onnx export

[GallonDeng]

great work! since radio could be used to get sota feature representation, it may be very useful for some tasks such as image retrieval or video retrieval. Also, is there example for onnx model export of radio, which could be easily integrated in downstream task just mentioned before.

[mranzinger]

We have an overly complicated example of doing an onnx export across the suite of models we analyzed: https://github.com/NVlabs/RADIO/blob/main/examples/count_params.py (Just uncomment the code around the onnx export). Have you run into problems using it? I think with opset=17, it should hopefully be pretty straightforward.

[GallonDeng]

thanks a lot @mranzinger ; I tried the 'count_paramas.py' and successfully exported onnx model for e-radio2. Here are some questions: 1. at least which version should be used for converting the onnx model to trt engine? it fails for trt 8.4; 2. the exported default e-radio2 onnx model still too large to show with netron tool, how to use a smaller version of e-radio2 since in the codabase there are many versions of registered e-radio2 for different configuration.

[mranzinger]

Good questions. @gheinrich is much more of an expert on e-radio + tensorrt things. Greg, could you shed some light on this?

[gheinrich]

Hello, I used TensorRT 9.0 to export E-RADIO2. If you're looking for a smaller version of E-RADIO, you can use eradio_xxxtiny though it might not be much easier to visualize. It is not shallower than the full-size version, only slimmer.