Load/save for other models than SGD/ALS

[seirl]

The README of this project: https://github.com/jfloff/pywFM states:

Make sure you are compiling source from libfm repository and at this specific commit, since pywFM needs the save_model. Beware that the installers and source code in libfm.org are both dated before this commit. I know this is extremely hacky, but since a fix was deployed it only allows the save_model option for SGD or ALS. I don't know why exactly, because it was working well before.

It seems weird to me that the author hasn't approached you to find a better solution than this hack, and I'm not familiar enough with the code to suggest a PR that would solve the problem cleanly. Besides, currently there's no explicit explanation as to why you forbid to load/save for things other than SGD or ALS, so I don't know where I could do that.

Therefore, I'm making this issue to see if we could find a better solution than this! :-)

@srendle Could you explain what the problem with other models are, and why this check is in place?

@jfloff Could you tell us why pywFM needs to be able to save/load different kinds of models than SGD and ALS?

Thanks!

[thierry-silbermann]

There is only 3 optimizations methods in libFM (SGD, ALS and MCMC).
When using SGD and ALS, you save the model parameters at the end and then you can load them again.
When using MCMC, you will have to save the model parameters at each iteration. The way it is implemented today, we are computing the error directly at each iteration on the test set which allow to avoid to keep the parameters at each iteration. If we were to save model parameters for each iteration, you would end up for huge dataset with a saved model in the order of hundred of GB which is not practical IMO. It will also need some changes from the original code to load those parameters in a smart way and process new data compared to ALS/SGD.

[jfloff]

Thank you for starting this discussion @seirl !

Reason is really simple: I use the save_model to ensure that libfm stores the model into a file, so pywfm can then read/parse it to get info like feature weights. (See implementation here).

Does it make sense to have an option/flag for MCMC to only save the last iteration into the model? Or, by default you get just the last iteration model, and if you pass a special option/flag you get the full list (with a warning regarding the potential size of the file)?