fastrtext Add autotune option

Hi there,

I saw on the Fasttext page here they've added an autotune feature, which automatically optimizes the various hyperparameters.

Seems it can be activated with the -autotune-validation option, which isn't currently supported. Wondered if this could be added with the updates for CRAN?

https://fasttext.cc/docs/en/autotune.html

best

Alan

Sep 19 '19 10:09 alanault

Hi,

Will work on it since I need to fix something else to be back on Cran... Hope next week I will have some time to work on it.

Sep 20 '19 12:09 pommedeterresautee

Sounds great - intrigued to see the results of this output

Sep 24 '19 12:09 alanault

Hi there, just wondered how this was coming on?

I'm not familiar with Rcpp, but happy to help on any R-based areas if that's useful?

Oct 21 '19 08:10 alanault

Hello, I have updated the C++ code. Should work using command line. Do you think it would make sense to have a R API for this?

Oct 28 '19 08:10 pommedeterresautee

Hi: that's fantastic!

Yes, I was thinking it could just be exactly the same as the other calls, just with the -autotune-validation passed as an option in the execute command, along with the validation file.

So, to use the supervised example: execute(commands = c("supervised", "-input", train_tmp_file_txt, "-output", tmp_file_model, "-autotune-validation", valid_tmp_file_text))

That way everything is consistent within the execute command?

Oct 28 '19 08:10 alanault

I think so. Not yet tried. If you check, can you let me know if it works?

Oct 28 '19 11:10 pommedeterresautee

Installed 0.3.4 So the call seems to work just fine... however each time I try it I get a C stack usage crash (is too close to the limit).

I'm only training on 100k short sentences and it crashes within seconds at (0.8% of completion), so wonder if something else is going on?

Oct 28 '19 15:10 alanault

does it crash in other situations?

Oct 28 '19 17:10 pommedeterresautee

No - if I remove the -autotune-validation argument and just make the same call train, then it runs just fine on the same data, so must be related to the validation...

Oct 28 '19 18:10 alanault

Did a test using the command line version and this worked fine, so the issue must be somewhere inside the rcpp wrapper

Oct 30 '19 08:10 alanault

Any luck with this?

Nov 19 '19 14:11 alanault