447 lines
15 KiB
Python
447 lines
15 KiB
Python
# Author: Arnaud Joly, Joel Nothman, Hamzeh Alsalhi
|
|
#
|
|
# License: BSD 3 clause
|
|
"""
|
|
Multi-class / multi-label utility function
|
|
==========================================
|
|
|
|
"""
|
|
from __future__ import division
|
|
from itertools import chain
|
|
|
|
from scipy.sparse import issparse
|
|
from scipy.sparse.base import spmatrix
|
|
from scipy.sparse import dok_matrix
|
|
from scipy.sparse import lil_matrix
|
|
|
|
import numpy as np
|
|
|
|
from ..externals.six import string_types
|
|
from ..utils.fixes import _Sequence as Sequence
|
|
from .validation import check_array
|
|
|
|
|
|
def _unique_multiclass(y):
|
|
if hasattr(y, '__array__'):
|
|
return np.unique(np.asarray(y))
|
|
else:
|
|
return set(y)
|
|
|
|
|
|
def _unique_indicator(y):
|
|
return np.arange(check_array(y, ['csr', 'csc', 'coo']).shape[1])
|
|
|
|
|
|
_FN_UNIQUE_LABELS = {
|
|
'binary': _unique_multiclass,
|
|
'multiclass': _unique_multiclass,
|
|
'multilabel-indicator': _unique_indicator,
|
|
}
|
|
|
|
|
|
def unique_labels(*ys):
|
|
"""Extract an ordered array of unique labels
|
|
|
|
We don't allow:
|
|
- mix of multilabel and multiclass (single label) targets
|
|
- mix of label indicator matrix and anything else,
|
|
because there are no explicit labels)
|
|
- mix of label indicator matrices of different sizes
|
|
- mix of string and integer labels
|
|
|
|
At the moment, we also don't allow "multiclass-multioutput" input type.
|
|
|
|
Parameters
|
|
----------
|
|
*ys : array-likes,
|
|
|
|
Returns
|
|
-------
|
|
out : numpy array of shape [n_unique_labels]
|
|
An ordered array of unique labels.
|
|
|
|
Examples
|
|
--------
|
|
>>> from sklearn.utils.multiclass import unique_labels
|
|
>>> unique_labels([3, 5, 5, 5, 7, 7])
|
|
array([3, 5, 7])
|
|
>>> unique_labels([1, 2, 3, 4], [2, 2, 3, 4])
|
|
array([1, 2, 3, 4])
|
|
>>> unique_labels([1, 2, 10], [5, 11])
|
|
array([ 1, 2, 5, 10, 11])
|
|
"""
|
|
if not ys:
|
|
raise ValueError('No argument has been passed.')
|
|
# Check that we don't mix label format
|
|
|
|
ys_types = set(type_of_target(x) for x in ys)
|
|
if ys_types == set(["binary", "multiclass"]):
|
|
ys_types = set(["multiclass"])
|
|
|
|
if len(ys_types) > 1:
|
|
raise ValueError("Mix type of y not allowed, got types %s" % ys_types)
|
|
|
|
label_type = ys_types.pop()
|
|
|
|
# Check consistency for the indicator format
|
|
if (label_type == "multilabel-indicator" and
|
|
len(set(check_array(y, ['csr', 'csc', 'coo']).shape[1]
|
|
for y in ys)) > 1):
|
|
raise ValueError("Multi-label binary indicator input with "
|
|
"different numbers of labels")
|
|
|
|
# Get the unique set of labels
|
|
_unique_labels = _FN_UNIQUE_LABELS.get(label_type, None)
|
|
if not _unique_labels:
|
|
raise ValueError("Unknown label type: %s" % repr(ys))
|
|
|
|
ys_labels = set(chain.from_iterable(_unique_labels(y) for y in ys))
|
|
|
|
# Check that we don't mix string type with number type
|
|
if (len(set(isinstance(label, string_types) for label in ys_labels)) > 1):
|
|
raise ValueError("Mix of label input types (string and number)")
|
|
|
|
return np.array(sorted(ys_labels))
|
|
|
|
|
|
def _is_integral_float(y):
|
|
return y.dtype.kind == 'f' and np.all(y.astype(int) == y)
|
|
|
|
|
|
def is_multilabel(y):
|
|
""" Check if ``y`` is in a multilabel format.
|
|
|
|
Parameters
|
|
----------
|
|
y : numpy array of shape [n_samples]
|
|
Target values.
|
|
|
|
Returns
|
|
-------
|
|
out : bool,
|
|
Return ``True``, if ``y`` is in a multilabel format, else ```False``.
