HopeJW

By default, libspconv is running on dynamic input size. The input size (5, 1) does not affect the pipeline. It's just a placeholder!

Yes, but you need to export the SwinT model first.

```c++ typedef unsigned short half; static inline half __internal_float2half(const float f) { unsigned int x; unsigned int u; unsigned int result; (void)memcpy(&x, &f, sizeof(f)); u = (x & 0x7fffffffU); sign...

I will check this issue. Thanks a lot.

Try to remove the `-Wall` flags: CUDA-BEVFusion/CMakeLists.txt ```CMake set(CMAKE_CXX_FLAGS_RELEASE "-std=c++14 -Wextra -Wall -Wno-missing-field-initializers -Wno-deprecated-declarations -O3 -DENABLE_TEXT_BACKEND_STB") set(CMAKE_CXX_FLAGS_DEBUG "-std=c++14 -O0 -g -DENABLE_TEXT_BACKEND_STB") set(CUDA_NVCC_FLAGS_RELEASE "-Werror=all-warnings -Xcompiler -std=c++14,-Wextra,-Wall,-Wno-deprecated-declarations,-O3 -DENABLE_TEXT_BACKEND_STB") set(CUDA_NVCC_FLAGS_DEBUG "-std=c++14 -O0 -g...

The 2d SPCONV is a specific form of 3d SPCONV. You just set z to 1 at indices which can achieve the 2d SPCONV.

@HuangVictorAuto could you give me a reproduce data? For example a simple pytorch code and a simple input data(features, indices, grid_size).

You can try the below commands: ```bash git clone https://github.com/mit-han-lab/bevfusion cd bevfusion git checkout db75150717a9462cb60241e36ba28d65f6908607 ```

yep, mmdet3d also supported by [export-scn](https://github.com/NVIDIA-AI-IOT/Lidar_AI_Solution/blob/master/CUDA-BEVFusion/qat/export-scn.py).

It seems the wrong transformation matrix is used. Here is a recommended video that I hope can help you: https://www.bilibili.com/video/BV1ZX4y187yh