How to do inference with trained model?

[013292]

First of all, appreciate for sharing the source code.
Following the instructions of glue-factory, we've got the trained model saved as checkpoint_best.tar which includes parameters of super-point and light-glue. However, the example code in this repo loads the parameters of the extractor and the matcher individually.
Is there any suggestion to close the gap between these two repo?
Any script of loading the trained models for inference?

Thank you :-)

[Phil26AT]

Hi @013292, great that you got your model running with glue-factory! Indeed this small step is missing. What you need to do is open the tar (the content is a dict), and extract the "matcher" item there. These weights should be directly re-usable with LightGlue. You can load the tar with the utils in glue-factory: https://github.com/cvg/glue-factory/blob/1f56839db2242929960d70f85bfac6c19ef2821c/gluefactory/utils/experiments.py#L65-L91

[aidongmandexiaowowo]

Hello, I would like to ask you, I also want to use the trained model for inference, but I still don’t know how to use the experiments.py file in glue-factory to export the matchers weights and extractor weights. Looking forward to your reply