|
|
|
|
Examples
|
|
--------
|
|
>>> import numpy as np
|
|
>>> from sklearn.utils.multiclass import is_multilabel
|
|
>>> is_multilabel([0, 1, 0, 1])
|
|
False
|
|
>>> is_multilabel([[1], [0, 2], []])
|
|
False
|
|
>>> is_multilabel(np.array([[1, 0], [0, 0]]))
|
|
True
|
|
>>> is_multilabel(np.array([[1], [0], [0]]))
|
|
False
|
|
>>> is_multilabel(np.array([[1, 0, 0]]))
|
|
True
|
|
"""
|
|
if hasattr(y, '__array__'):
|
|
y = np.asarray(y)
|
|
if not (hasattr(y, "shape") and y.ndim == 2 and y.shape[1] > 1):
|
|
return False
|
|
|
|
if issparse(y):
|
|
if isinstance(y, (dok_matrix, lil_matrix)):
|
|
y = y.tocsr()
|
|
return (len(y.data) == 0 or np.unique(y.data).size == 1 and
|
|
(y.dtype.kind in 'biu' or # bool, int, uint
|
|
_is_integral_float(np.unique(y.data))))
|
|
else:
|
|
labels = np.unique(y)
|
|
|
|
return len(labels) < 3 and (y.dtype.kind in 'biu' or # bool, int, uint
|
|
_is_integral_float(labels))
|
|
|
|
|
|
def check_classification_targets(y):
|
|
"""Ensure that target y is of a non-regression type.
|
|
|
|
Only the following target types (as defined in type_of_target) are allowed:
|
|
'binary', 'multiclass', 'multiclass-multioutput',
|
|
'multilabel-indicator', 'multilabel-sequences'
|
|
|
|
Parameters
|
|
----------
|
|
y : array-like
|
|
"""
|
|
y_type = type_of_target(y)
|
|
if y_type not in ['binary', 'multiclass', 'multiclass-multioutput',
|
|
'multilabel-indicator', 'multilabel-sequences']:
|
|
raise ValueError("Unknown label type: %r" % y_type)
|
|
|
|
|
|
def type_of_target(y):
|
|
"""Determine the type of data indicated by the target.
|
|
|
|
Note that this type is the most specific type that can be inferred.
|
|
For example:
|
|
|
|
* ``binary`` is more specific but compatible with ``multiclass``.
|
|
* ``multiclass`` of integers is more specific but compatible with
|
|
``continuous``.
|
|
* ``multilabel-indicator`` is more specific but compatible with
|
|
``multiclass-multioutput``.
|
|
|
|
Parameters
|
|
----------
|
|
y : array-like
|
|
|
|
Returns
|
|
-------
|
|
target_type : string
|
|
One of:
|
|
|
|
* 'continuous': `y` is an array-like of floats that are not all
|
|
integers, and is 1d or a column vector.
|
|
* 'continuous-multioutput': `y` is a 2d array of floats that are
|
|
not all integers, and both dimensions are of size > 1.
|
|
* 'binary': `y` contains <= 2 discrete values and is 1d or a column
|
|
vector.
|
|
* 'multiclass': `y` contains more than two discrete values, is not a
|
|
sequence of sequences, and is 1d or a column vector.
|
|
* 'multiclass-multioutput': `y` is a 2d array that contains more
|
|
than two discrete values, is not a sequence of sequences, and both
|
|
dimensions are of size > 1.
|
|
* 'multilabel-indicator': `y` is a label indicator matrix, an array
|
|
of two dimensions with at least two columns, and at most 2 unique
|
|
values.
|
|
* 'unknown': `y` is array-like but none of the above, such as a 3d
|
|
array, sequence of sequences, or an array of non-sequence objects.
|
|
|
|
Examples
|
|
--------
|
|
>>> import numpy as np
|
|
>>> type_of_target([0.1, 0.6])
|
|
'continuous'
|
|
>>> type_of_target([1, -1, -1, 1])
|
|
'binary'
|
|
>>> type_of_target(['a', 'b', 'a'])
|
|
'binary'
|
|
>>> type_of_target([1.0, 2.0])
|
|
'binary'
|
|
>>> type_of_target([1, 0, 2])
|
|
'multiclass'
|
|
>>> type_of_target([1.0, 0.0, 3.0])
|
|
'multiclass'
|
|
>>> type_of_target(['a', 'b', 'c'])
|
|
'multiclass'
|
|
>>> type_of_target(np.array([[1, 2], [3, 1]]))
|
|
'multiclass-multioutput'
|
|
>>> type_of_target([[1, 2]])
|
|
'multiclass-multioutput'
|
|
>>> type_of_target(np.array([[1.5, 2.0], [3.0, 1.6]]))
|
|
'continuous-multioutput'
|
|
>>> type_of_target(np.array([[0, 1], [1, 1]]))
|
|
'multilabel-indicator'
|
|
"""
|
|
valid = ((isinstance(y, (Sequence, spmatrix)) or hasattr(y, '__array__'))
|
|
and not isinstance(y, string_types))
|
|
|
|
if not valid:
|
|
raise ValueError('Expected array-like (array or non-string sequence), '
|
|
'got %r' % y)
|
|
|
|
sparseseries = (y.__class__.__name__ == 'SparseSeries')
|
|
if sparseseries:
|
|
raise ValueError("y cannot be class 'SparseSeries'.")
