shaypal5 / skift

scikit-learn wrappers for Python fastText.

skift

scikit-learn wrappers for Python fastText.

>>> from skift import FirstColFtClassifier
>>> df = pandas.DataFrame([['woof', 0], ['meow', 1]], columns=['txt', 'lbl'])
>>> sk_clf = FirstColFtClassifier(lr=0.3, epoch=10)
>>> sk_clf.fit(df[['txt']], df['lbl'])
>>> sk_clf.predict([['woof']])
[0]

Installation

Dependencies:

• numpy
• scipy
• scikit-learn
• The fasttext Python package

pip install skift

Configuration

Because fasttext reads input data from files, skift has to dump the input data into temporary files for fasttext to use. A dedicated folder is created for those files on the filesystem. By default, this storage is allocated in the system temporary storage location (i.e. /tmp on *nix systems). To override this default location, use the SKIFT_TEMP_DIR environment variable:

export SKIFT_TEMP_DIR=/path/to/desired/temp/folder

NOTE: The directory will be created if it does not already exist.

Features

• Adheres to the scikit-learn classifier API, including predict_proba.
• Also caters to the common use case of pandas.DataFrame inputs.
• Enables easy stacking of fastText with other types of scikit-learn-compliant classifiers.
• Pickle-able classifier objects.
• Built around the official fasttext Python package.
• Pure python.
• Supports Python 3.5+.
• Fully tested on Linux, OSX and Windows operating systems.

Wrappers

fastText works only on text data, which means that it will only use a single column from a dataset which might contain many feature columns of different types. As such, a common use case is to have the fastText classifier use a single column as input, ignoring other columns. This is especially true when fastText is to be used as one of several classifiers in a stacking classifier, with other classifiers using non-textual features.

skift includes several scikit-learn-compatible wrappers (for the official fastText Python package) which cater to these use cases.

NOTICE: Any additional keyword arguments provided to the classifier constructor, besides those required, will be forwarded to the fastText.train_supervised method on every call to fit.

Standard wrappers

These wrappers do not make additional assumptions on input besides those commonly made by scikit-learn classifies; i.e. that input is a 2d ndarray object and such.

• FirstColFtClassifier - An sklearn classifier adapter for fasttext that takes the first column of input ndarray objects as input.

>>> from skift import FirstColFtClassifier
>>> df = pandas.DataFrame([['woof', 0], ['meow', 1]], columns=['txt', 'lbl'])
>>> sk_clf = FirstColFtClassifier(lr=0.3, epoch=10)
>>> sk_clf.fit(df[['txt']], df['lbl'])
>>> sk_clf.predict([['woof']])
[0]

• IdxBasedFtClassifier - An sklearn classifier adapter for fasttext that takes input by column index. This is set on object construction by providing the input_ix parameter to the constructor.

>>> from skift import IdxBasedFtClassifier
>>> df = pandas.DataFrame([[5, 'woof', 0], [83, 'meow', 1]], columns=['count', 'txt', 'lbl'])
>>> sk_clf = IdxBasedFtClassifier(input_ix=1, lr=0.4, epoch=6)
>>> sk_clf.fit(df[['count', 'txt']], df['lbl'])
>>> sk_clf.predict([['woof']])
[0]

pandas-dependent wrappers

These wrappers assume the X parameter given to fit, predict, and predict_proba methods is a pandas.DataFrame object:

• FirstObjFtClassifier - An sklearn adapter for fasttext using the first column of dtype == object as input.

>>> from skift import FirstObjFtClassifier
>>> df = pandas.DataFrame([['woof', 0], ['meow', 1]], columns=['txt', 'lbl'])
>>> sk_clf = FirstObjFtClassifier(lr=0.2)
>>> sk_clf.fit(df[['txt']], df['lbl'])
>>> sk_clf.predict([['woof']])
[0]

• ColLblBasedFtClassifier - An sklearn adapter for fasttext taking input by column label. This is set on object construction by providing the input_col_lbl parameter to the constructor.

>>> from skift import ColLblBasedFtClassifier
>>> df = pandas.DataFrame([['woof', 0], ['meow', 1]], columns=['txt', 'lbl'])
>>> sk_clf = ColLblBasedFtClassifier(input_col_lbl='txt', epoch=8)
>>> sk_clf.fit(df[['txt']], df['lbl'])
>>> sk_clf.predict([['woof']])
[0]

• SeriesFtClassifier - An sklearn adapter for fasttext taking a Pandas Series as input.

>>> from skift import SeriesFtClassifier
>>> df = pandas.DataFrame([['woof', 0], ['meow', 1]], columns=['txt', 'lbl'])
>>> sk_clf = SeriesFtClassifier(input_col_lbl='txt', epoch=8)
>>> sk_clf.fit(df['txt'], df['lbl'])
>>> sk_clf.predict(['woof'])
>>> sk_clf.predict(df['txt'])

Hyperparameter auto-tuning

It's possible to pass a validation set to fit() in order to optimize the hyper-parameters.

First, to adjust the auto-tune settings, the corresponding keyword arguments can be passed to the constructor (if none are passed the default settings are used):

>>> from skift import SeriesFtClassifier
>>> df_train = pandas.DataFrame([['woof', 0], ['meow', 1]], columns=['txt', 'lbl'])
>>> df_val = pandas.DataFrame([['woof woof', 0], ['meow meow', 1]], columns=['txt', 'lbl'])
>>> sk_clf = SeriesFtClassifier(epoch=8, autotuneDuration=5)

Then, the validation dataframe (or series, in this case, since we constructed a SeriesFtClassifier) and label column should be provided to the fit() method:

>>> sk_clf.fit(df_train['txt'], df_train['lbl'], X_validation=df_val['txt'], y_validation=df_val['lbl'])

Or simply by position:

>>> sk_clf.fit(df_train['txt'], df_train['lbl'], df_val['txt'], df_val['lbl'])

Using Pre-trained word vectors

This is done in the exact same way as with the Python module or the fastText CLI, but not setting the right vector dimensions in the constructor (identical to the dimensions of the pretrained vectors you are using) will crash fastText without explanation, so we provide an example:

from skift import SeriesFtClassifier
ft_clf = SeriesFtClassifier(
    autotuneDuration=900,
    pretrainedVectors='/Users/myuser/data/word_vectors/crawl-300d-2M.vec',
    dim=300,
)

In this case, not providing the constructor with dim=300 would bring about a crash when calling ft_clf.fit().

Contributing

Package author and current maintainer is Shay Palachy ([e-mail]); You are more than welcome to approach him for help. Contributions are very welcomed.

Installing for development

Clone:

git clone [e-mail]:shaypal5/skift.git

Install in development mode, including test dependencies:

cd skift
pip install -e '.[test]'

To also install fasttext, see instructions in the Installation section.

Running the tests

To run the tests use:

cd skift
pytest

Adding documentation

The project is documented using the numpy docstring conventions, which were chosen as they are perhaps the most widely-spread conventions that are both supported by common tools such as Sphinx and result in human-readable docstrings. When documenting code you add to this project, follow these conventions.

Additionally, if you update this README.rst file, use python setup.py checkdocs to validate it compiles.

Credits

Created by Shay Palachy ([e-mail]).

Contributions:

• Dimid Duchovny <https://github.com/dimidd>_ contributed the SeriesFtClassifier class and the hyperparameter auto-tuning capability.

Fixes: uniaz, crouffer, amirzamli and sgt.