torchgeo icon indicating copy to clipboard operation
torchgeo copied to clipboard

Add plot method to IntersectionDataset

Open adamjstewart opened this issue 10 months ago • 1 comments

Summary

Add a plot method to IntersectionDataset that calls the plot method of all GeoDatasets and merges their output into a single figure.

Rationale

TorchGeo currently relies on the following fundamental design principle:

  • Each dataset has a plot method that returns a single Figure
  • Each data module has a plot method that calls dataset.plot and returns a single Figure
  • Each trainer calls this plot method during validation and logs a single Figure to tensorboard

This works great for most NonGeoDatasets and GeoDatasets, but falls apart as soon as you use an IntersectionDataset. Merging two Figures together is non-trivial, and so far we've simply ignored this use case by plotting the mask/prediction but not the image. This prevents proper visualization of training and makes IntersectionDatasets different from other GeoDatasets in terms of features.

Implementation

This idea can be implemented in multiple stages (possibly by multiple people).

First, we'll need to add an abstract plot method to GeoDataset. There are a few point datasets that do not yet have plot methods (EDDMapS, GBIF, INaturalist), so we'll have to add one. Might as well add one to NonGeoDataset as well to enforce consistency. Regardless, this is likely a good idea to do, it's just a prerequisite for what follows.

Second, we'll need to add a new parameter to all GeoDataset.plot methods containing an optional pointer. The idea is that normally, GeoDataset.plot will create its own Figure and returns it, but IntersectionDataset.plot will instead create a Figure and call each GeoDataset.plot method with the same sample and this additional pointer telling it which subplot to populate. Or maybe pass it an existing figure and have it select the next empty subplot to use? The number of subplots to create in IntersectionDataset.plot could either be based on the number of datasets (would exclude plotting prediction maps) or on the number of image/mask/prediction elements in the sample dict (relies on #985 if we want any guarantees). See Additional information for caveats.

Alternatives

We could instead change GeoDataset.plot and GeoDataModule.plot to return one or more Figures and our trainers to plot a list of figures. This may actually be significantly simpler to implement. The only downside is that not all datasets will be consistent, and it would be nice to be able to display a single Figure with all information.

Additional information

Some remaining TODOs to figure out:

  • How will this work for UnionDataset?
  • How will this work for IntersectionDataset with multiple images (e.g., Landsat & Sentinel)?

This is related to #1263 and could likely be done in parallel.

adamjstewart avatar Apr 01 '24 09:04 adamjstewart

It may actually be possible to add new subplots dynamically: https://stackoverflow.com/questions/12319796/dynamically-add-create-subplots-in-matplotlib

adamjstewart avatar Apr 01 '24 12:04 adamjstewart