1
0
Fork 0
QuaPy/quapy/data/base.py

566 lines
26 KiB
Python

import itertools
from functools import cached_property
from typing import Iterable
import numpy as np
from scipy.sparse import issparse
from scipy.sparse import vstack
from sklearn.model_selection import train_test_split, RepeatedStratifiedKFold
from numpy.random import RandomState
from quapy.functional import strprev
from quapy.util import temp_seed
class LabelledCollection:
"""
A LabelledCollection is a set of objects each with a label attached to each of them.
This class implements several sampling routines and other utilities.
:param instances: array-like (np.ndarray, list, or csr_matrix are supported)
:param labels: array-like with the same length of instances
:param classes: optional, list of classes from which labels are taken. If not specified, the classes are inferred
from the labels. The classes must be indicated in cases in which some of the labels might have no examples
(i.e., a prevalence of 0)
"""
def __init__(self, instances, labels, classes=None):
if issparse(instances):
self.instances = instances
elif isinstance(instances, list) and len(instances) > 0 and isinstance(instances[0], str):
# lists of strings occupy too much as ndarrays (although python-objects add a heavy overload)
self.instances = np.asarray(instances, dtype=object)
else:
self.instances = np.asarray(instances)
self.labels = np.asarray(labels)
n_docs = len(self)
if classes is None:
self.classes_ = np.unique(self.labels)
self.classes_.sort()
else:
self.classes_ = np.unique(np.asarray(classes))
self.classes_.sort()
if len(set(self.labels).difference(set(classes))) > 0:
raise ValueError(f'labels ({set(self.labels)}) contain values not included in classes_ ({set(classes)})')
self.index = {class_: np.arange(n_docs)[self.labels == class_] for class_ in self.classes_}
@classmethod
def load(cls, path: str, loader_func: callable, classes=None, **loader_kwargs):
"""
Loads a labelled set of data and convert it into a :class:`LabelledCollection` instance. The function in charge
of reading the instances must be specified. This function can be a custom one, or any of the reading functions
defined in :mod:`quapy.data.reader` module.
:param path: string, the path to the file containing the labelled instances
:param loader_func: a custom function that implements the data loader and returns a tuple with instances and
labels
:param classes: array-like, the classes according to which the instances are labelled
:param loader_kwargs: any argument that the `loader_func` function needs in order to read the instances, i.e.,
these arguments are used to call `loader_func(path, **loader_kwargs)`
:return: a :class:`LabelledCollection` object
"""
return LabelledCollection(*loader_func(path, **loader_kwargs), classes)
def __len__(self):
"""
Returns the length of this collection (number of labelled instances)
:return: integer
"""
return self.instances.shape[0]
def prevalence(self):
"""
Returns the prevalence, or relative frequency, of the classes in the codeframe.
:return: a np.ndarray of shape `(n_classes)` with the relative frequencies of each class, in the same order
as listed by `self.classes_`
"""
return self.counts() / len(self)
def counts(self):
"""
Returns the number of instances for each of the classes in the codeframe.
:return: a np.ndarray of shape `(n_classes)` with the number of instances of each class, in the same order
as listed by `self.classes_`
"""
return np.asarray([len(self.index[class_]) for class_ in self.classes_])
@property
def n_classes(self):
"""
The number of classes
:return: integer
"""
return len(self.classes_)
@property
def binary(self):
"""
Returns True if the number of classes is 2
:return: boolean
"""
return self.n_classes == 2
def sampling_index(self, size, *prevs, shuffle=True, random_state=None):
"""
Returns an index to be used to extract a random sample of desired size and desired prevalence values. If the
prevalence values are not specified, then returns the index of a uniform sampling.
For each class, the sampling is drawn with replacement if the requested prevalence is larger than
the actual prevalence of the class, or without replacement otherwise.
