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 #19058 from sl-sergei:cuda_1d

Browse Source 
Conv1D and Pool1D for CUDA backend

* CUDA-independent changes

* Add Conv1D and Pool1D for CUDA backend

* CUDA-independent changes

* Fix typo

* fix comment

* Update fix

* make changes more correct for pooling layer

* Minor fixes for review

* Split skip blocks
This commit is contained in:
Sergei Slashchinin
2021-01-22 01:16:56 +03:00
committed by GitHub
parent 7a790d0d35
commit ea41f89b40
6 changed files with 74 additions and 17 deletions
Download Patch File Download Diff File Expand all files Collapse all files
modules/dnn/src/layers/convolution_layer.cpp
+20 -2
View File
@@ -125,6 +125,9 @@ public:
{
kernel_size.assign(1, kernel_size[0]);
strides.assign(1, strides[0]);
dilations.assign(1, dilations[0]);
pads_begin.assign(1, pads_begin[0]);
pads_end.assign(1, pads_end[0]);
}
CV_Assert(weightShape.dims() == kernel_size.size() + 2);
for (int i = 0; i < kernel_size.size(); i++) {
@@ -311,8 +314,8 @@ public:
#ifdef HAVE_CUDA
if (backendId == DNN_BACKEND_CUDA)
{
/* only convolution 2d and 3d supported */
if (ksize == 2 || ksize == 3)
/* only 1d, 2d and 3d convolutions supported */
if (ksize > 0 && ksize <= 3)
return true;
return false;
@@ -2001,6 +2004,21 @@ public:
const auto groups = input_feature_maps / input_feature_maps_per_group;
ConvolutionConfiguration config;
if (input_shape.size() == 3)
{
// Conv1D
// We add an extra dim for input and output tensors, because CuDNN doesn't support convolution with 3D tensors
input_shape.insert(std::end(input_shape) - 1, 1);
output_shape.insert(std::end(output_shape) - 1, 1);
// Do the similar thing for the other parameters
pads_begin.insert(std::begin(pads_begin), 0);
pads_end.insert(std::begin(pads_end), 0);
strides.insert(std::begin(strides), 1);
dilations.insert(std::begin(dilations), 1);
kernel_size.insert(std::begin(kernel_size), 1);
}
config.kernel_size.assign(std::begin(kernel_size), std::end(kernel_size));
config.dilations.assign(std::begin(dilations), std::end(dilations));
config.strides.assign(std::begin(strides), std::end(strides));
modules/dnn/src/layers/pooling_layer.cpp
+18 -6
View File
@@ -178,14 +178,13 @@ public:
if (inputs[0].dims == 3)
{
//Pool1D
kernel_size.erase(kernel_size.begin() + 1);
strides.erase(strides.begin() + 1);
pads_begin.erase(pads_begin.begin() + 1);
pads_end.erase(pads_end.begin() + 1);
// Pool1D
kernel_size.assign(1, kernel_size[0]);
strides.assign(1, strides[0]);
pads_begin.assign(1, pads_begin[0]);
pads_end.assign(1, pads_end[0]);
}
#ifdef HAVE_OPENCL
poolOp.release();
#endif
@@ -392,6 +391,19 @@ public:
return make_cuda_node<cuda4dnn::MaxPoolingOp>(preferableTarget, std::move(context->stream), config);
}
if (input_shape.size() == 3)
{
// Pool1D
// We add an extra dim for input tensor, because CuDNN support pooling only with 2 and 3 spatial dimensions
input_shape.insert(std::end(input_shape) - 1, 1);
// Do the similar thing for the other parameters
pads_begin.insert(std::begin(pads_begin), 0);
pads_end.insert(std::begin(pads_end), 0);
strides.insert(std::begin(strides), 1);
kernel_size.insert(std::begin(kernel_size), 1);
}
PoolingConfiguration config;
if (type == MAX)
{