Skip to main content
Help us improve PyPI by participating in user testing. All experience levels needed!

Well, that's no ordinary rabbit.

Project description

Rabbit of Caerbannog

Well, that’s no ordinary rabbit - that’s the most foul, cruel, and bad-tempered rodent you ever set eyes on!

—Tim the Enchanter

This module is a high-level interface for the Vowpal Wabbit machine learning system. Currently it relies on the wabbit_wappa module for lower-level interaction, but strives to provide a more high-level object-oriented interface.

There are currently 3 kinds of Rabbit`s you can `import from caerbannog:

Your standard rabbit instance. By default runs Vowpal Wabbit using pipes for stdin/stdout
Runs Vowpal Wabbit in active learning mode, using TCP socket
The initializer expects the argument fp which is an open file with ‘wt’ mode. the inputs fed to teach will be written to this file for offline processing.

Movie Review Sentiments - Active learning demo with caerbannog

Import relevant modules here. We are using the ActiveRabbit for active online learning

from caerbannog import ActiveRabbit, Example
from itertools import islice
import random
import nltk
from nltk.corpus import movie_reviews

Create an active bunny, with active mellowness of 0.01

rabbit = ActiveRabbit(loss_function='logistic', active_mellowness=0.01)

Load the documents from NLTK movie review corpus (note that you need to download these first by For each document, make a tuple (document_words, category) where category is either ‘pos’ or ‘neg’ and document_words is a list of words from tokenizer.

documents = [(list(movie_reviews.words(fileid)), category)
              for category in movie_reviews.categories()
              for fileid in movie_reviews.fileids(category)]

The feature extractor function. First filters out all tokens that are non-alphanumeric. Then make the ‘w’ namespace consist of all the words in the review; ‘n’ consists of 2-4 ngrams of the document words.

def document_features(document_words):
    document_words = list(filter(str.isalnum, document_words))
    example = Example()
    ngrams = set()
    for j in range(2, 5):
        ngrams.update('_'.join(i) for i in nltk.ngrams(document_words, j))

    return example

Vowpal Wabbit expects labels to be -1 and 1 for logistic binary classifier

def convert_sent(sent):
    return {'pos': 1, 'neg': -1}[sent]

Convert the sentiment value and extract features.

examples = [ (convert_sent(sent), document_features(doc)) for (doc, sent) in documents ]

Train with 1500 first examples and keep the remaining ones for verification

teach, test = examples[:1500], examples[1500:]

Teach the filter. We ask for prediction for each example; if the importance is over 1 we “label” the example and teach it to the classifier. We repeat the classification 40 times to ensure that the classifier has had enough to adjust the weights.

taught = 0
predicted = 0
labelled = set()
for i in range(10):
    for sent, ex in teach:
        predicted += 1
        if rabbit.predict(example=ex).importance >= 1:
            rabbit.teach(label=sent, example=ex)
            taught += 1

print("Predicted {}, taught {} (ratio {}). {} unique inputs labelled"
      .format(predicted, taught, taught/predicted, len(labelled)))
Predicted 15000, taught 1057 (ratio 0.07046666666666666). 1042 unique inputs labelled

Test with the testing set. For each correctly labelled example, increase the counter

correct = 0
for sent, ex in test:
    prediction = rabbit.predict(example=ex)
    if prediction.label == sent:
        correct += 1

print("{} inputs predicted. {} correct; ratio {}".format(len(test), correct, correct / len(test)))
500 inputs predicted. 418 correct; ratio 0.836


MIT license.

Project details

Release history Release notifications

This version
History Node


Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Filename, size & hash SHA256 hash help File type Python version Upload date
caerbannog-0.1.tar.gz (5.1 kB) Copy SHA256 hash SHA256 Source None Oct 29, 2014

Supported by

Elastic Elastic Search Pingdom Pingdom Monitoring Google Google BigQuery Sentry Sentry Error logging CloudAMQP CloudAMQP RabbitMQ AWS AWS Cloud computing Fastly Fastly CDN DigiCert DigiCert EV certificate StatusPage StatusPage Status page