如何使GridSeachCV与管道中的自定义转换器一起工作？

Question

如果我排除自定义转换器，则GridSearchCV可以正常运行，但是会出错。这是一个伪数据集：

import pandas
import numpy
from sklearn_pandas import DataFrameMapper
from sklearn_pandas import cross_val_score
from sklearn.pipeline import Pipeline
from sklearn.grid_search import GridSearchCV
from sklearn.base import TransformerMixin
from sklearn.preprocessing import LabelBinarizer
from sklearn.ensemble import RandomForestClassifier
import sklearn_pandas
from sklearn.preprocessing import MinMaxScaler

df = pandas.DataFrame({"Letter":["a","b","c","d","a","b","c","d","a","b","c","d","a","b","c","d"],
                       "Number":[1,2,3,4,1,2,3,4,1,2,3,4,1,2,3,4], 
                       "Label":["G","G","B","B","G","G","B","B","G","G","B","B","G","G","B","B"]})

class MyTransformer(TransformerMixin):

    def transform(self, x, **transform_args):
        x["Number"] = x["Number"].apply(lambda row: row*2)
        return x

    def fit(self, x, y=None, **fit_args):
        return self

x_train = df
y_train = x_train.pop("Label")    

mapper = DataFrameMapper([
    ("Number", MinMaxScaler()),
    ("Letter", LabelBinarizer()),
    ])

pipe = Pipeline([
    ("custom", MyTransformer()),
    ("mapper", mapper),
    ("classifier", RandomForestClassifier()),
    ])


param_grid = {"classifier__min_samples_split":[10,20], "classifier__n_estimators":[2,3,4]}

model_grid = sklearn_pandas.GridSearchCV(pipe, param_grid, verbose=2, scoring="accuracy")

model_grid.fit(x_train, y_train)

错误是

list indices must be integers, not str

当管道中有自定义转换器时，如何使GridSearchCV工作？

Answer 1

我知道这个答案来得相当晚，但我在 sklearn 和BaseSearchCV衍生类中遇到了同样的行为。这个问题实际上似乎源于_PartitionIteratorsklearn cross_validation 模块中的类，因为它假设TransformerMixin管道中每个类发出的所有内容都将类似于数组，因此它会生成用于索引传入的索引切片Xargs 以类似数组的方式。方法如下__iter__：

def __iter__(self):
    ind = np.arange(self.n)
    for test_index in self._iter_test_masks():
        train_index = np.logical_not(test_index)
        train_index = ind[train_index]
        test_index = ind[test_index]
        yield train_index, test_index 

网格BaseSearchCV搜索元类调用 cross_validation 的_fit_and_score，它使用名为 的方法safe_split。这是相关行：

X_subset = [X[idx] for idx in indices]

如果 X 是您从transform函数中发出的 pandas 数据框，这绝对会产生意外的结果。

我发现有两种方法可以解决这个问题：

1. 确保从变压器返回一个数组：

   return x.as_matrix()
   

2. 这是一个黑客行为。如果变压器管道要求下一个变压器的输入是 DataFrame，就像我的情况一样，您可以编写一个实用程序脚本，该脚本本质上与 sklearn 模块相同，但包含一些在以下方法grid_search中调用的巧妙验证方法_fit班上BaseSearchCV：

   def _validate_X(X):
       """Returns X if X isn't a pandas frame, otherwise 
       the underlying matrix in the frame. """
       return X if not isinstance(X, pd.DataFrame) else X.as_matrix()
   
   def _validate_y(y):
       """Returns y if y isn't a series, otherwise the array"""
       if y is None:
           return y
   
       # if it's a series
       elif isinstance(y, pd.Series):
           return np.array(y.tolist())
   
       # if it's a dataframe:
       elif isinstance(y, pd.DataFrame):
           # check it's X dims
           if y.shape[1] > 1:
               raise ValueError('matrix provided as y')
           return y[y.columns[0]].tolist()
   
       # bail and let the sklearn function handle validation
       return y
   

作为示例，这是我的“自定义 grid_search 模块”。