scikit-learnでグリッドサーチ

グリッドサーチ

グリッドサーチは、パラメータをチューニングしてモデルの汎化性能を向上させる方法。

sklearnでグリッドサーチ

パラメータの辞書を作ります。
GridSearchCVに、

• モデル
• パラメータ辞書
• 交差検証の分割数

渡します。

後の、学習と推論（予測）の作業はいつも通りの、fitとpredictで実行します。

from sklearn.datasets import load_iris
from sklearn.model_selection import train_test_split
from sklearn.svm import SVC
from sklearn.model_selection import GridSearchCV
from sklearn.metrics import accuracy_score


iris = load_iris()
train_X, val_X, train_y, val_y = train_test_split(iris.data, iris.target, test_size=0.2, random_state=0)


params_grid = {'C': [0.001, 0.01, 0.1, 1, 10, 100],
                'gamma':[0.001, 0.01, 0.1, 1, 10, 100]}

grid_search = GridSearchCV(SVC(), params_grid, cv=5)

grid_search = grid_search.fit(train_X, train_y)

params = grid_search.best_params_
print("best_paramerters is ",params)
print("best_score is ", grid_search.best_score_)

pred = grid_search.predict(val_X)

print("accuracy is ", accuracy_score(pred, val_y))

上のgrid_searchのparamsに最適なパラメータが格納されているので、シンプルにSVCを再実装して、予測を行うこともできます。

svc = SVC(**params)
svc.fit(train_X, train_y)
pred = svc.predict(val_X)
print(accuracy_score(pred, val_y))

グリッドサーチ×1つ抜き交差検証法

言うまでもなく、1つ抜き交差検証法はデータ数が大きくなると膨大な時間がかかる。

from sklearn.datasets import load_iris
from sklearn.model_selection import train_test_split
from sklearn.svm import SVC
from sklearn.model_selection import GridSearchCV
from sklearn.metrics import accuracy_score

from sklearn.model_selection import KFold, LeaveOneOut

iris = load_iris()
train_X, val_X, train_y, val_y = train_test_split(iris.data, iris.target, test_size=0.2, random_state=0)


params_grid = {'C': [0.001, 0.01, 0.1, 1, 10, 100],
                'gamma':[0.001, 0.01, 0.1, 1, 10, 100]}

loo = LeaveOneOut()
grid_search = GridSearchCV(SVC(), params_grid, cv=loo)

grid_search = grid_search.fit(train_X, train_y)

params = grid_search.best_params_
print("best_paramerters is ",params)
print("best_score is ", grid_search.best_score_)

pred = grid_search.predict(val_X)

print("accuracy is ", accuracy_score(pred, val_y))

参考

https://scikit-learn.org/stable/modules/generated/sklearn.model_selection.GridSearchCV.html