BayesFusion Support Forum

To switch the inference algorithm to EPIS use set_bayesian_algorithm with pysmile.BayesianAlgorithmType.EPIS_SAMPLING:
net.set_bayesian_algorithm(pysmile.BayesianAlgorithmType.EPIS_SAMPLING)

Regarding explicit relevance targets in the network, here's the info from the wrappers manual
Network ...

The inference in fully continuous and hybrid networks works the same way. If there's no evidence for nodes with parents which do not have evidence, forward sampling is used (this includes the hybrid case).

If there's evidence in nodes which have parents without evidence, the network will be ...

Network.set_controlled_value works just like set_evidence - you specify the node and the outcome. Calling this function is the equivalent of using 'Control Value' in GeNIe - in GeNIe the Control Value submenu has the entries for all node outcomes.

Error -42 is caused by insufficient memory in exact inference algorithm. This algorithm creates temporary data structure (the join tree), which can become very large, depending on the connections in the network.

You can switch to approximate sampling inference (EPIS), or set explicit targets in the ...

BTW, if you want to move this conversation into private space, email me or send me a forum PM.

One more thing, we're now working on additional structure learning algorithms and there will be a way to calculate BIC for network+dataset directly.

Consider using Network.set_controlled_value. For more info on arc cutting/value control, please refer to GeNIe manual, chapter "Controlling values". In the PDF version of the manual, the chapter number is 6.2.8.

BTW, you can use GeNIe to calculate loglikelihood for a dataset/network without learning. This functionality is not exposed in PySMILE, however.

You can obtain the log likelihood after EM by calling the EM.get_last_score() method.

BayesianSearch.get_last_score() returns the Bayesian score after learning.

Currently there's no method to calculate BIC from the dataset and network without learning.

I confirm that SMILE does not support parameter learning in hybrid networks.

Correct, the .so file from the .tar.gz archive is for the scenario where you do not use pip.

I have sent you a download link using private forum message.

Later we call the EM algorithm for parameter learning on this constructed network `cate_net`, the initial numbers like 0.5, 0.5, 0, 0 etc does not matter right?

If you have the complete dataset (no missing entries), the choice of initial parameters does not matter, unless you set the equivalent ...

If you are attempting to learn parameters using a dataset without missing data, the selection of initial parameters (uniformize vs randomize vs keep original) does not matter - the node probabilities are calculated by counting records in the data. Just instantiate the EM object and call em.Learn ...

Which Linux distro runs on your cluster?