NotImplementedError when training FarSeg model from torchgeo.models

Description of Issue When attempting to use one of the models built into torchgeo, from torchgeo.models import FarSeg I receive a NotImplementedError when running the cell in Jupyter Notebook that performs the fitting of the model using a PyTorch Lightning Trainer. I have uploaded the source code for the CloudFarSegModel class and the CloudDataset class to this repository here. The competition dataset is nearly 30 gb so it would have to be downloaded from the AzureBlobStorage source.

FarSeg_NotImplementedError.txt

Credit to Katie Wetstone for source code from the blog/notebook to train a benchmark Unet model on the dataset: https://www.drivendata.co/blog/cloud-cover-benchmark/

[isaaccorley]

@KennSmithDS feel free to make a separate issue if you run into any problems with training on your custom datamodule using one of ours as template so we can document this for other users.

Also feel free to make a PR for the the Cloud Cover challenge dataset and datamodule. It would be a good first contribution and it would be great to have it included in torchgeo.

[calebrob6]

do I need to create a custom data module for this cloud cover dataset then to pass into the datamodule= parameter for the Trainer?

Yes, exactly! We have a list of (pytorch lightning modules (i.e. trainers), datamodules) here that train.py uses to run experiments with. These pairs are exactly what get fed into the pl.Trainer object.