search

kevinsbello / dagma

A Python 3 package for learning Bayesian Networks (DAGs) from data. Official implementation of the paper "DAGMA: Learning DAGs via M-matrices and a Log-Determinant Acyclicity Characterization"
https://dagma.readthedocs.io/en/latest/
Apache License 2.0
102 stars 19 forks source link

Hyper parameters for large scale experiments #1

Closed zzhang1987 closed 1 year ago

zzhang1987 commented 1 year ago

Hi there, would you please provide the hyper parameters for large scale experiments? I tries to use the hyper parameter for small scale exps, but it did not work.

Best,

Zhen Zhang

kevinsbello commented 1 year ago

Hi, I believe I only changed the L1 coefficient to 0.05, it could also help to increase the final max number of iterations a bit, e.g., form 6e4 to 7e4. Let me know if it works. -K

zzhang1987 commented 1 year ago

L1 coefficient might be the reason. I believe smaller L1 coefficient may result in better performance.

zzhang1987 commented 1 year ago

Also in the paper you have mentioned that you are using an 8-core E5-2860v4. I believe E5-2860v4 has more than 8 core. Do you mean 8-way (i.e. a very high-end machine with 8 E5-2860v4 cpu) or it is just a typo?

kevinsbello commented 1 year ago

Right, E5-2860v4 has more than 8 cores, however, I was restricted to use only 8-cores as I was using a shared resource :)

zzhang1987 commented 1 year ago

Thanks for your helpful response. Now I believe the performance should be aligned with the ones reported in the paper.