Memory leak in BinaryBalancerAlgorithm #312
Description
Use of the BinaryBalancerAlgorithm in e.g. the LPDebiaserBinary means that at one point, the following gets executed.
self.__roc_stats = [ tools.loss_from_roc(y_true[ids], probs[ids], self.rocs[i]) for i, ids in enumerate(group_ids) ]
If you save the returned model to MLFlow (using the tutorial provided at https://www.holisticai.com/blog/holistic-ai-library-and-mlflow-anti-bias-toolkit-for-data-science), the final model blow up in size (data set of 779649 rows*48 columns amounts to 94077045492 bytes).
I have set the final attribute mitigator.algorithm._BinaryBalancerAlgorithm__roc_stats = None , as this attribute isn't used after the fit method. Yet this feels like it should be adressed.