:param size: integer, the requested size
:param prevs: the prevalence for each class; the prevalence value for the last class can be lead empty since
it is constrained. E.g., for binary collections, only the prevalence `p` for the first class (as listed in
`self.classes_` can be specified, while the other class takes prevalence value `1-p`
:param shuffle: if set to True (default), shuffles the index before returning it
:param random_state: seed for reproducing sampling
:return: a np.ndarray of shape `(size)` with the indexes
"""
if len(prevs) == 0: # no prevalence was indicated; returns an index for uniform sampling
return self.uniform_sampling_index(size, random_state=random_state)
if len(prevs) == self.n_classes - 1:
prevs = prevs + (1 - sum(prevs),)
assert len(prevs) == self.n_classes, 'unexpected number of prevalences'
assert sum(prevs) == 1, f'prevalences ({prevs}) wrong range (sum={sum(prevs)})'
# Decide how many instances should be taken for each class in order to satisfy the requested prevalence
# accurately, and the number of instances in the sample (exactly). If int(size * prevs[i]) (which is
# <= size * prevs[i]) examples are drawn from class i, there could be a remainder number of instances to take
# to satisfy the size constrain. The remainder is distributed along the classes with probability = prevs.
# (This aims at avoiding the remainder to be placed in a class for which the prevalence requested is 0.)
n_requests = {class_: round(size * prevs[i]) for i, class_ in enumerate(self.classes_)}
remainder = size - sum(n_requests.values())
with temp_seed(random_state):
# due to rounding, the remainder can be 0, >0, or <0
if remainder > 0:
# when the remainder is >0 we randomly add 1 to the requests for each class;
# more prevalent classes are more likely to be taken in order to minimize the impact in the final prevalence
for rand_class in np.random.choice(self.classes_, size=remainder, p=prevs):
n_requests[rand_class] += 1
elif remainder < 0:
# when the remainder is <0 we randomly remove 1 from the requests, unless the request is 0 for a chosen
# class; we repeat until remainder==0
while remainder!=0:
rand_class = np.random.choice(self.classes_, p=prevs)
if n_requests[rand_class] > 0:
n_requests[rand_class] -= 1
remainder += 1
indexes_sample = []
for class_, n_requested in n_requests.items():
n_candidates = len(self.index[class_])
index_sample = self.index[class_][
np.random.choice(n_candidates, size=n_requested, replace=(n_requested > n_candidates))
] if n_requested > 0 else []
indexes_sample.append(index_sample)
indexes_sample = np.concatenate(indexes_sample).astype(int)
if shuffle:
indexes_sample = np.random.permutation(indexes_sample)
return indexes_sample
def uniform_sampling_index(self, size, random_state=None):
"""
Returns an index to be used to extract a uniform sample of desired size. The sampling is drawn
with replacement if the requested size is greater than the number of instances, or without replacement
otherwise.
:param size: integer, the size of the uniform sample
:param random_state: if specified, guarantees reproducibility of the split.
:return: a np.ndarray of shape `(size)` with the indexes
"""
if random_state is not None:
ng = RandomState(seed=random_state)
else:
ng = np.random
return ng.choice(len(self), size, replace=size > len(self))
def sampling(self, size, *prevs, shuffle=True, random_state=None):
"""
Return a random sample (an instance of :class:`LabelledCollection`) of desired size and desired prevalence
values. For each class, the sampling is drawn without replacement if the requested prevalence is larger than
the actual prevalence of the class, or with replacement otherwise.
:param size: integer, the requested size
:param prevs: the prevalence for each class; the prevalence value for the last class can be lead empty since
it is constrained. E.g., for binary collections, only the prevalence `p` for the first class (as listed in
`self.classes_` can be specified, while the other class takes prevalence value `1-p`
:param shuffle: if set to True (default), shuffles the index before returning it
:param random_state: seed for reproducing sampling
:return: an instance of :class:`LabelledCollection` with length == `size` and prevalence close to `prevs` (or
prevalence == `prevs` if the exact prevalence values can be met as proportions of instances)
"""
prev_index = self.sampling_index(size, *prevs, shuffle=shuffle, random_state=random_state)
return self.sampling_from_index(prev_index)
def uniform_sampling(self, size, random_state=None):
"""
Returns a uniform sample (an instance of :class:`LabelledCollection`) of desired size. The sampling is drawn
with replacement if the requested size is greater than the number of instances, or without replacement
otherwise.
:param size: integer, the requested size
:param random_state: if specified, guarantees reproducibility of the split.
:return: an instance of :class:`LabelledCollection` with length == `size`
"""
unif_index = self.uniform_sampling_index(size, random_state=random_state)
return self.sampling_from_index(unif_index)
def sampling_from_index(self, index):
"""
Returns an instance of :class:`LabelledCollection` whose elements are sampled from this collection using the
index.
