Assertion failed when fitting with l2_reg_w=0 l2_reg_V=0 about fastfm

[ibayer]

There are a few situations where this is expected. For example if you have a feature that's zero for all examples. One could add input checks for this situations or just always use some regularization (could be very small).
Maybe adding a warning that using FM without regularization is not recommenced would be sensible way to deal with this issue.
I think that regularization is (nearly) always necessary for good performance.

[merrellb]

I agree it may not be a great idea but given that it seems to be failing at an assertion I wondered if it might be possible to do it more gracefully (or less cryptically). Sorting out the cause is particularly challenging given that these values are given in one of the website examples where the cause isn't as clear as tuning parameters one by one.

http://ibayer.github.io/fastFM/guide.html#how-to-choose-the-right-solver

[ibayer]

The assertion is in the C code which makes error handling quite a bit difficult. I also don't like that the assert crashes python and am open to suggestions on how to do this better (best with PR 😄 ).
Thanks for pointing out the example, this definitely needs to be fixed
#34