g8a9 / l3wrapper Goto Github PK

Currently, rules are generated with a human-readable format if the relative parameter is set to True. However, items do not have always the same order in different rules. This makes the model harder to read.
It might be useful to assign an initial order to items (lexicographical?) and stick with it when generating the rules files.

Given an instance of L3Classifier, the parallelization of the fitting (e.g. via GridSearchCV or joblib.Parallel is not working due to the inner functioning of L3. Its binaries create intermediate result files during for example DBCoverage and more. These are not labeled with a stem or token, hence they come in contrast if multiple runs of the same executable are active in the same working directory.

The problem might be overcome by creating a new folder on the fly (yet, I know, another IO operation), moving into it and perform training.

Currently, the classifier allows saving rules in a human-readable format. This is done for visualization purposes solely.
It might be useful to exploit these two resulting files or even another (possibly compressed) format of them to dump the fitted model and store it for later uses.
The enhancement can be seen as a way to create a checkpoint of a fitted model and to allow users to reconstruct the model for further inferences.