|
|
|
|
if is_multilabel(y):
|
|
return 'multilabel-indicator'
|
|
|
|
try:
|
|
y = np.asarray(y)
|
|
except ValueError:
|
|
# Known to fail in numpy 1.3 for array of arrays
|
|
return 'unknown'
|
|
|
|
# The old sequence of sequences format
|
|
try:
|
|
if (not hasattr(y[0], '__array__') and isinstance(y[0], Sequence)
|
|
and not isinstance(y[0], string_types)):
|
|
raise ValueError('You appear to be using a legacy multi-label data'
|
|
' representation. Sequence of sequences are no'
|
|
' longer supported; use a binary array or sparse'
|
|
' matrix instead.')
|
|
except IndexError:
|
|
pass
|
|
|
|
# Invalid inputs
|
|
if y.ndim > 2 or (y.dtype == object and len(y) and
|
|
not isinstance(y.flat[0], string_types)):
|
|
return 'unknown' # [[[1, 2]]] or [obj_1] and not ["label_1"]
|
|
|
|
if y.ndim == 2 and y.shape[1] == 0:
|
|
return 'unknown' # [[]]
|
|
|
|
if y.ndim == 2 and y.shape[1] > 1:
|
|
suffix = "-multioutput" # [[1, 2], [1, 2]]
|
|
else:
|
|
suffix = "" # [1, 2, 3] or [[1], [2], [3]]
|
|
|
|
# check float and contains non-integer float values
|
|
if y.dtype.kind == 'f' and np.any(y != y.astype(int)):
|
|
# [.1, .2, 3] or [[.1, .2, 3]] or [[1., .2]] and not [1., 2., 3.]
|
|
return 'continuous' + suffix
|
|
|
|
if (len(np.unique(y)) > 2) or (y.ndim >= 2 and len(y[0]) > 1):
|
|
return 'multiclass' + suffix # [1, 2, 3] or [[1., 2., 3]] or [[1, 2]]
|
|
else:
|
|
return 'binary' # [1, 2] or [["a"], ["b"]]
|
|
|
|
|
|
def _check_partial_fit_first_call(clf, classes=None):
|
|
"""Private helper function for factorizing common classes param logic
|
|
|
|
Estimators that implement the ``partial_fit`` API need to be provided with
|
|
the list of possible classes at the first call to partial_fit.
|
|
|
|
Subsequent calls to partial_fit should check that ``classes`` is still
|
|
consistent with a previous value of ``clf.classes_`` when provided.
|
|
|
|
This function returns True if it detects that this was the first call to
|
|
``partial_fit`` on ``clf``. In that case the ``classes_`` attribute is also
|
|
set on ``clf``.
|
|
|
|
"""
|
|
if getattr(clf, 'classes_', None) is None and classes is None:
|
|
raise ValueError("classes must be passed on the first call "
|
|
"to partial_fit.")
|
|
|
|
elif classes is not None:
|
|
if getattr(clf, 'classes_', None) is not None:
|
|
if not np.array_equal(clf.classes_, unique_labels(classes)):
|
|
raise ValueError(
|
|
"`classes=%r` is not the same as on last call "
|
|
"to partial_fit, was: %r" % (classes, clf.classes_))
|
|
|
|
else:
|
|
# This is the first call to partial_fit
|
|
clf.classes_ = unique_labels(classes)
|
|
return True
|
|
|
|
# classes is None and clf.classes_ has already previously been set:
|
|
# nothing to do
|
|
return False
|
|
|
|
|
|
def class_distribution(y, sample_weight=None):
|
|
"""Compute class priors from multioutput-multiclass target data
|
|
|
|
Parameters
|
|
----------
|
|
y : array like or sparse matrix of size (n_samples, n_outputs)
|
|
The labels for each example.
|
|
|
|
sample_weight : array-like of shape = (n_samples,), optional
|
|
Sample weights.
|
|
|
|
Returns
|
|
-------
|
|
classes : list of size n_outputs of arrays of size (n_classes,)
|
|
List of classes for each column.
