阅读源码系列：sklearn中train_test_split使用及源码学习

引言

• 最近用到了sklearn中的train_test_split,感觉很好用，想来实现也不是特别复杂，于是想着看一下源码，学习一下大佬的写法

使用方法

>>> import numpy as np
>>> from sklearn.model_selection import train_test_split
>>> X, y = np.arange(10).reshape((5, 2)), range(5)
>>> X
array([[0, 1],
       [2, 3],
       [4, 5],
       [6, 7],
       [8, 9]])
>>> list(y)
[0, 1, 2, 3, 4]
>>> X_train, X_test, y_train, y_test = train_test_split(
...     X, y, test_size=0.33, random_state=42)
...
>>> X_train
array([[4, 5],
       [0, 1],
       [6, 7]])
>>> y_train
[2, 0, 3]
>>> X_test
array([[2, 3],
       [8, 9]])
>>> y_test
[1, 4]
>>> train_test_split(y, shuffle=False)
[[0, 1, 2], [3, 4]]

学习到东西

• 作者在进行每一个操作前，都确保所得到的数据是想要的，尤其是类型判断，所以说写一个合格的接口，绝非易事
• 合理的异常捕捉和错误提示
• 一个函数尽量只做一件事，以及函数的命名都很讲究，清楚明了
• 迭代器的相关用法，比如chain相关模块需要进一步了解

源码剥离

• 从sklearn中剥离出train_test_split的相关代码
• 主要如下（可以简单运行）：

from itertools import chain, compress

import numpy as np
import scipy.sparse as sp
from math import ceil, floor
import numbers


def check_random_state(seed):
    """Turn seed into a np.random.RandomState instance

    Parameters
    ----------
    seed : None, int or instance of RandomState
        If seed is None, return the RandomState singleton used by np.random.
        If seed is an int, return a new RandomState instance seeded with seed.
        If seed is already a RandomState instance, return it.
        Otherwise raise ValueError.
    """
    if seed is None or seed is np.random:
        return np.random.mtrand._rand
    if isinstance(seed, numbers.Integral):
        return np.random.RandomState(seed)
    if isinstance(seed, np.random.RandomState):
        return seed
    raise ValueError('%r cannot be used to seed a numpy.random.RandomState'
                     ' instance' % seed)


class BaseShuffleSplit():
    """Base class for ShuffleSplit and StratifiedShuffleSplit"""

    def __init__(self, n_splits=10, *, test_size=None, train_size=None,
                 random_state=None):
        self.n_splits = n_splits
        self.test_size = test_size
        self.train_size = train_size
        self.random_state = random_state
        self._default_test_size = 0.1

    def split(self, X, y=None, groups=None):
        """Generate indices to split data into training and test set.

        Parameters
        ----------
        X : array-like of shape (n_samples, n_features)
            Training data, where n_samples is the number of samples
            and n_features is the number of features.

        y : array-like of shape (n_samples,)
            The target variable for supervised learning problems.

        groups : array-like of shape (n_samples,), default=None
            Group labels for the samples used while splitting the dataset into
            train/test set.

        Yields
        ------
        train : ndarray
            The training set indices for that split.

        test : ndarray
            The testing set indices for that split.

        Notes
        -----
        Randomized CV splitters may return different results for each call of
        split. You can make the results identical by setting `random_state`
        to an integer.
        """
        X, y, groups = indexable(X, y, groups)
        for train, test in self._iter_indices(X, y, groups):
            yield train, test

    def _iter_indices(self, X, y=None, groups=None):
        """Generate (train, test) indices"""

    def get_n_splits(self, X=None, y=None, groups=None):
        """Returns the number of splitting iterations in the cross-validator

        Parameters
        ----------
        X : object
            Always ignored, exists for compatibility.

        y : object
            Always ignored, exists for compatibility.

        groups : object
            Always ignored, exists for compatibility.

        Returns
        -------
        n_splits : int
            Returns the number of splitting iterations in the cross-validator.
        """
        return self.n_splits

    def __repr__(self):
        return _build_repr(self)


class ShuffleSplit(BaseShuffleSplit):
    """Random permutation cross-validator

    Yields indices to split data into training and test sets.

    Note: contrary to other cross-validation strategies, random splits
    do not guarantee that all folds will be different, although this is
    still very likely for sizeable datasets.

    Read more in the :ref:`User Guide <cross_validation>`.

