GetParam());
const int ch = 4;
Mat mean(1, ch, CV_32F), var(1, ch, CV_32F), weights(1, ch, CV_32F);
randu(mean, 0, 1);
randu(var, 0, 1);
randu(weights, 0, 1);
Net net;
{
LayerParams lp;
lp.type = "BatchNorm";
lp.name = "bn";
lp.set("has_weight", false);
lp.set("has_bias", false);
lp.blobs.push_back(mean);
lp.blobs.push_back(var);
net.addLayerToPrev(lp.name, lp.type, lp);
}
{
LayerParams lp;
lp.type = "Scale";
lp.name = "scale";
lp.set("has_bias", false);
lp.blobs.push_back(weights);
net.addLayerToPrev(lp.name, lp.type, lp);
}
Mat inp(4, 5, CV_32FC(ch));
randu(inp, 0, 1);
net.setPreferableBackend(backendId);
net.setPreferableTarget(targetId);
net.setInput(blobFromImage(inp));
Mat ref = net.forward();
net.setInput(blobFromImages(std::vector<Mat>(2, inp)));
Mat out = net.forward();
for (int i = 0; i < 2; ++i)
{
std::vector<Range> ranges(4, Range::all());
ranges[0].start = i;
ranges[0].end = i + 1;
normAssert(out(ranges), ref);
}
}
INSTANTIATE_TEST_CASE_P(/**/, Layer_Test_BatchNorm, dnnBackendsAndTargets());
}} // namespace