static spatiotemporal train/val/test splits for dataset

[CoenvdE]

Is your feature request related to a problem?

statis train val and test splits for testing DL models

Describe the solution you'd like

Is your feature request related to a problem?
Hi everyone,

I’d like to open the discussion about adding functionality to create static spatiotemporal train, validation, and test splits from a single large xarray dataset (e.g., global-scale data).

The goal is to generate these splits based on:

The spatial and temporal size of each sample
The stride between samples
A list of user-defined validation and test regions (as static spatiotemporal holdouts)
Potentially a land mask, used to ignore samples which only contain ocean, or land data.
The expected output would be, for each split, a list of dictionaries containing the coordinates (start and end) of each slice. This structure would make it straightforward to iterate over samples in a dataloader.

Keeping the validation and test regions static is important to ensure consistent model evaluation and comparability across experiments.
I am already working on a prototype for this with my work, and wanted to open the discussion to gather feedback and seeing if and to which library this could be a contribution!

Describe alternatives you've considered

Making a custom, its not done anywhere

Additional context

No response

[s-kganz]

Hi, thanks for this suggestion! Splitting data is important. However, I'm not sure whether that functionality needs to go into xbatcher itself because xarray gives us the tools to do that already. For example, you could do something like:

# da is a data array that I want to make batches from
da_train = da.isel(time=slice(None, 2005))
da_valid = da.isel(time=slice(2005, None))

train_bgen = BatchGenerator(da_train, ...)
valid_bgen = BatchGenerator(da_valid, ...)

I could imagine similar approaches for spatial slices, or using rioxarray to clip to a specific geometry. Does that accomplish what you are looking for?