Ulam models can't be fit when saved with a blank dataframe

[robbiekno]

This is a weird one, but I wanted to make sure you knew.

I'm using ulam models in a project, usually saving them in a Rdata file alongside the out-of-sample predictions from the models. However, there are some cases when there are no predictions to be saved because the model is being fit on live data, so the model was being saved alongside a blank dataframe (x <- dataframe() ). When I tried to bring in those models in another script, they weren't working and I was getting errors that looked like this when running the link function:

The exact name of the file was slightly different each time, but the errors were exactly the same otherwise. However, when I altered my script just slightly, replacing dataframe() with a 0-row dataframe featuring the usual columns of the data, suddenly everything was fine.

This whole thing freaked me out for a while and made me wonder if I was encountering major problems with my models before I figured it out, and I wanted to make sure anyone running into something similar knew. I also don't know if there is something to be fixed in terms of the rethinking package to prevent errors like this.

UPDATE: Now things aren't working again, so I'm trying to figure out next steps. Any help would be appreciated.

[robbiekno]

Following up on this with a final resolution, the key was to save the model alongside a dataframe that contained predictions from the model (using the link function). A dataframe of any size, containing real rows or fake rows but not predictions from the model, didn't do the trick. It seems like having a dataframe that contained real predictions provided some cmdstan files / metadeta that was necessary to predict from the model in the future.

[rmcelreath]

This doesn't yet make sense to me. If you have a script with a minimal example, I can probably isolate the problem. Although sometimes the cmdstanr internals change on me and these bugs are hard to figure out. But I can try.