search

edwardzhou130 / PolarSeg

Implementation for PolarNet: An Improved Grid Representation for Online LiDAR Point Clouds Semantic Segmentation (CVPR 2020)
BSD 3-Clause "New" or "Revised" License
373 stars 80 forks source link

ONNX conversion: input tensor size #21

Open zhangnju opened 3 years ago

zhangnju commented 3 years ago

Hi, when doing ONNX conversion for pytorch models , we need to create a dummy input variable , like (3,224,224) for Resnet50. if need to do onnx conversion for polarseg, how shall we create this dummy input variable? what is the size of input tensor?

Thanks

edwardzhou130 commented 3 years ago

I am not very familiar with ONNX. But we design PolarNet to be able to take a random size of the input point cloud feature (K*N, K is the feature channel number and N is the number of points). So, this might be a problem for the ONNX conversion. If it needs to be a fixed size, you can try to modify our BEV encoder network (pointnet + scatter max pooling) to a similar structure in PointPillar. They first create a fixed-sized tensor by padding the empty space with zero before feed it to the network.

kaxapatel commented 1 year ago

HI @zhangnju were you able to export the model? I am working with TensorRT implementation. Any help can be appreciated.