5. Dataset loading utilities¶
sklearn.datasets package embeds some small toy datasets
as introduced in the Getting Started section.
To evaluate the impact of the scale of the dataset (
n_features) while controlling the statistical properties of the data
(typically the correlation and informativeness of the features), it is
also possible to generate synthetic data.
This package also features helpers to fetch larger datasets commonly used by the machine learning community to benchmark algorithm on data that comes from the ‘real world’.
5.1. General dataset API¶
There are three distinct kinds of dataset interfaces for different types of datasets. The simplest one is the interface for sample images, which is described below in the Sample images section.
The dataset generation functions and the svmlight loader share a simplistic
interface, returning a tuple
(X, y) consisting of a
n_features numpy array
X and an array of length
containing the targets
The toy datasets as well as the ‘real world’ datasets and the datasets
fetched from mldata.org have more sophisticated structure.
These functions return a dictionary-like object holding at least two items:
an array of shape
n_features with key
(except for 20newsgroups)
and a numpy array of length
n_samples, containing the target values,
The datasets also contain a description in
DESCR and some contain
See the dataset descriptions below for details.
5.2. Toy datasets¶
scikit-learn comes with a few small standard datasets that do not require to download any file from some external website.
||Load and return the boston house-prices dataset (regression).|
||Load and return the iris dataset (classification).|
||Load and return the diabetes dataset (regression).|
||Load and return the digits dataset (classification).|
||Load and return the linnerud dataset (multivariate regression).|
These datasets are useful to quickly illustrate the behavior of the various algorithms implemented in the scikit. They are however often too small to be representative of real world machine learning tasks.
5.3. Sample images¶
The scikit also embed a couple of sample JPEG images published under Creative Commons license by their authors. Those image can be useful to test algorithms and pipeline on 2D data.
||Load sample images for image manipulation.|
||Load the numpy array of a single sample image|
The default coding of images is based on the
uint8 dtype to
spare memory. Often machine learning algorithms work best if the
input is converted to a floating point representation first. Also,
if you plan to use
pylab.imshow don’t forget to scale to the range
0 - 1 as done in the following example.
5.4. Sample generators¶
In addition, scikit-learn includes various random sample generators that can be used to build artificial datasets of controlled size and complexity.
5.4.1. Generators for classification and clustering¶
These generators produce a matrix of features and corresponding discrete targets.
126.96.36.199. Single label¶
make_classification create multiclass
datasets by allocating each class one or more normally-distributed clusters of
make_blobs provides greater control regarding the centers and
standard deviations of each cluster, and is used to demonstrate clustering.
make_classification specialises in introducing noise by way of:
correlated, redundant and uninformative features; multiple Gaussian clusters
per class; and linear transformations of the feature space.
make_moons generate 2d binary classification
datasets that are challenging to certain algorithms (e.g. centroid-based
clustering or linear classification), including optional Gaussian noise.
They are useful for visualisation. produces Gaussian
data with a spherical decision boundary for binary classification.
make_multilabel_classification generates random samples with multiple
labels, reflecting a bag of words drawn from a mixture of topics. The number of
topics for each document is drawn from a Poisson distribution, and the topics
themselves are drawn from a fixed random distribution. Similarly, the number of
words is drawn from Poisson, with words drawn from a multinomial, where each
topic defines a probability distribution over words. Simplifications with
respect to true bag-of-words mixtures include:
- Per-topic word distributions are independently drawn, where in reality all would be affected by a sparse base distribution, and would be correlated.
- For a document generated from multiple topics, all topics are weighted equally in generating its bag of words.
- Documents without labels words at random, rather than from a base distribution.
5.4.2. Generators for regression¶
make_regression produces regression targets as an optionally-sparse
random linear combination of random features, with noise. Its informative
features may be uncorrelated, or low rank (few features account for most of the
Other regression generators generate functions deterministically from
make_sparse_uncorrelated produces a target as a
linear combination of four features with fixed coefficients.
Others encode explicitly non-linear relations:
make_friedman1 is related by polynomial and sine transforms;
make_friedman2 includes feature multiplication and reciprocation; and
make_friedman3 is similar with an arctan transformation on the target.
