Memory leak in sklearn_pmml_model.ensemble.PMMLForestClassifier

[icicle4]

Description

When using sklearn_pmml_model.ensemble.PMMLForestClassifier load pmml weight file, when call predict_proba in loop, the memory usage will get a continuous growth.

Steps/Code to Reproduce

from sklearn_pmml_model.ensemble import PMMLForestClassifier
import numpy as np
import logging
from random import random
from tqdm import tqdm
import gc
import time

logging.basicConfig(level=logging.INFO, format='%(asctime)s %(levelname)-8s %(message)s',)
mlog = logging.getLogger('myLogger')
level = logging.getLevelName('INFO')
mlog.setLevel(level)

class MergeClassifier(object):
    def __init__(self, config):
        self._config = config

    @staticmethod
    def _clip_data(x, max_val=10000.0):
        for j in range(len(x)):
            x[j] = min(max_val, x[j])
            x[j] = max(-max_val, x[j])
        return x

    def __call__(self, feature, threshold=None):
        raise NotImplementedError

class RFMergeClassifier(MergeClassifier):
    def __init__(self, config):
        super(RFMergeClassifier, self).__init__(config)
        model_file = config["merge_model"]

        self._classifier = PMMLForestClassifier(pmml=model_file, n_jobs=1)
        self._threshold = 0.10799543425155324
        self._keep_pool_threshold = 0.06698819140047929
        self._feature_names = ['reid_dist_p10', 'reid_dist_p20', 'reid_dist_p25', 
                               'reid_dist_p30', 'reid_dist_p50', 'region_in_out_match', 
                               'has_shared_companion', 'tpid_overlap_duration', 
                               'min_tpid_duration', 'max_tpid_duration', 
                               'min_spatial_distance', 'max_spatial_distance', 
                               'spatial_match_count', 'confident_spatial_match_count', 
                               'spatial_match_candidate_count', 'static_area_match', 
                               'exact_door_match', 'other_door_match', 'end_door_distance', 
                               'start_door_distance', 'min_staff_prob', 'max_staff_prob']

        if "merge_model_threshold" in config:
            self._threshold = config["merge_model_threshold"]
        if "merge_model_keep_threshold" in config:
            self._keep_pool_threshold = config["merge_model_keep_threshold"]
    
    def build_feature(self):
        return {key: random() * 10 for key in self._feature_names}

    @property
    def feature_keys(self):
        return self._feature_names

    def _get_score(self, feature, t1, t2, threshold=None):
        x = self._clip_data([feature[key] for key in self._feature_names])
        score = self._classifier.predict_proba(np.asarray([x], dtype=np.float32))[0, 1]
        if "face_match_score" in feature:
            score += feature["face_match_score"]
        return score

    def __call__(self, feature, t1, t2, threshold=None):
        if threshold is None:
            threshold = self._threshold
        score = self._get_score(feature, t1, t2)

        return score > threshold, 1 - score, \
               score > self._keep_pool_threshold, "match prob {:0.3f}".format(score)


if __name__ == "__main__":
    config = {
        "merge_model": "./models/random_forest.pmml",
        "merge_model_threshold": 0.5, 
        "_keep_pool_threshold": 2
    }

    _classifer = RFMergeClassifier(config)

    feature = _classifer.build_feature()

    for i in tqdm(range(10000)):
        _classifer(feature, 100, 100, 0.5)

Actual Results

there is the memray report for above code

Versions

macOS-10.16-x86_64-i386-64bit
Python 3.9.7 (default, Sep 16 2021, 08:50:36)
[Clang 10.0.0 ]
NumPy 1.23.0
SciPy 1.8.1
Scikit-Learn 1.3.0
sklearn-pmml-model 1.0.2

[iamDecode]

Thanks for the detailed report! If its not too much effort and you still have this setup ready, could you make sure using scikit-learn's RandomForestClassifier does not exhibit this leak? I don't expect so, but its useful to exclude early on.

[iamDecode]

I managed to reproduce the error and trace it back to a typo in the originally forked cython code. I released a new version 1.0.3 with the fix. Let me know if this resolved your issue :)

[icicle4]

Thank you very much! The issue was solved!!! Below is the new memory usage curve.