:param index: np.ndarray
:return: an instance of :class:`LabelledCollection`
"""
documents = self.instances[index]
labels = self.labels[index]
return LabelledCollection(documents, labels, classes=self.classes_)
def split_stratified(self, train_prop=0.6, random_state=None):
"""
Returns two instances of :class:`LabelledCollection` split with stratification from this collection, at desired
proportion.
:param train_prop: the proportion of elements to include in the left-most returned collection (typically used
as the training collection). The rest of elements are included in the right-most returned collection
(typically used as a test collection).
:param random_state: if specified, guarantees reproducibility of the split.
:return: two instances of :class:`LabelledCollection`, the first one with `train_prop` elements, and the
second one with `1-train_prop` elements
"""
tr_docs, te_docs, tr_labels, te_labels = train_test_split(
self.instances, self.labels, train_size=train_prop, stratify=self.labels, random_state=random_state
)
training = LabelledCollection(tr_docs, tr_labels, classes=self.classes_)
test = LabelledCollection(te_docs, te_labels, classes=self.classes_)
return training, test
def split_random(self, train_prop=0.6, random_state=None):
"""
Returns two instances of :class:`LabelledCollection` split randomly from this collection, at desired
proportion.
:param train_prop: the proportion of elements to include in the left-most returned collection (typically used
as the training collection). The rest of elements are included in the right-most returned collection
(typically used as a test collection).
:param random_state: if specified, guarantees reproducibility of the split.
:return: two instances of :class:`LabelledCollection`, the first one with `train_prop` elements, and the
second one with `1-train_prop` elements
"""
indexes = np.random.RandomState(seed=random_state).permutation(len(self))
if isinstance(train_prop, int):
assert train_prop < len(self), \
'argument train_prop cannot be greater than the number of elements in the collection'
splitpoint = train_prop
elif isinstance(train_prop, float):
assert 0 < train_prop < 1, \
'argument train_prop out of range (0,1)'
splitpoint = int(np.round(len(self)*train_prop))
left, right = indexes[:splitpoint], indexes[splitpoint:]
training = self.sampling_from_index(left)
test = self.sampling_from_index(right)
return training, test
def __add__(self, other):
"""
Returns a new :class:`LabelledCollection` as the union of this collection with another collection.
Both labelled collections must have the same classes.
:param other: another :class:`LabelledCollection`
:return: a :class:`LabelledCollection` representing the union of both collections
"""
if not all(np.sort(self.classes_)==np.sort(other.classes_)):
raise NotImplementedError(f'unsupported operation for collections on different classes; '
f'expected {self.classes_}, found {other.classes_}')
return LabelledCollection.join(self, other)
@classmethod
def join(cls, *args: Iterable['LabelledCollection']):
"""
Returns a new :class:`LabelledCollection` as the union of the collections given in input.
:param args: instances of :class:`LabelledCollection`
:return: a :class:`LabelledCollection` representing the union of both collections
"""
args = [lc for lc in args if lc is not None]
assert len(args) > 0, 'empty list is not allowed for mix'
assert all([isinstance(lc, LabelledCollection) for lc in args]), \
'only instances of LabelledCollection allowed'
first_instances = args[0].instances
first_type = type(first_instances)
assert all([type(lc.instances)==first_type for lc in args[1:]]), \
'not all the collections are of instances of the same type'
if issparse(first_instances) or isinstance(first_instances, np.ndarray):
first_ndim = first_instances.ndim
assert all([lc.instances.ndim == first_ndim for lc in args[1:]]), \
'not all the ndarrays are of the same dimension'
if first_ndim > 1:
first_shape = first_instances.shape[1:]
assert all([lc.instances.shape[1:] == first_shape for lc in args[1:]]), \
'not all the ndarrays are of the same shape'
if issparse(first_instances):
instances = vstack([lc.instances for lc in args])
else:
instances = np.concatenate([lc.instances for lc in args])
elif isinstance(first_instances, list):
instances = list(itertools.chain(lc.instances for lc in args))
else:
raise NotImplementedError('unsupported operation for collection types')
labels = np.concatenate([lc.labels for lc in args])
classes = np.unique(labels).sort()
return LabelledCollection(instances, labels, classes=classes)
@property
def Xy(self):
"""
Gets the instances and labels. This is useful when working with `sklearn` estimators, e.g.:
>>> svm = LinearSVC().fit(*my_collection.Xy)
:return: a tuple `(instances, labels)` from this collection
"""
return self.instances, self.labels
@property
def Xp(self):
"""
Gets the instances and the true prevalence. This is useful when implementing evaluation protocols from
a :class:`LabelledCollection` object.
