The following parameters are all set to default values and can be modified selectively.
top_m
After the calculation of evidence support is completed, the latent variables with the largest support of the first m evidence are selected as candidates, and their approximate probability and approximate entropy are calculated. The default setting is 2000.
top_k
After the approximate entropy of the m hidden variables is calculated, the first k hidden variables with the smallest approximate entropy are selected to construct k subgraphs. The default setting is 10.
top_n
After the k subgraphs are inferred, the entropy is calculated according to the actual inference probability, and the first n variables with the smallest entropy are selected for labeling. The default setting is 1.
update_proportion
In order to speed up the reasoning, the evidence support is recalculated only after the hidden variables of the update_proportion ratio are marked. The default setting is 0.01.
optimization_threshold
In order to speed up the reasoning, latent variables whose approximate entropy is less than or equal to optimization_threshold are directly marked and no longer reasoned. The default setting is 0, and a negative value means that this optimization is not required.
balance
When creating a subgraph, the 0-1 variables are balanced. The default setting is False.
learning_epoches
The number of parameter learning rounds, the default setting is 500.
inference_epoches
The number of inference rounds for factor graphs, the default setting is 500.
learning_method
The parameter learning method currently supports both stochastic gradient descent (sgd) and batch gradient descent (bgd). The default setting is sgd.
n_process
The number of multi-process acceleration processes, the default setting is 1.
out
Whether it is necessary to output the probability and label of hidden variable inference to the file in real time, the default is False.