python - GridSearchCV 评分和 grid_scores_

Question

我正在尝试了解如何获取 GridSearchCV 的记分员值.下面的示例代码在文本数据上设置了一个小型管道。

然后它在不同的 ngram 上设置网格搜索。

评分是通过 f1 度量完成的:

#setup the pipeline
tfidf_vec = TfidfVectorizer(analyzer='word', min_df=0.05, max_df=0.95)
linearsvc = LinearSVC()
clf = Pipeline([('tfidf_vec', tfidf_vec), ('linearsvc', linearsvc)])

# setup the grid search
parameters = {'tfidf_vec__ngram_range': [(1, 1), (1, 2)]}
gs_clf = GridSearchCV(clf, parameters, n_jobs=-1, scoring='f1')
gs_clf = gs_clf.fit(docs_train, y_train)

现在我可以打印分数:

打印 gs_clf.grid_scores_

[mean: 0.81548, std: 0.01324, params: {'tfidf_vec__ngram_range': (1, 1)},
 mean: 0.82143, std: 0.00538, params: {'tfidf_vec__ngram_range': (1, 2)}]

打印 gs_clf.grid_scores_[0].cv_validation_scores

array([ 0.83234714,  0.8       ,  0.81409002])

我不清楚 documentation :

1. gs_clf.grid_scores_[0].cv_validation_scores 是一个数组，其分数通过评分参数定义，每折(在这种情况下，f1 测量每折)？如果不是，那又是什么？

2. 如果我改为选择另一个 metric ，例如 scoring='f1_micro'，gs_clf.grid_scores_[i].cv_validation_scores 中的每个数组都将包含用于特定网格搜索参数选择的折叠的 f1_micro 指标？

Answer 1

我编写了以下函数将 grid_scores_ 对象转换为 pandas.DataFrame。希望数据框 View 有助于消除您的困惑，因为它是一种更直观的格式:

def grid_scores_to_df(grid_scores):
    """
    Convert a sklearn.grid_search.GridSearchCV.grid_scores_ attribute to a tidy
    pandas DataFrame where each row is a hyperparameter-fold combinatination.
    """
    rows = list()
    for grid_score in grid_scores:
        for fold, score in enumerate(grid_score.cv_validation_scores):
            row = grid_score.parameters.copy()
            row['fold'] = fold
            row['score'] = score
            rows.append(row)
    df = pd.DataFrame(rows)
    return df

你必须有以下导入才能工作:import pandas as pd。