Method run

persper/classifier.py:314–408

Perform experiment in a leave-one-out style Args: label_func: A function, takes a data point as input and return its target label. estimator: A function, return a classifier which supports 'fit' and 'predict' method

(self, label_func, estimator, ngram_range=(1, 1),
            text_feature='', use_text=True, use_frc=False,
            dp_filter=lambda dp: True, tokenizer=None, 
            max_features=None, min_df=1, use_bns=False, k=None,
            n_jobs=1, use_description=False, use_comment=False, 
            use_svm=False, svm_type='linear', 
            use_rf=False, num_estimators = 100, val_max_features=0.5, 
             is_jira=False)

        self.classifiers = {}

    def run(self, label_func, estimator, ngram_range=(1, 1),
            text_feature='', use_text=True, use_frc=False,
            dp_filter=lambda dp: True, tokenizer=None, 
            max_features=None, min_df=1, use_bns=False, k=None,
            n_jobs=1, use_description=False, use_comment=False, 
            use_svm=False, svm_type='linear', 
            use_rf=False, num_estimators = 100, val_max_features=0.5, 
             is_jira=False):
        """Perform experiment in a leave-one-out style
        
        Args:
            label_func: A function, takes a data point as input and
                return its target label.
            estimator: A function, return a classifier which supports
                'fit' and 'predict' method
            ngram_range: A tuple of two integers, specify what range of 
                ngram to use
            text_feature: A string, 
                can be either 'title' or 'description' or 'comment' for jira issue, 
                and can be either 'message' or 'subject' for fs patch.
                If set to None, then texts will not be used.
            use_text: A boolean flag, whether to use text feature
            use_frc: A boolean flag, whether to use frc
            dp_filter: A function decides which data point to exclude.
            tokenizer: A function takes a string and return a list of tokens.
            max_features: An int or None. If not None, only consider
                top max_features ordered by term frequency across the corpus.
            min_df: An int, ignore terms when building vocabulary if their
                document frequency is strictly lower than this threshold.
            use_bns: A boolean flag, use BNS if True, otherwise use IDF
            k: An int, number of top features to keep during feature selection. 
        """
        self._clean()
        
        if text_feature == '': 
            if is_jira:
                text_feature = 'title'
            else:
                text_feature = 'message'
        
        iss = FeatureLabelExtractor(self.datasets, text_feature, label_func, dp_filter, is_jira)
        
        for jr in self.file_list:
            ofs_list = [ojr for ojr in self.file_list if ojr != jr]

            if is_jira:
                train_X, train_y = iss.jira_issue_transform(ofs_list, use_description, use_comment)
                test_X, test_y = iss.jira_issue_transform([jr], use_description, use_comment)
            else:
                train_X, train_y = iss.fs_patch_transform(ofs_list)
                test_X, test_y = iss.fs_patch_transform([jr])

            cv = CountVectorizer(tokenizer=tokenizer,
                                 ngram_range=ngram_range,
                                 max_features=max_features,
                                 min_df=min_df)
            
            train_X['count'] = cv.fit_transform(train_X['text'])