Generating INT8 Calibration Table

According to documentation, we need a Calibration Table to export the .onnx file as an INT8 optimized .plan file

export{.exe} model.onnx engine.plan INT8CalibrationTable

I’m assuming that in order to generate the calibration table, I need to do the export on the host machine.

My machine currently uses a RTX 3090 and the 20.03 repo is running in a docker container. However, every time I try to do the export on the host machine, the program aborts. I’ve tried to vary the --size and --calibration-batches but still no luck.

Here is an example of the output that I usually get.

odtk export /models/retinanet-kitti-rn50fpn_fp.pth /models/retinanet-kitti-rn50fpn_8int.plan --int8 --calibration-images /dataset/Kitti-Coco/validate/data --calibration-batches 12  --size 640
Loading model from retinanet-kitti-rn50fpn_fp.pth...
     model: RetinaNet
  backbone: ResNet50FPN
   classes: 8, anchors: 9
Exporting to ONNX...
Building INT8 core model...
Building accelerated plugins...
Applying optimizations and building TRT CUDA engine...
Half2 support requested on hardware without native FP16 support, performance will be negatively affected.
Int8 support requested on hardware without native Int8 support, performance will be negatively affected.
Assertion failed: Unsupported SM.
../rtSafe/cuda/caskUtils.cpp:80
Aborting...

../rtSafe/cuda/caskUtils.cpp (80) - Assertion Error in trtSmToCask: 0 (Unsupported SM.)
Writing to /models/retinanet-kitti-rn50fpn_8int.plan...
Segmentation fault (core dumped)

Any help would be greatly appreciated.