Multiple roi in Sampler

[Modexus]

I have an issue with multiple roi's passed to a Sampler.

I am trying to use the same train/val/test set as Tile2Vec. In the paper they split a singular roi in a 12x12 grid. Every grid cell is then assigned randomly to one of the three datasets.

The issue lies in that the sampler only take a single roi as an argument (the datasets as well). I can create the 12x12 grid but I am unable to assign multiple non-contigous regions to the train/val/test set.

Is it possible to somehow pass multiple regions as the roi argument (I am not familiar with rtree)? My current workaround is changing GeoSampler to handle mutiple regions.

[adamjstewart]

I think you're right that it isn't possible with the current setup. I'm open to suggestions, either allowing multiple roi in a GeoSampler, or some kind of dataset_split method that supports grids.

[Modexus]

Splitting the dataset directly does not seem feasible at the moment as the dataset does not allow for multiple rois (this would also be a good feature, for example using multiple regions of NAIP distributed over the country).

As mentioned my current workaround involves changing the GeoSampler to support multiple rois as well as using a roi_split_grid function that is passed to the sampler. This seems nice as the dataset does not have to be changed, just the samplers roi restricted (more generically as a roi_split function can be passed).

This does not work yet as it interferes with #537 but once I remove these it should.

[adamjstewart]

This feature was added by #866, see torchgeo.datasets.random_grid_cell_assignment.