|
|
|
|
n_classes : list of integers of size n_outputs
|
|
Number of classes in each column
|
|
|
|
class_prior : list of size n_outputs of arrays of size (n_classes,)
|
|
Class distribution of each column.
|
|
|
|
"""
|
|
classes = []
|
|
n_classes = []
|
|
class_prior = []
|
|
|
|
n_samples, n_outputs = y.shape
|
|
|
|
if issparse(y):
|
|
y = y.tocsc()
|
|
y_nnz = np.diff(y.indptr)
|
|
|
|
for k in range(n_outputs):
|
|
col_nonzero = y.indices[y.indptr[k]:y.indptr[k + 1]]
|
|
# separate sample weights for zero and non-zero elements
|
|
if sample_weight is not None:
|
|
nz_samp_weight = np.asarray(sample_weight)[col_nonzero]
|
|
zeros_samp_weight_sum = (np.sum(sample_weight) -
|
|
np.sum(nz_samp_weight))
|
|
else:
|
|
nz_samp_weight = None
|
|
zeros_samp_weight_sum = y.shape[0] - y_nnz[k]
|
|
|
|
classes_k, y_k = np.unique(y.data[y.indptr[k]:y.indptr[k + 1]],
|
|
return_inverse=True)
|
|
class_prior_k = np.bincount(y_k, weights=nz_samp_weight)
|
|
|
|
# An explicit zero was found, combine its weight with the weight
|
|
# of the implicit zeros
|
|
if 0 in classes_k:
|
|
class_prior_k[classes_k == 0] += zeros_samp_weight_sum
|
|
|
|
# If an there is an implicit zero and it is not in classes and
|
|
# class_prior, make an entry for it
|
|
if 0 not in classes_k and y_nnz[k] < y.shape[0]:
|
|
classes_k = np.insert(classes_k, 0, 0)
|
|
class_prior_k = np.insert(class_prior_k, 0,
|
|
zeros_samp_weight_sum)
|
|
|
|
classes.append(classes_k)
|
|
n_classes.append(classes_k.shape[0])
|
|
class_prior.append(class_prior_k / class_prior_k.sum())
|
|
else:
|
|
for k in range(n_outputs):
|
|
classes_k, y_k = np.unique(y[:, k], return_inverse=True)
|
|
classes.append(classes_k)
|
|
n_classes.append(classes_k.shape[0])
|
|
class_prior_k = np.bincount(y_k, weights=sample_weight)
|
|
class_prior.append(class_prior_k / class_prior_k.sum())
|
|
|
|
return (classes, n_classes, class_prior)
|
|
|
|
|
|
def _ovr_decision_function(predictions, confidences, n_classes):
|
|
"""Compute a continuous, tie-breaking ovr decision function.
|
|
|
|
It is important to include a continuous value, not only votes,
|
|
to make computing AUC or calibration meaningful.
|
|
|
|
Parameters
|
|
----------
|
|
predictions : array-like, shape (n_samples, n_classifiers)
|
|
Predicted classes for each binary classifier.
|
|
|
|
confidences : array-like, shape (n_samples, n_classifiers)
|
|
Decision functions or predicted probabilities for positive class
|
|
for each binary classifier.
|
|
|
|
n_classes : int
|
|
Number of classes. n_classifiers must be
|
|
``n_classes * (n_classes - 1 ) / 2``
|
|
"""
|
|
n_samples = predictions.shape[0]
|
|
votes = np.zeros((n_samples, n_classes))
|
|
sum_of_confidences = np.zeros((n_samples, n_classes))
|
|
|
|
k = 0
|
|
for i in range(n_classes):
|
|
for j in range(i + 1, n_classes):
|
|
sum_of_confidences[:, i] -= confidences[:, k]
|
|
sum_of_confidences[:, j] += confidences[:, k]
|
|
votes[predictions[:, k] == 0, i] += 1
|
|
votes[predictions[:, k] == 1, j] += 1
|
|
k += 1
|
|
|
|
max_confidences = sum_of_confidences.max()
|
|
min_confidences = sum_of_confidences.min()
|
|
|
|
if max_confidences == min_confidences:
|
|
return votes
|
|
|
|
# Scale the sum_of_confidences to (-0.5, 0.5) and add it with votes.
|
|
# The motivation is to use confidence levels as a way to break ties in
|
|
# the votes without switching any decision made based on a difference
|
|
# of 1 vote.
|
|
eps = np.finfo(sum_of_confidences.dtype).eps
|
|
max_abs_confidence = max(abs(max_confidences), abs(min_confidences))
|
|
scale = (0.5 - eps) / max_abs_confidence
|
|
return votes + sum_of_confidences * scale
|