5.4.3. Generators for manifold learning¶
||Generate an S curve dataset.|
||Generate a swiss roll dataset.|
5.4.4. Generators for decomposition¶
||Generate a mostly low rank matrix with bell-shaped singular values|
||Generate a signal as a sparse combination of dictionary elements.|
||Generate a random symmetric, positive-definite matrix.|
||Generate a sparse symmetric definite positive matrix.|
5.5. Datasets in svmlight / libsvm format¶
scikit-learn includes utility functions for loading
datasets in the svmlight / libsvm format. In this format, each line
takes the form
<feature-id>:<feature-value> .... This format is especially suitable for sparse datasets.
In this module, scipy sparse CSR matrices are used for
X and numpy arrays are used for
You may load a dataset like as follows:
>>> from sklearn.datasets import load_svmlight_file >>> X_train, y_train = load_svmlight_file("/path/to/train_dataset.txt") ...
You may also load two (or more) datasets at once:
>>> X_train, y_train, X_test, y_test = load_svmlight_files( ... ("/path/to/train_dataset.txt", "/path/to/test_dataset.txt")) ...
In this case,
X_test are guaranteed to have the same number
of features. Another way to achieve the same result is to fix the number of
>>> X_test, y_test = load_svmlight_file( ... "/path/to/test_dataset.txt", n_features=X_train.shape) ...
Public datasets in svmlight / libsvm format: http://www.csie.ntu.edu.tw/~cjlin/libsvmtools/datasets/
Faster API-compatible implementation: https://github.com/mblondel/svmlight-loader
5.6. The Olivetti faces dataset¶
This dataset contains a set of face images taken between April 1992 and April
1994 at AT&T Laboratories Cambridge. The website describing the original
dataset is now defunct, but archived copies can be accessed through
the Internet Archive’s Wayback Machine. The
sklearn.datasets.fetch_olivetti_faces function is the data
fetching / caching function that downloads the data
archive from AT&T.
As described on the original website:
There are ten different images of each of 40 distinct subjects. For some subjects, the images were taken at different times, varying the lighting, facial expressions (open / closed eyes, smiling / not smiling) and facial details (glasses / no glasses). All the images were taken against a dark homogeneous background with the subjects in an upright, frontal position (with tolerance for some side movement).
The image is quantized to 256 grey levels and stored as unsigned 8-bit integers; the loader will convert these to floating point values on the interval [0, 1], which are easier to work with for many algorithms.
The “target” for this database is an integer from 0 to 39 indicating the identity of the person pictured; however, with only 10 examples per class, this relatively small dataset is more interesting from an unsupervised or semi-supervised perspective.
The original dataset consisted of 92 x 112, while the version available here consists of 64x64 images.
When using these images, please give credit to AT&T Laboratories Cambridge.
5.7. The 20 newsgroups text dataset¶
The 20 newsgroups dataset comprises around 18000 newsgroups posts on 20 topics split in two subsets: one for training (or development) and the other one for testing (or for performance evaluation). The split between the train and test set is based upon a messages posted before and after a specific date.
This module contains two loaders. The first one,
returns a list of the raw texts that can be fed to text feature
extractors such as
with custom parameters so as to extract feature vectors.
The second one,
returns ready-to-use features, i.e., it is not necessary to use a feature
sklearn.datasets.fetch_20newsgroups function is a data
fetching / caching functions that downloads the data archive from
the original 20 newsgroups website, extracts the archive contents
~/scikit_learn_data/20news_home folder and calls the
sklearn.datasets.load_files on either the training or
testing set folder, or both of them:
>>> from sklearn.datasets import fetch_20newsgroups >>> newsgroups_train = fetch_20newsgroups(subset='train') >>> from pprint import pprint >>> pprint(list(newsgroups_train.target_names)) ['alt.atheism', 'comp.graphics', 'comp.os.ms-windows.misc', 'comp.sys.ibm.pc.hardware', 'comp.sys.mac.hardware', 'comp.windows.x', 'misc.forsale', 'rec.autos', 'rec.motorcycles', 'rec.sport.baseball', 'rec.sport.hockey', 'sci.crypt', 'sci.electronics', 'sci.med', 'sci.space', 'soc.religion.christian', 'talk.politics.guns', 'talk.politics.mideast', 'talk.politics.misc', 'talk.religion.misc']
The real data lies in the
target attributes. The target
attribute is the integer index of the category:
