ShallowLearn

A collection of supervised learning models based on shallow neural network approaches (e.g., word2vec and fastText) with some additional exclusive features. Written in Python and fully compatible with scikit-learn.

Discussion group for users and developers: https://groups.google.com/d/forum/shallowlearn

Getting Started

Install the latest version:

pip install cython
pip install shallowlearn

Import models from shallowlearn.models, they implement the standard methods for supervised learning in scikit-learn, e.g., fit(X, y), predict(X), predict_proba(X), etc.

Data is raw text, each sample in the iterable X is a list of tokens (words of a document), while each element in the iterable y (corresponding to an element in X) can be a single label or a list in case of a multi-label training set. Obviously, y must be of the same size of X.

Models

GensimFastText

Choose this model if your goal is classification with fastText! (it is going to be the most stable and rich feature-wise)

A supervised learning model based on the fastText algorithm [1]. The code is mostly taken and rewritten from Gensim, it takes advantage of its optimizations (e.g. Cython) and support.

It is possible to choose the Softmax loss function (default) or one of its two "approximations": Hierarchical Softmax and Negative Sampling.

The parameter bucket configures the feature hashing space, i.e., the hashing trick described in [1]. Using the hashing trick together with partial_fit(X, y) yields a powerful online text classifier (see Online learning).

It is possible to load pre-trained word vectors at initialization, passing a Gensim Word2Vec or a ShallowLearn LabeledWord2Vec instance (the latter is retrievable from a GensimFastText model by the attribute classifier). With method fit_embeddings(X) it is possible to pre-train word vectors, using the current parameter values of the model.

Constructor argument names are a mix between the ones of Gensim and the ones of fastText (see this class docstring).

>>> from shallowlearn.models import GensimFastText
>>> clf = GensimFastText(size=100, min_count=0, loss='hs', iter=3, seed=66)
>>> clf.fit([('i', 'am', 'tall'), ('you', 'are', 'fat')], ['yes', 'no'])
>>> clf.predict([('tall', 'am', 'i')])
['yes']

FastText

The supervised algorithm of fastText implemented in fastText.py , which exposes an interface on the original C++ code. The current advantages of this class over GensimFastText are the subwords and the n-gram features implemented via the hashing trick. The constructor arguments are equivalent to the original supervised model, except for input_file, output and label_prefix.

WARNING: The only way of loading datasets in fastText.py is through the filesystem (as of version 0.8.2), so data passed to fit(X, y) will be written in temporary files on disk.

>>> from shallowlearn.models import FastText
>>> clf = FastText(dim=100, min_count=0, loss='hs', epoch=3, bucket=5, word_ngrams=2)
>>> clf.fit([('i', 'am', 'tall'), ('you', 'are', 'fat')], ['yes', 'no'])
>>> clf.predict([('tall', 'am', 'i')])
['yes']

DeepInverseRegression

TODO: Based on https://radimrehurek.com/gensim/models/word2vec.html#gensim.models.word2vec.Word2Vec.score

DeepAveragingNetworks

TODO: Based on https://github.com/miyyer/dan

Exclusive Features

Next cool features will be listed as Issues in Github, for now:

Persistence

Any model can be serialized and de-serialized with the two methods save and load. They overload the SaveLoad interface of Gensim, so it is possible to control the cost on disk usage of the models, instead of simply pickling the objects. The original interface also allows to use compression on the serialization outputs.

save may create multiple files with names prefixed by the name given to the serialized model.

>>> from shallowlearn.models import GensimFastText
>>> clf = GensimFastText(size=100, min_count=0, loss='hs', iter=3, seed=66)
>>> clf.save('./model')
>>> loaded = GensimFastText.load('./model') # it also creates ./model.CLF

Benchmarks

Text classification

The script scripts/document_classification_20newsgroups.py refers to this scikit-learn example in which text classifiers are compared on a reference dataset; we added our models to the comparison. The current results, even if still preliminary, are comparable with other approaches, achieving the best performance in speed.

Results as of release 0.0.5, with chi2_select option set to 80%. The times take into account of tf-idf vectorization in the “classic” classifiers, and the I/O operations for the training of fastText.py. The evaluation measure is macro F1.

Online learning

The script scripts/plot_out_of_core_classification.py computes a benchmark on some scikit-learn classifiers which are able to learn incrementally, a batch of examples at a time. These classifiers can learn online by using the scikit-learn method partial_fit(X, y). The original example describes the approach through feature hashing, which we set with parameter bucket.

The results are decent but there is room for improvement. We configure our classifier with iter=1, size=100, alpha=0.1, sample=0, min_count=0, so to keep the model fast and small, and to not cut off words from the few samples we have.

References