I thought I'd share some info to help get started. We also just did a live stream, so if you want a video reference for connecting to the VM you can check it out here: https://www.twitch.tv/videos/648634668. But assuming that like me you're a text person, here's the juice:

- The training locations all fall within a single image tile, which I copied to the VM using `aws s3 cp s3://eohackathon-covid19/Hackthon_Data/Gauteng/2528C.tif 2528C.tif` (you need to set secret keys and things - see the README.yaml file for instructions). The test locations all fall within 2930D.tif. SO: you don't need to copy all the imagery from s3 to get going - just these two tiles (4GB each).

- The GP settlement layer is available as a shapefile, which means you can generate more training data if you want. BUT: careful with the class balance. Both train and (spoiler alert) test have ~20-30% positives (informal settlements) - randomly sampling locations will likely get you closer to 2% positives, and thus might give a model that does worse on the test set.

- The test set comes from an entirely different province. Check out some of the imagery and you'll notice it's from a fairly urban area, but with a good mix of land-use classes. If you're doing a random split on the training data, you'll see high accuracy (95%) locally but will get a nasty surprise when your model is scored on the leaderboard (eg 75% accuracy, and log_loss of ~0.7 or whatever, higher than the 0.2 seen in training). THINK ABOUT HOW TO MAKE A BETTER LOCAL VALIDATION SET. Maybe split by latitude, or generate a new validation set from a different location in GP...

- Make sure to save your notebook to your local machine after making a submission - the VMs disappear at the end of the weekend, and you don't want to be stuck with no code to submit :)

- Finally, although this is a competition, we're all trying to learn things. If you find a nice way to speed up something like image access, or have a nifty trick for generating more training locations, or you've done an image segmentation model using the shapefile as a mask, or you figured out how to install unrar... SHARE :) It's so nice as a beginner to get help from others and see how they've overcome challenges. Add tips in this thread or start your own.

Good luck, and have fun :)

JW