:return: a tuple `(instances, prevalence)` from this collection
"""
return self.instances, self.prevalence()
@property
def X(self):
"""
An alias to self.instances
:return: self.instances
"""
return self.instances
@property
def y(self):
"""
An alias to self.labels
:return: self.labels
"""
return self.labels
@property
def p(self):
"""
An alias to self.prevalence()
:return: self.prevalence()
"""
return self.prevalence()
def stats(self, show=True):
"""
Returns (and eventually prints) a dictionary with some stats of this collection. E.g.,:
>>> data = qp.datasets.fetch_reviews('kindle', tfidf=True, min_df=5)
>>> data.training.stats()
>>> #instances=3821, type=<class 'scipy.sparse.csr.csr_matrix'>, #features=4403, #classes=[0 1], prevs=[0.081, 0.919]
:param show: if set to True (default), prints the stats in standard output
:return: a dictionary containing some stats of this collection. Keys include `#instances` (the number of
instances), `type` (the type representing the instances), `#features` (the number of features, if the
instances are in array-like format), `#classes` (the classes of the collection), `prevs` (the prevalence
values for each class)
"""
ninstances = len(self)
instance_type = type(self.instances[0])
if instance_type == list:
nfeats = len(self.instances[0])
elif instance_type == np.ndarray or issparse(self.instances):
nfeats = self.instances.shape[1]
else:
nfeats = '?'
stats_ = {'instances': ninstances,
'type': instance_type,
'features': nfeats,
'classes': self.classes_,
'prevs': strprev(self.prevalence())}
if show:
print(f'#instances={stats_["instances"]}, type={stats_["type"]}, #features={stats_["features"]}, '
f'#classes={stats_["classes"]}, prevs={stats_["prevs"]}')
return stats_
def kFCV(self, nfolds=5, nrepeats=1, random_state=None):
"""
Generator of stratified folds to be used in k-fold cross validation.
:param nfolds: integer (default 5), the number of folds to generate
:param nrepeats: integer (default 1), the number of rounds of k-fold cross validation to run
:param random_state: integer (default 0), guarantees that the folds generated are reproducible
:return: yields `nfolds * nrepeats` folds for k-fold cross validation
"""
kf = RepeatedStratifiedKFold(n_splits=nfolds, n_repeats=nrepeats, random_state=random_state)
for train_index, test_index in kf.split(*self.Xy):
train = self.sampling_from_index(train_index)
test = self.sampling_from_index(test_index)
yield train, test
class Dataset:
"""
Abstraction of training and test :class:`LabelledCollection` objects.
:param training: a :class:`LabelledCollection` instance
:param test: a :class:`LabelledCollection` instance
:param vocabulary: if indicated, is a dictionary of the terms used in this textual dataset
:param name: a string representing the name of the dataset
"""
def __init__(self, training: LabelledCollection, test: LabelledCollection, vocabulary: dict = None, name=''):
assert set(training.classes_) == set(test.classes_), 'incompatible labels in training and test collections'
self.training = training
self.test = test
self.vocabulary = vocabulary
self.name = name
@classmethod
def SplitStratified(cls, collection: LabelledCollection, train_size=0.6):
"""
Generates a :class:`Dataset` from a stratified split of a :class:`LabelledCollection` instance.
See :meth:`LabelledCollection.split_stratified`
:param collection: :class:`LabelledCollection`
:param train_size: the proportion of training documents (the rest conforms the test split)
:return: an instance of :class:`Dataset`
"""
return Dataset(*collection.split_stratified(train_prop=train_size))
@property
def classes_(self):
"""
The classes according to which the training collection is labelled
:return: The classes according to which the training collection is labelled
"""
return self.training.classes_
@property
def n_classes(self):
"""
The number of classes according to which the training collection is labelled
:return: integer
"""
return self.training.n_classes
@property
def binary(self):
"""
Returns True if the training collection is labelled according to two classes
:return: boolean
"""
return self.training.binary
@classmethod
def load(cls, train_path, test_path, loader_func: callable, classes=None, **loader_kwargs):
"""
Loads a training and a test labelled set of data and convert it into a :class:`Dataset` instance.
