机器学习入门实例-加州房价预测-3（选择与训练模型+调参）

news2024/11/15 19:49:22

选择与训练模型

使用线性回归

	from sklearn.linear_model import LinearRegression
    from sklearn.metrics import mean_squared_error

    lin_reg = LinearRegression()
    lin_reg.fit(housing_prepared, housing_labels)
    housing_predictions = lin_reg.predict(housing_prepared)
    lin_mse = mean_squared_error(housing_labels, housing_predictions)
    lin_rmse = np.sqrt(lin_mse)
    # 太高的表示Underfit
    print(lin_rmse) # 68627.87390018745

在这里插入图片描述

使用决策树

    from sklearn.metrics import mean_squared_error
    from sklearn.tree import DecisionTreeRegressor

    tree_reg = DecisionTreeRegressor()
    tree_reg.fit(housing_prepared, housing_labels)
    housing_predictions = tree_reg.predict(housing_prepared)
    tree_mse = mean_squared_error(housing_labels, housing_predictions)
    tree_rmse = np.sqrt(tree_mse)
    # 太低的表示overfit
    print(tree_rmse) # 0.0

k-fold cross-validation

fold指不同的数据子集。k-fold cross-validation就是随机产生k个fold，每次选一个fold来评估效果，其他k-1个fold用来训练。这样会得到k个评估分数。
由于scikit-learn的cross-validation用的是utility function（越大越好）而非cost function（越小越好），所以这里选用了负数的mse。

def display_scores(scores):
    print("Scores: ", scores)
    print("Mean: ", scores.mean())
    print("Standard deviation: ", scores.std())

...
	from sklearn.tree import DecisionTreeRegressor
	from sklearn.linear_model import LinearRegression
	from sklearn.model_selection import cross_val_score
	
    tree_reg = DecisionTreeRegressor()
    scores = cross_val_score(tree_reg, housing_prepared, housing_labels,
                             scoring="neg_mean_squared_error", cv=10)
    tree_rmse_scores = np.sqrt(-scores)
    print(display_scores(tree_rmse_scores))

    scores2 = cross_val_score(LinearRegression(), housing_prepared, housing_labels,
                              scoring="neg_mean_squared_error", cv=10)
    lin_rmse_scores = np.sqrt(-scores2)
    print(display_scores(lin_rmse_scores))

    from sklearn.ensemble import RandomForestRegressor
    forest_reg = RandomForestRegressor()
    scores3 = cross_val_score(forest_reg, housing_prepared, housing_labels,
                              scoring="neg_mean_squared_error", cv=10)
    forest_rmse_scores = np.sqrt(-scores3)
    print(display_scores(forest_rmse_scores))

Scores:  [72747.0548719  69975.38915452 68277.88913022 71825.88436612
 69330.68736727 76568.19953125 72764.63740705 72219.04604848
 69031.97373315 70374.45146712]
Mean:  71311.52130770871
Standard deviation:  2321.1947983407713

Scores:  [71762.76364394 64114.99166359 67771.17124356 68635.19072082
 66846.14089488 72528.03725385 73997.08050233 68802.33629334
 66443.28836884 70139.79923956]
Mean:  69104.07998247063
Standard deviation:  2880.3282098180657

Scores:  [51627.53671267 49369.27783646 46517.4011464  51846.1271283
 47431.17962375 51357.65695338 52565.55993554 49896.31387898
 48580.76091798 53639.76973388]
Mean:  50283.15838673507
Standard deviation:  2192.6847023341056

可以看到，随机森林的效果好一点点。

保存模型可以用joblib包：

    import joblib
    joblib.dump(forest_reg, "forest_regression.pkl")

使用时：

import joblib
my_model_loaded = joblib.load("forest_regression.pkl")
y_pred = my_model_loaded.predict(X_test)

Fine-tune（微调）

Grid Search

    from sklearn.ensemble import RandomForestRegressor
    from sklearn.model_selection import GridSearchCV
    forest_reg = RandomForestRegressor()
    param_grid = [
        {'n_estimators': [3, 10, 30, 50], 'max_features': [2, 4, 6, 8, None]},
        {'bootstrap': [False], 'n_estimators': [3, 10, 30], 'max_features': [2, 3, 4, 8]}
    ]
    grid_search = GridSearchCV(forest_reg, param_grid, cv=5,
                               scoring="neg_mean_squared_error",
                               return_train_score=True)
    grid_search.fit(housing_prepared, housing_labels)
    print(grid_search.best_params_)
    print(grid_search.best_estimator_)
    print(np.sqrt(-grid_search.best_score_))
    # 输出训练结果
    cvres = grid_search.cv_results_
    for mean_score, params in zip(cvres["mean_test_score"], cvres["params"]):
        print(np.sqrt(-mean_score), params)


