YOLO-NAS: Running onnx/tensorrt inference with dynamic image size

[janmarczak]

💡 Your Question

Hello,

I was wondering if there is any way to export the YOLO-NAS model to onnx with dynamic image size axes and then convert to TensrRT with dynamic shapes (if needed I can exclude pre and post processing from the onnx)

Basically currently I am using model.predict(img, skip_image_resizing=True) and I would like to have something like this but for onnx/TensorRt.

Versions

No response

[BloodAxe]

The way export currently implemented, exported ONNX graph has static shapes and hence cannot take dynamic input. However it should be possible in theory to export it as dynamic by manually editing the logic of export code and specifying dynamic axis attributes here and there when the model is exported.
One should be really careful with nms sub-graph that is attached to the main model.