cboettig
commented
1 year ago Definitely an interesting idea, torch seems pretty compelling here. Looks like interesting things being built on it recently, e.g. thought you might like to take a look at cito https://cran.r-project.org/web/packages/cito/vignettes/Introduction_to_cito.html if you haven't seen it.
At the moment we are looking at investing our time into getting tensorflow 2.0.0 working with greta seamlessly.
However, in the future it might be something of interest to look at an alternative approach, such as torch.
One of the main advantages with using torch is that it has a simpler installation process, and once you install the R version of torch (https://github.com/mlverse/torch), it will just work, and doesn't require a separate installation as done with tensorflow.
The obvious large disadvantage here is that all the code is currently written with tensorflow and would require a large overhaul.
Another option is that we could do both, which would require abstracting away specific parts of tensorflow / torch so that they could be turned on, or swapped out.
A lot of this discussion on this topic was done here: https://github.com/greta-dev/greta/issues/247 by Davis and Nick Golding, but I just thought it might be useful to have a specific issue mentioning changing backends, as while it isn't something we want to do right now, it might be something that happens in the future.