TensorRT group convolution get wrong results

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

Description

When run convolution with groups != 1 in TensorRT8.2.0, get wrong results.

Environment

TensorRT Version: 8.2.0
GPU Type: 2070super
Operating System + Version: Ubuntu 18.04 LTS
Python Version (if applicable): 3.8

Relevant Files

go GitHub - iilialsa/groupConvolution: Test group convolution in trt8.2.0 ., see my test code.

Steps To Reproduce

When set convolution channel_in=28, channel_out=28, groups=4, it gets right results.
When set convolution channel_in=32, channel_out=32, groups=4, it gets wrong results.

I test this in trt7.x.x, it always get right results. Seems like a bug in trt8.2.0.

2

Hi, Please refer to the below links to perform inference in INT8

Thanks!

4

I am not using int8, my problem happens in fp32.

5

Hi,

Could you please provide issue repro test data and steps, onnx model to try from our end for better debugging.

Thank you.

6

Hi, @spolisetty , sorry for late reply.
My test onnx model just has a convolution layer.
You can run following code to see the different result between pytorch and tensorrt.
code: common.py (6.6 KB)
groupconv.py (3.0 KB)
onnx:groupconv.onnx (9.4 KB)
results:result.txt (8.7 KB)

7

It works well in docker, so my local tensorrt may not be installed well. Sorry to ask this problem.