Fail to convert peoplenet to int8 engine

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

Environment

TensorRT Image: nvcr.io/nvidia/tensorrt:24.03-py3
GPU Type: NVIDIA GeForce RTX 3050
Operating System + Version: WSL
PeopleNet ONNX and calibration file: PeopleNet | NVIDIA NGC

Steps to reproduce:

  1. docker run --gpus all -it --rm -v /home/CV-Framework:/workspace/CV-Framework nvcr.io/nvidia/tensorrt:24.03-py3
  2. trtexec --onnx=/workspace/CV-Framework/model_repository/peoplenet_int8_onnx/1/resnet34_peoplenet_int8.onnx --maxShapes=“input_1:0”:16x3x544x960 --minShapes=“input_1:0”:1x3x544x960 --optShapes=“input_1:0”:8x3x544x960 --calib=/workspace/CV-Framework/model_repository/peoplenet_int8_onnx/1/resnet34_peoplenet_int8.txt --best --saveEngine=/workspace/CV-Framework/model_repository/peoplenet_int8_onnx/1/resnet34_peoplenet_int8.plan
    Note: The trtexec command works when --fp16 is flagged but fails when --int8 or --best is flagged

Error:

image
image1340×820 24.9 KB

Appreciate your advice on this. Thanks!

3

Hi @shtan6 ,
Can you please share the verbose logs with us.

Thanks

4

Hi,

I flag the verbose option in trtexec command and the logs is attached.

Thank you
logs.txt (15.8 KB)

5

Hi,

Appreciate your help on this matter.

Thx!