search

allenai / rslearn

A tool for developing remote sensing datasets and models.
Apache License 2.0
5 stars 0 forks source link

Move Quickstart back into README and walk through model training and prediction #77

Closed favyen2 closed 4 days ago

favyen2 commented 1 week ago

There are some TODOs that can be completed later I think.

pbeukema commented 4 days ago

I have been through almost (still waiting on a materialize step to see the results) all of the steps.

My main feedback, as I mentioned yesterday, is that this is not a quick start -- it is quite long to execute and somewhat tedious with all of the commands. I would call it instead installation/setup/tutorial/something along those lines so that the reader is clear on what to expect. That said, I think that's fine depending on the target audience.

But if the goal is to have a true "quickstart" I think that can be accomplished, but it needs to be much faster and easier to execute all of these steps (and that may be out of scope to do before the intended OS release, or not desired/impossible).

I don't know what is the best approach but some combination of the following might be useful:

  1. A docker image which includes the dataset and the updated configuration files and a script that did all of the required steps without having the user manually executing each one.
  2. Stricter limits on how much data is getting processed. In particular materialize takes a lot of time.
  3. Recommendation for a much larger machine (CPUs).
  4. Moving the world_cover sample image to a GCP repository so that we can leverage faster downloads (such as gcloud alpha storage cp)
favyen2 commented 4 days ago

OK thanks! I changed it to Example Usage for now. I think in the future we can have one that's faster with less data and just using one of the WorldCover GeoTIFFs.