Galileo config for pastis

[gabrieltseng]

Depends on allenai/rslearn#272

[robmarkcole]

Is this just using the dataset as is from https://zenodo.org/records/5012942 ? I guess not, in which case it would be nice to know how it is generated, presumably with rslearn

[favyen2]

Is this just using the dataset as is from https://zenodo.org/records/5012942 ? I guess not, in which case it would be nice to know how it is generated, presumably with rslearn

Here is the script used to convert the data:

https://github.com/allenai/rslearn_projects/blob/master/one_off_projects/2025_05_13_pastis/convert.py

It is just converted to rslearn dataset format, but the images still come from the original dataset (not generated by rslearn).

[robmarkcole]

@gabrieltseng it is also necessary to include the required config.json, might be worth documenting these.

What machine did you use for training? On an H100 I get Tried to allocate 112.50 GiB. GPU 0 has a total capacity of 79.44 GiB of which 50.23 GiB is free with batch_size 1

[gabrieltseng]

@robmarkcole I have been using GPUs with less memory than a H100 for testing. It may be worth increasing the patch size or decreasing the model size if you are running into memory issues