_do_forward_pass, _do_backward_pass, _compute_posteriors not defined in skweak.aggregation

[ManuBohra]

skweak/aggregation.py", line 405, in fit
logprob, fwdlattice = self._do_forward_pass(framelogprob)
AttributeError: 'HMM' object has no attribute '_do_forward_pass'

[mnschmit]

I installed via pip and got the same error when trying to fit the HMM.

[david-waterworth]

_do_forward_pass comes from the hmmlearn dependency. It's working for me, I have hmmlearn==0.2.6 and scikit-learn==1.0.2. I wonder if you have an older version installed?

[ManuBohra]

I have hmmlearn==0.2.7 and meet with all requirements that is stated in requirements.txt

*python 3.6.9

[mnschmit]

@david-waterworth thanks a lot for the hint! It seems that hmmlearn renamed the "missing" functions to _do_forward_log_pass, _do_backward_log_pass, and _compute_posteriors_log when they introduced the faster alternative called scaling.
Probably, skweak could simply rename the calls to these three functions and it would work with the newer version of hmmlearn but to avoid any other issue, I simply downgraded by explicitly stating hmmlearn==0.2.6 (I also had .7 before just like @ManuBohra).

Now it works for me. I only get a lot of RuntimeWarning: invalid value encountered in subtract during model inference. I can ignore them with numpy.errstate but should I worry about them? Do you get these, too, @david-waterworth ?

[plison]

Yes, the error seems to be due to a change in hmmlearn. We are soon going to release a new version of skweak which fixes this issue along with other improvements, stay tuned :-) In the meantime, yes, as suggested by @mnschmit , you can either downgrade to hmmlearn==0.2.6, or edit the aggregation.py module to use the new function names from hmmlearn.
Regarding the warning message ('invalid value encountered in substract'), I don't really know.

[mnschmit]

Thanks for the confirmation @plison I'll watch the repository to avoid missing the new release!

The warning message occurs here. hmmlearn seems to return invalid values here. Maybe this is also due to the older version I am using now. I hope it'll be fixed with the new release, too!

[ljvmiranda921]

Agree, I was able to fix this by pinning the following dependencies in requirements.txt

skweak==0.2.13
hmmlearn==0.2.6

[martarho]

It seems that additionally to the functions above, _compute_log_xi_sum is also not supported anymore, or at least can't be found.

Do you know when will you release the new version? And if not soon, what would be a suitable fix for this?

[plison]

I hope to be able to release the new version by next week. But a quick fix would be to downgrade hmmlearn to a previous version, which does include _compute_log_xi_sum, _do_forward_pass and those other functions.

[plison]

Should be fixed in the new version that was just released!