Configuration for recreating geo bench results.
Hi - Thank you for this work. I have to ask a question - I was trying to recreate the results on geo bench - however in the paper - you only mention that you do a grid search over [0.5, 1, 10, 20] for learning rate but don't mention the actual learning rate used which could make a huge difference given the wide range. Additionally, you mention using LARS as the optimiser but the implementation of same doesn't exist in PyTorch. Could you tell if you implemented yours or if you used some library? I don't see it in the repo. The repo doesn't seem to have the training code - only architecture. Am I correct in understanding this or am I missing something?
Also while going through the code I notice for segmentation you have - frozen_exclude=['all'], so basically you are fine-tuning the whole model for segmentation with a UPerNet head? Am i correct in understanding that?
You are correct.
We are in the process of adding more experiments in the full fine-tuning setting and updating the Arxiv paper.
Pre-training codes and experiments on the geobench will be uploaded soon.
Thanks for your interest. We will let you know when updated.
Please stay tuned!