I don't think it's that expensive if you don't count paying yourself for the time.
You can develop these in a shared Google hosted Jupiter notebook and run pytorch locally to test while you go.
But to train you would need to rent a well appointed resource, like a dgx slice on digital ocean or wherever.
I've been informed by csb that tensor flow and pytorch are old technologies now.
Matt Hamilton (npub1gx2…4lmh)