{'bootstrap': False, 'max_features': 4, 'n_estimators': 30}
RandomForestRegressor(bootstrap=False, max_features=4, n_estimators=30)
49184.012820612865
63296.60697816219 {'max_features': 2, 'n_estimators': 3}
55506.98954834032 {'max_features': 2, 'n_estimators': 10}
52588.29130799578 {'max_features': 2, 'n_estimators': 30}
52209.351970524105 {'max_features': 2, 'n_estimators': 50}
61124.66160352609 {'max_features': 4, 'n_estimators': 3}
52392.16874785449 {'max_features': 4, 'n_estimators': 10}
50273.44009993824 {'max_features': 4, 'n_estimators': 30}
50073.39902683726 {'max_features': 4, 'n_estimators': 50}
58727.56972424551 {'max_features': 6, 'n_estimators': 3}
51782.31672170441 {'max_features': 6, 'n_estimators': 10}
49980.86986567165 {'max_features': 6, 'n_estimators': 30}
49621.90325596429 {'max_features': 6, 'n_estimators': 50}
59108.772714035185 {'max_features': 8, 'n_estimators': 3}
52103.26690162754 {'max_features': 8, 'n_estimators': 10}
50132.82702279907 {'max_features': 8, 'n_estimators': 30}
49781.109586154096 {'max_features': 8, 'n_estimators': 50}
59531.39144640905 {'max_features': None, 'n_estimators': 3}
53194.75190489223 {'max_features': None, 'n_estimators': 10}
51329.73377108047 {'max_features': None, 'n_estimators': 30}
50826.59885373242 {'max_features': None, 'n_estimators': 50}
61849.029391012125 {'bootstrap': False, 'max_features': 2, 'n_estimators': 3}
54126.80686429009 {'bootstrap': False, 'max_features': 2, 'n_estimators': 10}
51848.68218742629 {'bootstrap': False, 'max_features': 2, 'n_estimators': 30}
60098.910109841345 {'bootstrap': False, 'max_features': 3, 'n_estimators': 3}
52334.13150446245 {'bootstrap': False, 'max_features': 3, 'n_estimators': 10}
50341.87869180483 {'bootstrap': False, 'max_features': 3, 'n_estimators': 30}
57650.86470796738 {'bootstrap': False, 'max_features': 4, 'n_estimators': 3}
51682.89592908345 {'bootstrap': False, 'max_features': 4, 'n_estimators': 10}
49184.012820612865 {'bootstrap': False, 'max_features': 4, 'n_estimators': 30}
57308.84929909555 {'bootstrap': False, 'max_features': 8, 'n_estimators': 3}
51285.68101404841 {'bootstrap': False, 'max_features': 8, 'n_estimators': 10}
49469.879650849194 {'bootstrap': False, 'max_features': 8, 'n_estimators': 30}

解释：param_grid表示要测试两组参数，第一组是n_estimators和max_features的组合，所以共有4 x 5 = 20种；第二组是 3 x 4 = 12种，所以一共测试了20 + 12 = 32种参数设置方式。每一种训练5次。训练结束后可以通过best_params_获取参数组合。

n_estimators：随机森林中树的数量，默认为 100
max_features：选择特征的方式，可以是整数、浮点数、字符串或 None。默认为 “auto”，即使用所有特征。如果是整数，则表示每次分裂只考虑部分特征；如果是浮点数，则表示每次分裂只考虑部分特征的比例；如果是字符串 “sqrt” 或 “log2”，则表示每次分裂只考虑特征数的平方根或对数；如果是 None，则表示每次分裂都使用所有特征
bootstrap 参数是指是否使用自助法（bootstrap）从原始样本中进行有放回的抽样来生成新的训练集。默认值为 True，即使用自助法。如果将其设为 False，则表示不使用自助法，即不进行有放回的抽样。

gs也可以用自定义的参数。

from sklearn.pipeline import Pipeline
from sklearn.ensemble import RandomForestRegressor
from sklearn.model_selection import GridSearchCV

# 定义自定义的 Transformer
class MyTransformer(BaseEstimator, TransformerMixin):
    def __init__(self, param1=1, param2=1):
        self.param1 = param1
        self.param2 = param2

    def fit(self, X, y=None):
        # ...
        return self

    def transform(self, X, y=None):
        # ...
        return X_transformed

# 定义 Pipeline 和 param_grid
pipeline = Pipeline([
    ('my_transformer', MyTransformer()),
    ('rf', RandomForestRegressor())
])

param_grid = {
    'my_transformer__param1': [1, 2, 3],
    'my_transformer__param2': [0.1, 0.2, 0.3],
    'rf__n_estimators': [10, 50, 100],
    'rf__max_depth': [None, 5, 10],
}
# 进行网格搜索
grid_search = GridSearchCV(pipeline, param_grid=param_grid, cv=5, n_jobs=-1)
grid_search.fit(X, y)