>>> newsgroups_train.filenames.shape (11314,) >>> newsgroups_train.target.shape (11314,) >>> newsgroups_train.target[:10] array([12, 6, 9, 8, 6, 7, 9, 2, 13, 19])
It is possible to load only a sub-selection of the categories by passing the
list of the categories to load to the
>>> cats = ['alt.atheism', 'sci.space'] >>> newsgroups_train = fetch_20newsgroups(subset='train', categories=cats) >>> list(newsgroups_train.target_names) ['alt.atheism', 'sci.space'] >>> newsgroups_train.filenames.shape (1073,) >>> newsgroups_train.target.shape (1073,) >>> newsgroups_train.target[:10] array([1, 1, 1, 0, 1, 0, 0, 1, 1, 1])
5.7.2. Converting text to vectors¶
In order to feed predictive or clustering models with the text data,
one first need to turn the text into vectors of numerical values suitable
for statistical analysis. This can be achieved with the utilities of the
sklearn.feature_extraction.text as demonstrated in the following
example that extract TF-IDF vectors of unigram tokens
from a subset of 20news:
>>> from sklearn.feature_extraction.text import TfidfVectorizer >>> categories = ['alt.atheism', 'talk.religion.misc', ... 'comp.graphics', 'sci.space'] >>> newsgroups_train = fetch_20newsgroups(subset='train', ... categories=categories) >>> vectorizer = TfidfVectorizer() >>> vectors = vectorizer.fit_transform(newsgroups_train.data) >>> vectors.shape (2034, 34118)
The extracted TF-IDF vectors are very sparse, with an average of 159 non-zero components by sample in a more than 30000-dimensional space (less than .5% non-zero features):
>>> vectors.nnz / float(vectors.shape) 159.01327433628319
sklearn.datasets.fetch_20newsgroups_vectorized is a function which returns
ready-to-use tfidf features instead of file names.
5.7.3. Filtering text for more realistic training¶
It is easy for a classifier to overfit on particular things that appear in the 20 Newsgroups data, such as newsgroup headers. Many classifiers achieve very high F-scores, but their results would not generalize to other documents that aren’t from this window of time.
For example, let’s look at the results of a multinomial Naive Bayes classifier, which is fast to train and achieves a decent F-score:
>>> from sklearn.naive_bayes import MultinomialNB >>> from sklearn import metrics >>> newsgroups_test = fetch_20newsgroups(subset='test', ... categories=categories) >>> vectors_test = vectorizer.transform(newsgroups_test.data) >>> clf = MultinomialNB(alpha=.01) >>> clf.fit(vectors, newsgroups_train.target) >>> pred = clf.predict(vectors_test) >>> metrics.f1_score(newsgroups_test.target, pred, average='weighted') 0.88251152461278892
(The example Classification of text documents using sparse features shuffles the training and test data, instead of segmenting by time, and in that case multinomial Naive Bayes gets a much higher F-score of 0.88. Are you suspicious yet of what’s going on inside this classifier?)
Let’s take a look at what the most informative features are:
>>> import numpy as np >>> def show_top10(classifier, vectorizer, categories): ... feature_names = np.asarray(vectorizer.get_feature_names()) ... for i, category in enumerate(categories): ... top10 = np.argsort(classifier.coef_[i])[-10:] ... print("%s: %s" % (category, " ".join(feature_names[top10]))) ... >>> show_top10(clf, vectorizer, newsgroups_train.target_names) alt.atheism: sgi livesey atheists writes people caltech com god keith edu comp.graphics: organization thanks files subject com image lines university edu graphics sci.space: toronto moon gov com alaska access henry nasa edu space talk.religion.misc: article writes kent people christian jesus sandvik edu com god
You can now see many things that these features have overfit to:
- Almost every group is distinguished by whether headers such as
Distribution:appear more or less often.
- Another significant feature involves whether the sender is affiliated with a university, as indicated either by their headers or their signature.
- The word “article” is a significant feature, based on how often people quote previous posts like this: “In article [article ID], [name] <[e-mail address]> wrote:”
- Other features match the names and e-mail addresses of particular people who were posting at the time.
With such an abundance of clues that distinguish newsgroups, the classifiers barely have to identify topics from text at all, and they all perform at the same high level.
For this reason, the functions that load 20 Newsgroups data provide a
parameter called remove, telling it what kinds of information to strip out
of each file. remove should be a tuple containing any subset of
('headers', 'footers', 'quotes'), telling it to remove headers, signature
blocks, and quotation blocks respectively.