    Parameters
    ----------
    n_splits : int, default=10
        Number of re-shuffling & splitting iterations.

    test_size : float or int, default=None
        If float, should be between 0.0 and 1.0 and represent the proportion
        of the dataset to include in the test split. If int, represents the
        absolute number of test samples. If None, the value is set to the
        complement of the train size. If ``train_size`` is also None, it will
        be set to 0.1.

    train_size : float or int, default=None
        If float, should be between 0.0 and 1.0 and represent the
        proportion of the dataset to include in the train split. If
        int, represents the absolute number of train samples. If None,
        the value is automatically set to the complement of the test size.

    random_state : int, RandomState instance or None, default=None
        Controls the randomness of the training and testing indices produced.
        Pass an int for reproducible output across multiple function calls.
        See :term:`Glossary <random_state>`.

    Examples
    --------
    # >>> import numpy as np
    # >>> from sklearn.model_selection import ShuffleSplit
    # >>> X = np.array([[1, 2], [3, 4], [5, 6], [7, 8], [3, 4], [5, 6]])
    # >>> y = np.array([1, 2, 1, 2, 1, 2])
    # >>> rs = ShuffleSplit(n_splits=5, test_size=.25, random_state=0)
    # >>> rs.get_n_splits(X)
    # 5
    # >>> print(rs)
    # ShuffleSplit(n_splits=5, random_state=0, test_size=0.25, train_size=None)
    # >>> for train_index, test_index in rs.split(X):
    # ...     print("TRAIN:", train_index, "TEST:", test_index)
    # TRAIN: [1 3 0 4] TEST: [5 2]
    # TRAIN: [4 0 2 5] TEST: [1 3]
    # TRAIN: [1 2 4 0] TEST: [3 5]
    # TRAIN: [3 4 1 0] TEST: [5 2]
    # TRAIN: [3 5 1 0] TEST: [2 4]
    # >>> rs = ShuffleSplit(n_splits=5, train_size=0.5, test_size=.25,
    # ...                   random_state=0)
    # >>> for train_index, test_index in rs.split(X):
    # ...     print("TRAIN:", train_index, "TEST:", test_index)
    TRAIN: [1 3 0] TEST: [5 2]
    TRAIN: [4 0 2] TEST: [1 3]
    TRAIN: [1 2 4] TEST: [3 5]
    TRAIN: [3 4 1] TEST: [5 2]
    TRAIN: [3 5 1] TEST: [2 4]
    """

    def __init__(self, n_splits=10, *, test_size=None, train_size=None,
                 random_state=None):
        super().__init__(
            n_splits=n_splits,
            test_size=test_size,
            train_size=train_size,
            random_state=random_state)
        self._default_test_size = 0.1

    def _iter_indices(self, X, y=None, groups=None):
        n_samples = _num_samples(X)
        n_train, n_test = _validate_shuffle_split(
            n_samples, self.test_size, self.train_size,
            default_test_size=self._default_test_size)

        rng = check_random_state(self.random_state)
        for i in range(self.n_splits):
            # random partition
            permutation = rng.permutation(n_samples)
            ind_test = permutation[:n_test]
            ind_train = permutation[n_test:(n_test + n_train)]
            yield ind_train, ind_test


def _num_samples(x):
    """Return number of samples in array-like x."""
    message = 'Expected sequence or array-like, got %s' % type(x)
    if hasattr(x, 'fit') and callable(x.fit):
        # Don't get num_samples from an ensembles length!
        raise TypeError(message)

    if not hasattr(x, '__len__') and not hasattr(x, 'shape'):
        if hasattr(x, '__array__'):
            x = np.asarray(x)
        else:
            raise TypeError(message)

    if hasattr(x, 'shape') and x.shape is not None:
        if len(x.shape) == 0:
            raise TypeError("Singleton array %r cannot be considered"
                            " a valid collection." % x)
        # Check that shape is returning an integer or default to len
        # Dask dataframes may not return numeric shape[0] value
        if isinstance(x.shape[0], numbers.Integral):
            return x.shape[0]

    try:
        return len(x)
    except TypeError as type_error:
        raise TypeError(message) from type_error


def check_consistent_length(*arrays):
    """Check that all arrays have consistent first dimensions.

    Checks whether all objects in arrays have the same shape or length.

    Parameters
    ----------
    *arrays : list or tuple of input objects.
        Objects that will be checked for consistent length.
    """
    lengths = [_num_samples(X) for X in arrays if X is not None]
    uniques = np.unique(lengths)
    if len(uniques) > 1:
        raise ValueError("Found input variables with inconsistent numbers of"
                         " samples: %r" % [int(l) for l in lengths])


def _make_indexable(iterable):
    """Ensure iterable supports indexing or convert to an indexable variant.

    Convert sparse matrices to csr and other non-indexable iterable to arrays.
    Let `None` and indexable objects (e.g. pandas dataframes) pass unchanged.

