scatterDense operator output is all zeros in centerpoint's exported scn.onnx(sparseconvolution)

AI & Data Science Deep Learning (Training & Inference) TensorRT
1

Description

According to Nvidia github project CUDA-centerpoint(here), I exported my custom centerpoint scn.onnx model, the onnx exporting procedure is similar to functions described in export-scn.py,exptool.py,funcs.py. the Nvidia CUDA-centerpoint’s input of scn.onnx is 1x5(trained on nuscenes dataset ), output si**ze is 1x256x200x200. my model’s scn.onnx’s input is 1x4(trained on KITTI dataset),output of bev size is 1x256x200x176. both offical scn.onnx and my custom scn.onnx conatain scatterDense-> reshape->1x256x200x200 procedure. The tensRT infrence of offical scn.onnx is successful, BUT my scn.onnx’s scanDense’s output are all zeros

I check data flow from input 1*4 to last part of sparseconvolotions, the results are correct before scatterDense operator, it seems there are some problems in my scn.onnx exporting process or scatterDense operator itself.

Environment

TensorRT Version: 8.6.1.6
GPU Type: RTX3060
Nvidia Driver Version: 535
CUDA Version: 11.3
CUDNN Version: 8.2.1
Operating System + Version: ubuntu 20.04

Relevant Files

CUDA-centerpoint
my export-scn.py,exptool.py,funcs.py are( *.py are not allowed, use txt here )
exptool.txt (26.2 KB)
export_scn.txt (9.7 KB)
funcs.txt (23.0 KB)

SCN_ONNX
SCN_ONNX1745×999 76.1 KB

Qutestion

why scatterDense’s output is all zeros?

4

Hi ,

I would recommend -

  • Checking the ONNX Model: Verify that the ONNX model is correctly exported and that the scatterDense operator is properly defined.

  • Comparing Model Implementations: Compare the implementation of the scatterDense operator in the official SCN model and the custom model to identify any differences.

  • TensorRT Logging: Enable verbose logging in TensorRT to gather more information about the inference process and the scatterDense operator output.

  • *Input and Output Sizes: Check that the input and output sizes of the custom model match the expected sizes for the scatterDense operator.

    If issue persist, please share the onnx model with us.*

5

I have checked the ONNX model, and have compared the official model and my custom model, have compared pth model results and my custom model results, The problem still remained.

here are my custom ONNX model and modified deployment source code
centerpoint_scn_onnx.zip (1.3 MB)
The ONNX exporting source code is attatched in the 1st post