>>> newsgroups_test = fetch_20newsgroups(subset='test', ... remove=('headers', 'footers', 'quotes'), ... categories=categories) >>> vectors_test = vectorizer.transform(newsgroups_test.data) >>> pred = clf.predict(vectors_test) >>> metrics.f1_score(pred, newsgroups_test.target, average='weighted') 0.78409163025839435
This classifier lost over a lot of its F-score, just because we removed metadata that has little to do with topic classification. It loses even more if we also strip this metadata from the training data:
>>> newsgroups_train = fetch_20newsgroups(subset='train', ... remove=('headers', 'footers', 'quotes'), ... categories=categories) >>> vectors = vectorizer.fit_transform(newsgroups_train.data) >>> clf = BernoulliNB(alpha=.01) >>> clf.fit(vectors, newsgroups_train.target) >>> vectors_test = vectorizer.transform(newsgroups_test.data) >>> pred = clf.predict(vectors_test) >>> metrics.f1_score(newsgroups_test.target, pred, average='weighted') 0.73160869205141166
Some other classifiers cope better with this harder version of the task. Try
running Sample pipeline for text feature extraction and evaluation with and without
--filter option to compare the results.
When evaluating text classifiers on the 20 Newsgroups data, you
should strip newsgroup-related metadata. In scikit-learn, you can do this by
remove=('headers', 'footers', 'quotes'). The F-score will be
lower because it is more realistic.
5.8. Downloading datasets from the mldata.org repository¶
sklearn.datasets package is able to directly download data
sets from the repository using the function
For example, to download the MNIST digit recognition database:
>>> from sklearn.datasets import fetch_mldata >>> mnist = fetch_mldata('MNIST original', data_home=custom_data_home)
The MNIST database contains a total of 70000 examples of handwritten digits of size 28x28 pixels, labeled from 0 to 9:
>>> mnist.data.shape (70000, 784) >>> mnist.target.shape (70000,) >>> np.unique(mnist.target) array([ 0., 1., 2., 3., 4., 5., 6., 7., 8., 9.])
After the first download, the dataset is cached locally in the path
specified by the
data_home keyword argument, which defaults to
>>> os.listdir(os.path.join(custom_data_home, 'mldata')) ['mnist-original.mat']
Data sets in mldata.org do not adhere to a strict
naming or formatting convention.
able to make sense of the most common cases, but allows to tailor the
defaults to individual datasets:
The data arrays in mldata.org are most often shaped as
(n_features, n_samples). This is the opposite of the
sklearn.datasets.fetch_mldatatransposes the matrix by default. The
transpose_datakeyword controls this behavior:
>>> iris = fetch_mldata('iris', data_home=custom_data_home) >>> iris.data.shape (150, 4) >>> iris = fetch_mldata('iris', transpose_data=False, ... data_home=custom_data_home) >>> iris.data.shape (4, 150)
For datasets with multiple columns,
sklearn.datasets.fetch_mldatatries to identify the target and data columns and rename them to
data. This is done by looking for arrays named
datain the dataset, and failing that by choosing the first array to be
targetand the second to be
data. This behavior can be changed with the
data_namekeywords, setting them to a specific name or index number (the name and order of the columns in the datasets can be found at its mldata.org under the tab “Data”:
>>> iris2 = fetch_mldata('datasets-UCI iris', target_name=1, data_name=0, ... data_home=custom_data_home) >>> iris3 = fetch_mldata('datasets-UCI iris', target_name='class', ... data_name='double0', data_home=custom_data_home)
5.9. The Labeled Faces in the Wild face recognition dataset¶
This dataset is a collection of JPEG pictures of famous people collected over the internet, all details are available on the official website:
Each picture is centered on a single face. The typical task is called Face Verification: given a pair of two pictures, a binary classifier must predict whether the two images are from the same person.
An alternative task, Face Recognition or Face Identification is: given the picture of the face of an unknown person, identify the name of the person by referring to a gallery of previously seen pictures of identified persons.
