Possible to export model with non-square input?

[rbavery]

I want to run inference on rectangular inputs since resizing to square degrades performance. Has anyone edited export.py so that imgsz doesn't expect an int and takes a tuple instead? and successfully run yolov5 inference on non square images with onnx or torchscript? looking at the code it looks like if the argparse can take it, it can process imgsz that is not a square.