Subclassing fastai classes for different data sources?

I’m working with satellite data (multi-spectral). Intuitively it makes sense to me to use fastai by subclassing what is already there — either at the level of the application (i.e. making a ‘multispectral’ application in parallel to the existing vision, etc.), or by subclassing classes in the vision application to add more channels.

It is a pretty large amount of code either way, so I’m looking for advice on what approach to take before I simply jump in.

Has anybody had experience doing this — or is there even maybe a blog article somewhere on it? (I’ve looked but not found any.)


There is a tutorial here on how to create a custom ItemList. There is also an blog post about subclassing the classes of the data block API (though not in vision I believe).

1 Like

Thank you, these are both super helpful!