Also renamed parameter names for better consistency and added estimator_enabled parameter which disables the weak labelling pathway.
Changes to get_infer_data
When infer_labeled = True, fixing_domain_gen performed inference on ALL cells (clean and DK). This meant that un-detected errors (clean cells that had errors) were also being repaired. This overestimated the # of correct/total repair and thus it is not recommend to set infer_labeled = True when performing experiments.
When infer_labeled = False, fixing_domain_gen only inferred on non-weak labelled cells. If during weak labelling a DK cell was weak labelled, it would not be inferred. This is fixed in https://github.com/HoloClean/holoclean/pull/56 so all DK cells are inferred regardless of their labelling.
Also renamed parameter names for better consistency and added
estimator_enabledparameter which disables the weak labelling pathway.Changes to
get_infer_dataWhen infer_labeled = True, fixing_domain_gen performed inference on ALL cells (clean and DK). This meant that un-detected errors (clean cells that had errors) were also being repaired. This overestimated the # of correct/total repair and thus it is not recommend to set infer_labeled = True when performing experiments.
When infer_labeled = False, fixing_domain_gen only inferred on non-weak labelled cells. If during weak labelling a DK cell was weak labelled, it would not be inferred. This is fixed in https://github.com/HoloClean/holoclean/pull/56 so all DK cells are inferred regardless of their labelling.