    Parameters
    ----------
    iterable : {list, dataframe, ndarray, sparse matrix} or None
        Object to be converted to an indexable iterable.
    """
    if sp.issparse(iterable):
        return iterable.tocsr()
    elif hasattr(iterable, "__getitem__") or hasattr(iterable, "iloc"):
        return iterable
    elif iterable is None:
        return iterable
    return np.array(iterable)


def indexable(*iterables):
    """Make arrays indexable for cross-validation.

    Checks consistent length, passes through None, and ensures that everything
    can be indexed by converting sparse matrices to csr and converting
    non-interable objects to arrays.

    Parameters
    ----------
    *iterables : {lists, dataframes, ndarrays, sparse matrices}
        List of objects to ensure sliceability.
    """
    result = [_make_indexable(X) for X in iterables]
    check_consistent_length(*result)
    return result


def _validate_shuffle_split(n_samples, test_size, train_size,
                            default_test_size=None):
    """
    Validation helper to check if the test/test sizes are meaningful wrt to the
    size of the data (n_samples)
    """
    if test_size is None and train_size is None:
        test_size = default_test_size

    # 获得数据类型
    test_size_type = np.asarray(test_size).dtype.kind
    train_size_type = np.asarray(train_size).dtype.kind

    if (test_size_type == 'i' and (test_size >= n_samples or test_size <= 0)
            or test_size_type == 'f' and (test_size <= 0 or test_size >= 1)):
        raise ValueError('test_size={0} should be either positive and smaller'
                         ' than the number of samples {1} or a float in the '
                         '(0, 1) range'.format(test_size, n_samples))

    if (train_size_type == 'i' and (train_size >= n_samples or train_size <= 0)
            or train_size_type == 'f' and (train_size <= 0 or train_size >= 1)):
        raise ValueError('train_size={0} should be either positive and smaller'
                         ' than the number of samples {1} or a float in the '
                         '(0, 1) range'.format(train_size, n_samples))

    if train_size is not None and train_size_type not in ('i', 'f'):
        raise ValueError("Invalid value for train_size: {}".format(train_size))
    if test_size is not None and test_size_type not in ('i', 'f'):
        raise ValueError("Invalid value for test_size: {}".format(test_size))

    if (train_size_type == 'f' and test_size_type == 'f' and
            train_size + test_size > 1):
        raise ValueError(
            'The sum of test_size and train_size = {}, should be in the (0, 1)'
            ' range. Reduce test_size and/or train_size.'
                .format(train_size + test_size))

    if test_size_type == 'f':
        n_test = ceil(test_size * n_samples)
    elif test_size_type == 'i':
        n_test = float(test_size)

    if train_size_type == 'f':
        n_train = floor(train_size * n_samples)
    elif train_size_type == 'i':
        n_train = float(train_size)

    if train_size is None:
        n_train = n_samples - n_test
    elif test_size is None:
        n_test = n_samples - n_train

    if n_train + n_test > n_samples:
        raise ValueError('The sum of train_size and test_size = %d, '
                         'should be smaller than the number of '
                         'samples %d. Reduce test_size and/or '
                         'train_size.' % (n_train + n_test, n_samples))

    n_train, n_test = int(n_train), int(n_test)

    if n_train == 0:
        raise ValueError(
            'With n_samples={}, test_size={} and train_size={}, the '
            'resulting train set will be empty. Adjust any of the '
            'aforementioned parameters.'.format(n_samples, test_size,
                                                train_size)
        )

    return n_train, n_test


def _list_indexing(X, key, key_dtype):
    """Index a Python list."""
    if np.isscalar(key) or isinstance(key, slice):
        # key is a slice or a scalar
        return X[key]
    if key_dtype == 'bool':
        # key is a boolean array-like
        return list(compress(X, key))
    # key is a integer array-like of key
    return [X[idx] for idx in key]


def _determine_key_type(key, accept_slice=True):
    """Determine the data type of key.

    Parameters
    ----------
    key : scalar, slice or array-like
        The key from which we want to infer the data type.

    accept_slice : bool, default=True
        Whether or not to raise an error if the key is a slice.

