ONNX GPU Cuda

Please change the version of onnxruntime-gpu within requirements.txt if you need more up-to-date version of CUDA!

ONNX Weight Conversion

Please read: note that we never want to export NMS metadata! Why?

• Injects NMS logic into the ONNX graph
• Forces fixed IoU + confidence thresholds
• Breaks threshold sweeps
• Makes per-class F1 analysis painful

Yolov7 Weights

Download the WongKinYiu/yolov7 github repo and open the requirements.txt in the root part of the repo... uncomment the installs for exporting.

Use the export.py script with the following flags below:

python export.py \
  --weights yolov7.pt \
  --img 640 \
  --batch 1 \  
  --dynamic \
  --simplify \ ## Graph cleanup only! Please use AI to ask what this does... before using this flag

The script will generate the yolov7.onnx file

References:

[FasterRCNN] Fasterrcnn submodule sourced from: https://github.com/sovit-123/fasterrcnn-pytorch-training-pipeline [Yolov7] Yolov7 submodule sourced from: https://github.com/WongKinYiu/yolov7