README.rst

seqlearn

seqlearn is a sequence classification toolkit for Python. It is designed to extend scikit-learn and offer as similar as possible an API.

Compiling and installing

Get NumPy >=1.6, SciPy >=0.11, Cython >=0.20.2 and a recent version of scikit-learn. Then issue:

python setup.py install

to install seqlearn.

If you want to use seqlearn from its source directory without installing, you have to compile first:

python setup.py build_ext --inplace

Getting started

The easiest way to start using seqlearn is to fetch a dataset in CoNLL 2000 format. Define a task-specific feature extraction function, e.g.:

>>> def features(sequence, i):
...     yield "word=" + sequence[i].lower()
...     if sequence[i].isupper():
...         yield "Uppercase"
...

Load the training file, say train.txt:

>>> from seqlearn.datasets import load_conll
>>> X_train, y_train, lengths_train = load_conll("train.txt", features)

Train a model:

>>> from seqlearn.perceptron import StructuredPerceptron
>>> clf = StructuredPerceptron()
>>> clf.fit(X_train, y_train, lengths_train)

Check how well you did on a validation set, say validation.txt:

>>> X_test, y_test, lengths_test = load_conll("validation.txt", features)
>>> from seqlearn.evaluation import bio_f_score
>>> y_pred = clf.predict(X_test, lengths_test)
>>> print(bio_f_score(y_test, y_pred))

For more information, see the documentation.

Debugging Notes

If you're using the .score() method, note that it might not be evaluating/scoring the right granularity (for example, evaluating letter by letter versus word by word when your model outputs words). Also check out whole_sequence_accuracy() method.