    Returns
    -------
    dtype : {'int', 'str', 'bool', None}
        Returns the data type of key.
    """
    err_msg = ("No valid specification of the columns. Only a scalar, list or "
               "slice of all integers or all strings, or boolean mask is "
               "allowed")

    dtype_to_str = {int: 'int', str: 'str', bool: 'bool', np.bool_: 'bool'}
    array_dtype_to_str = {'i': 'int', 'u': 'int', 'b': 'bool', 'O': 'str',
                          'U': 'str', 'S': 'str'}

    if key is None:
        return None
    if isinstance(key, tuple(dtype_to_str.keys())):
        try:
            return dtype_to_str[type(key)]
        except KeyError:
            raise ValueError(err_msg)
    if isinstance(key, slice):
        if not accept_slice:
            raise TypeError(
                'Only array-like or scalar are supported. '
                'A Python slice was given.'
            )
        if key.start is None and key.stop is None:
            return None
        key_start_type = _determine_key_type(key.start)
        key_stop_type = _determine_key_type(key.stop)
        if key_start_type is not None and key_stop_type is not None:
            if key_start_type != key_stop_type:
                raise ValueError(err_msg)
        if key_start_type is not None:
            return key_start_type
        return key_stop_type
    if isinstance(key, (list, tuple)):
        unique_key = set(key)
        key_type = {_determine_key_type(elt) for elt in unique_key}
        if not key_type:
            return None
        if len(key_type) != 1:
            raise ValueError(err_msg)
        return key_type.pop()
    if hasattr(key, 'dtype'):
        try:
            return array_dtype_to_str[key.dtype.kind]
        except KeyError:
            raise ValueError(err_msg)
    raise ValueError(err_msg)


def _safe_indexing(X, indices, *, axis=0):
    """Return rows, items or columns of X using indices.

    .. warning::

        This utility is documented, but **private**. This means that
        backward compatibility might be broken without any deprecation
        cycle.

    Parameters
    ----------
    X : array-like, sparse-matrix, list, pandas.DataFrame, pandas.Series
        Data from which to sample rows, items or columns. `list` are only
        supported when `axis=0`.
    indices : bool, int, str, slice, array-like
        - If `axis=0`, boolean and integer array-like, integer slice,
          and scalar integer are supported.
        - If `axis=1`:
            - to select a single column, `indices` can be of `int` type for
              all `X` types and `str` only for dataframe. The selected subset
              will be 1D, unless `X` is a sparse matrix in which case it will
              be 2D.
            - to select multiples columns, `indices` can be one of the
              following: `list`, `array`, `slice`. The type used in
              these containers can be one of the following: `int`, 'bool' and
              `str`. However, `str` is only supported when `X` is a dataframe.
              The selected subset will be 2D.
    axis : int, default=0
        The axis along which `X` will be subsampled. `axis=0` will select
        rows while `axis=1` will select columns.

    Returns
    -------
    subset
        Subset of X on axis 0 or 1.

    Notes
    -----
    CSR, CSC, and LIL sparse matrices are supported. COO sparse matrices are
    not supported.
    """
    if indices is None:
        return X

    if axis not in (0, 1):
        raise ValueError(
            "'axis' should be either 0 (to index rows) or 1 (to index "
            " column). Got {} instead.".format(axis)
        )

    indices_dtype = _determine_key_type(indices)

    if axis == 0 and indices_dtype == 'str':
        raise ValueError(
            "String indexing is not supported with 'axis=0'"
        )

    if axis == 1 and X.ndim != 2:
        raise ValueError(
            "'X' should be a 2D NumPy array, 2D sparse matrix or pandas "
            "dataframe when indexing the columns (i.e. 'axis=1'). "
            "Got {} instead with {} dimension(s).".format(type(X), X.ndim)
        )

    if axis == 1 and indices_dtype == 'str' and not hasattr(X, 'loc'):
        raise ValueError(
            "Specifying the columns using strings is only supported for "
            "pandas DataFrames"
        )

    return _list_indexing(X, indices, indices_dtype)


def train_test_split(*arrays,
                     test_size=None,
                     train_size=None,
                     random_state=None,
                     shuffle=True,
                     stratify=None):
    n_arrays = len(arrays)
    if n_arrays == 0:
        raise ValueError("At least one array required as input")

    # 已经判断都具有相同长度
    arrays = indexable(*arrays)

    n_samples = _num_samples(arrays[0])

    # 获得train和test合法数字
    n_train, n_test = _validate_shuffle_split(n_samples, test_size, train_size,
                                              default_test_size=0.25)

    if shuffle is False:
        if stratify is not None:
            raise ValueError(
                "Stratified train/test split is not implemented for "
                "shuffle=False")

        train = np.arange(n_train)
        test = np.arange(n_train, n_train + n_test)

    else:
        CVClass = ShuffleSplit

        cv = CVClass(test_size=n_test,
                     train_size=n_train,
                     random_state=random_state)

        train, test = next(cv.split(X=arrays[0], y=stratify))

    return list(chain.from_iterable(
        (_safe_indexing(a, train), _safe_indexing(a, test)) for a in arrays))


X, y = np.arange(10).reshape((5, 2)), range(5)

train_test_split(X, y)

参考资料

• 【sklearn源码】train_test_split