FM: model coefficients Nan after fit

[RiccardoNizzolo]

Hi!
I'm running this kernel on kaggle:
kernel

After the fit on a valid matrix (without NA) the status of the model still contains Nan.
During the fit there are neither exceptions nor warnings.

To see the status of de model I have called :
model.__getstate__()

The result is:

model FM_FTRL status:
(0.01, 0.01, 1e-05, 0.1, 0.01, 0.0, 1.0, 0.0, 1020688, 200, 17, array([ nan, 0., 0., ..., 1., 1., 1.]), array([ nan, nan, nan, ..., nan, nan, nan]), array([ nan, nan, nan, ..., nan, nan, nan]), array([ nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan,
nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan,
nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan,
nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan,
nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan,
nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan,
nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan,
nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan,
nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan,
nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan,
nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan,
nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan,
nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan,
nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan,
nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan,
nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan,
nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan,
nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan,
nan, nan]), array([ nan, nan, nan, ..., nan, nan, nan]), array([ nan, nan, nan, ..., nan, nan, nan]), 0, 0, True)

[anttttti]

I think this is an issue with your parameters.

" 1.0, 0.0, 1020688".. the 0.0 parameter in the middle means you've initialized the factor weights as 0. A init_fm=value of 0.01-0.001 usually works well.

[RiccardoNizzolo]

The init_fm parameter is initialized with value 0.2:
model = FM_FTRL(alpha=0.01, beta=0.01, L1=0.00001, L2=0.1, D=sparse_merge.shape[1], alpha_fm=0.01, L2_fm=0.1, init_fm=0.2, D_fm=200, e_noise=0.0001, iters=17, inv_link="identity", threads=4)

If i call model.__getstate__() just after the initialization I get init_fm=0 but the z array is randomly initialized:
log

The interesting fact is that the Nan came out dependently on the training set:
model.fit(train) -->converge
model.fit(train.sign()) --> get Nan values

[SpongebBob]

I have the same problem

If i call model.getstate() just after the initialization I get init_fm=0

All predict values are nan

[anttttti]

I could replicate this now. The problem is not with init_fm, that one you could set as 0, but with bad results. The issue is alpha_fm set too high compared to feature values, leading the updates to overflow. Try normalizing the input rows, or starting with very low alpha_fm, <0.01.

This could be fixed by gradient clipping, but in general you would still get low quality models in cases where this is happening. You should normalize the input rows or lower alpha_fm.