Both Face Verification and Face Recognition are tasks that are typically performed on the output of a model trained to perform Face Detection. The most popular model for Face Detection is called Viola-Jones and is implemented in the OpenCV library. The LFW faces were extracted by this face detector from various online websites.
scikit-learn provides two loaders that will automatically download,
cache, parse the metadata files, decode the jpeg and convert the
interesting slices into memmaped numpy arrays. This dataset size is more
than 200 MB. The first load typically takes more than a couple of minutes
to fully decode the relevant part of the JPEG files into numpy arrays. If
the dataset has been loaded once, the following times the loading times
less than 200ms by using a memmaped version memoized on the disk in the
~/scikit_learn_data/lfw_home/ folder using
The first loader is used for the Face Identification task: a multi-class classification task (hence supervised learning):
>>> from sklearn.datasets import fetch_lfw_people >>> lfw_people = fetch_lfw_people(min_faces_per_person=70, resize=0.4) >>> for name in lfw_people.target_names: ... print(name) ... Ariel Sharon Colin Powell Donald Rumsfeld George W Bush Gerhard Schroeder Hugo Chavez Tony Blair
The default slice is a rectangular shape around the face, removing most of the background:
>>> lfw_people.data.dtype dtype('float32') >>> lfw_people.data.shape (1288, 1850) >>> lfw_people.images.shape (1288, 50, 37)
Each of the
1140 faces is assigned to a single person id in the
>>> lfw_people.target.shape (1288,) >>> list(lfw_people.target[:10]) [5, 6, 3, 1, 0, 1, 3, 4, 3, 0]
The second loader is typically used for the face verification task: each sample is a pair of two picture belonging or not to the same person:
>>> from sklearn.datasets import fetch_lfw_pairs >>> lfw_pairs_train = fetch_lfw_pairs(subset='train') >>> list(lfw_pairs_train.target_names) ['Different persons', 'Same person'] >>> lfw_pairs_train.pairs.shape (2200, 2, 62, 47) >>> lfw_pairs_train.data.shape (2200, 5828) >>> lfw_pairs_train.target.shape (2200,)
Both for the
sklearn.datasets.fetch_lfw_pairs function it is
possible to get an additional dimension with the RGB color channels by
color=True, in that case the shape will be
(2200, 2, 62, 47, 3).
sklearn.datasets.fetch_lfw_pairs datasets is subdivided into
3 subsets: the development
train set, the development
test set and
10_folds set meant to compute performance metrics using a
10-folds cross validation scheme.
- Labeled Faces in the Wild: A Database for Studying Face Recognition in Unconstrained Environments. Gary B. Huang, Manu Ramesh, Tamara Berg, and Erik Learned-Miller. University of Massachusetts, Amherst, Technical Report 07-49, October, 2007.
5.10. Forest covertypes¶
The samples in this dataset correspond to 30×30m patches of forest in the US, collected for the task of predicting each patch’s cover type, i.e. the dominant species of tree. There are seven covertypes, making this a multiclass classification problem. Each sample has 54 features, described on the dataset’s homepage. Some of the features are boolean indicators, while others are discrete or continuous measurements.
sklearn.datasets.fetch_covtype will load the covertype dataset;
it returns a dictionary-like object
with the feature matrix in the
and the target values in
The dataset will be downloaded from the web if necessary.
5.11. RCV1 dataset¶
Reuters Corpus Volume I (RCV1) is an archive of over 800,000 manually categorized newswire stories made available by Reuters, Ltd. for research purposes. The dataset is extensively described in .
sklearn.datasets.fetch_rcv1 will load the following version: RCV1-v2, vectors, full sets, topics multilabels:
>>> from sklearn.datasets import fetch_rcv1 >>> rcv1 = fetch_rcv1()
It returns a dictionary-like object, with the following attributes:
The feature matrix is a scipy CSR sparse matrix, with 804414 samples and
47236 features. Non-zero values contains cosine-normalized, log TF-IDF vectors.
A nearly chronological split is proposed in : The first 23149 samples are the training set. The last 781265 samples are the testing set. This follows the official LYRL2004 chronological split.
The array has 0.16% of non zero values:
>>> rcv1.data.shape (804414, 47236)
The target values are stored in a scipy CSR sparse matrix, with 804414 samples and 103 categories. Each sample has a value of 1 in its categories, and 0 in others. The array has 3.15% of non zero values:
>>> rcv1.target.shape (804414, 103)
Each sample can be identified by its ID, ranging (with gaps) from 2286 to 810596:
>>> rcv1.sample_id[:3] array([2286, 2287, 2288], dtype=int32)
The target values are the topics of each sample. Each sample belongs to at least one topic, and to up to 17 topics.
There are 103 topics, each represented by a string. Their corpus frequencies span five orders of magnitude, from 5 occurrences for ‘GMIL’, to 381327 for ‘CCAT’:
>>> rcv1.target_names[:3].tolist() ['E11', 'ECAT', 'M11']
The dataset will be downloaded from the rcv1 homepage if necessary. The compressed size is about 656 MB.