GASSL ResNet50 Weights

[isaaccorley]

This PR adds the ResNet50 RGB weights trained on fMoW (mostly Maxar WorldView imagery) using the Geography Aware Self-supervised Learning (GASSL) method. This is basically a modified MoCo-v2 with Geolocation as a pretext task (Geo) as well as temporal pairs contrastive learning (TP).

Weights taken from this repo and uploaded to huggingface.

Adding these weights is essential since it's being compared as a baseline in both the SatMAE and Scale-MAE papers.

[adamjstewart]

Need to add to docs/api/resnet_pretrained_weights.csv. We should start thinking about how we want to handle multiple satellites. I'm thinking we'll want different tables for each satellite since the benchmark datasets will differ. We can punt that to the PR where we add all our new SSL4EO-L models.

[isaaccorley]

Should I add the transforms as well? They only train on the fmow RGB data which is uint8 and they just divide by 255 and use imagenet normalization. Nvm, needed to add transforms.

[isaaccorley]

No idea why coverage is failing

[calebrob6]

This LGTM, @adamjstewart any problems with merging this?

[adamjstewart]

Give me some time to review. How urgent is this? Would prefer to focus on the paper until the deadline, but can review sooner if needed. I mostly want to verify the transforms and filename and such.

[calebrob6]

I've checked the sha256sum and transforms so I think we're good

[adamjstewart]

Do we want to call it fMoW or Maxar? Might be more clear what the weights can be used for with the latter.

[calebrob6]

The current name fits with the existing scheme {model}_{dataset}_{bands}_{paper/source}_{hash} -- regardless, users aren't interacting with the actual file.

[adamjstewart]

The existing scheme uses image source, not dataset.

[isaaccorley]

Fmow RGB is more than just Maxar though.

[adamjstewart]

Why did you choose Apache 2.0 for the license?

[adamjstewart]

Also, should we confirm the license of the pre-trained model before we modify and redistribute it? If it's unlicensed then technically we can't use it. If @calebrob6 doesn't care then I don't care, it's Microsoft that will take the blame.

[calebrob6]

Where are you seeing "Apache 2.0"?

[adamjstewart]

https://huggingface.co/torchgeo/resnet50_fmow_rgb_gassl

[calebrob6]

It seems the weights aren't licensed -- but just to be clear, we are not uploading the weights to this repo. It is always good to ask the authors though.

[calebrob6]

It seems the weights are licensed here (https://zenodo.org/record/7379715/) as Creative Commons Attribution 4.0 International, which is quite open.

[isaaccorley]

I'll change the license on HF

[calebrob6]

Already did