Hyperparameter optimization

17 views

Skip to first unread message

Uldis Bojars

unread,

Apr 4, 2025, 7:08:25 AMApr 4

to Annif Users

Hi,

What Annif backends support hyperparameter optimization and can benefit from it?

I understand that not all backends support it, for example, trying to run hyperparameter optimization on the Omikuji backend results in an error:

AttributeError: module 'annif.backend' has no attribute 'hyperopt'

For the backends that support optimization, is it sufficient to run 'hyperopt' without parameters (except for --trials) or are there any specific parameters that should be / are recommended to be set when doing hyperparameter optimization?

Best regards,

Uldis Bojārs

juho.i...@helsinki.fi

unread,

Apr 7, 2025, 3:13:58 AMApr 7

to Annif Users

Hi,

Currently the hyperparameter optimization support is implemented for MLLM and (simple) ensemble backends. The parameters that are optimized in the process are here:

- https://github.com/NatLibFi/Annif/blob/38f5e8ae6e4019c2b3a732dcc3c22411c1968287/annif/backend/mllm.py#L46-L51

- https://github.com/NatLibFi/Annif/blob/38f5e8ae6e4019c2b3a732dcc3c22411c1968287/annif/backend/ensemble.py#L126-L129 (just the base project weights)

Note to self: We should make Annif give a clearer error message for the backends without hyperopt support than AttributeError.

I think there is no need to set any special parameter when running hyperopt, apart from the --trials as you said.