The function in charge of reading the instances must be specified. This function can be a custom one, or any of
the reading functions defined in :mod:`quapy.data.reader` module.
:param train_path: string, the path to the file containing the training instances
:param test_path: string, the path to the file containing the test instances
:param loader_func: a custom function that implements the data loader and returns a tuple with instances and
labels
:param classes: array-like, the classes according to which the instances are labelled
:param loader_kwargs: any argument that the `loader_func` function needs in order to read the instances.
See :meth:`LabelledCollection.load` for further details.
:return: a :class:`Dataset` object
"""
training = LabelledCollection.load(train_path, loader_func, classes, **loader_kwargs)
test = LabelledCollection.load(test_path, loader_func, classes, **loader_kwargs)
return Dataset(training, test)
@property
def vocabulary_size(self):
"""
If the dataset is textual, and the vocabulary was indicated, returns the size of the vocabulary
:return: integer
"""
return len(self.vocabulary)
@property
def train_test(self):
"""
Alias to `self.training` and `self.test`
:return: the training and test collections
:return: the training and test collections
"""
return self.training, self.test
def stats(self, show=True):
"""
Returns (and eventually prints) a dictionary with some stats of this dataset. E.g.,:
>>> data = qp.datasets.fetch_reviews('kindle', tfidf=True, min_df=5)
>>> data.stats()
>>> Dataset=kindle #tr-instances=3821, #te-instances=21591, type=<class 'scipy.sparse.csr.csr_matrix'>, #features=4403, #classes=[0 1], tr-prevs=[0.081, 0.919], te-prevs=[0.063, 0.937]
:param show: if set to True (default), prints the stats in standard output
:return: a dictionary containing some stats of this collection for the training and test collections. The keys
are `train` and `test`, and point to dedicated dictionaries of stats, for each collection, with keys
`#instances` (the number of instances), `type` (the type representing the instances),
`#features` (the number of features, if the instances are in array-like format), `#classes` (the classes of
the collection), `prevs` (the prevalence values for each class)
"""
tr_stats = self.training.stats(show=False)
te_stats = self.test.stats(show=False)
if show:
print(f'Dataset={self.name} #tr-instances={tr_stats["instances"]}, #te-instances={te_stats["instances"]}, '
f'type={tr_stats["type"]}, #features={tr_stats["features"]}, #classes={tr_stats["classes"]}, '
f'tr-prevs={tr_stats["prevs"]}, te-prevs={te_stats["prevs"]}')
return {'train': tr_stats, 'test': te_stats}
@classmethod
def kFCV(cls, data: LabelledCollection, nfolds=5, nrepeats=1, random_state=0):
"""
Generator of stratified folds to be used in k-fold cross validation. This function is only a wrapper around
:meth:`LabelledCollection.kFCV` that returns :class:`Dataset` instances made of training and test folds.
:param nfolds: integer (default 5), the number of folds to generate
:param nrepeats: integer (default 1), the number of rounds of k-fold cross validation to run
:param random_state: integer (default 0), guarantees that the folds generated are reproducible
:return: yields `nfolds * nrepeats` folds for k-fold cross validation as instances of :class:`Dataset`
"""
for i, (train, test) in enumerate(data.kFCV(nfolds=nfolds, nrepeats=nrepeats, random_state=random_state)):
yield Dataset(train, test, name=f'fold {(i % nfolds) + 1}/{nfolds} (round={(i // nfolds) + 1})')
def reduce(self, n_train=100, n_test=100):
"""
Reduce the number of instances in place for quick experiments. Preserves the prevalence of each set.
:param n_train: number of training documents to keep (default 100)
:param n_test: number of test documents to keep (default 100)
:return: self
"""
self.training = self.training.sampling(n_train, *self.training.prevalence())
self.test = self.test.sampling(n_test, *self.test.prevalence())
return self