Attempting to fix slow NaiveBayes

[jcalbert]

Three changes:

1. basic_extractor can accept a list of strings as well as a list of
   ('word','label') tuples.

2. BaseClassifier now has an instance variable _word_set which is a set
   of tokens seen by the classifier.

1+2) BaseClassifier.extract_features passes _word_set to extractor
rather than the training set.

3. NLTKClassifier.update adds new words to the _word_set.

[jcalbert]

As mentioned in a few issues, NaiveBayes is slow.

Test script, using the movie_reviews dataset:

import os, sys
from random import shuffle
pos_train = [(open('../movie_reviews/pos/'+fname,'r').read(),'pos') for fname in os.listdir('../movie_reviews/pos/')]
neg_train = [(open('../movie_reviews/neg/'+fname,'r').read(),'neg') for fname in os.listdir('../movie_reviews/neg/')]
all_train = pos_train + neg_train
shuffle(all_train)

#Old version
from textblob.classifiers_OLD import NaiveBayesClassifier as NBC

counts = [2**j for j in range(1,7)]
print 'Baseline: '
baseline_res = []
for n in counts:
    t = %timeit -o -q -n 1 -r 1 classifier = NBC(all_train[:n])
    s = str(n)
    print s + ' reviews' +  ' '*(5-len(s)) + str(round(t.best,3)) + ' sec'
    baseline_res.append(t.best)

from textblob.classifiers import NaiveBayesClassifier as NBC
print 'Modified: '
modified_res = []
for n in counts:
    t = %timeit -o -q -n 1 -r 1 classifier = NBC(all_train[:n])
    s = str(n)
    print s + ' reviews' +  ' '*(5-len(s)) + str(round(t.best,3)) + ' sec'
    modified_res.append(t.best)

with output:

Baseline: 
2 reviews    0.05 sec
4 reviews    0.137 sec
8 reviews    0.569 sec
16 reviews   2.57 sec
32 reviews   10.669 sec
64 reviews   42.275 sec

Modified: 
2 reviews    0.032 sec
4 reviews    0.056 sec
8 reviews    0.137 sec
16 reviews   0.338 sec
32 reviews   0.759 sec
64 reviews   1.803 sec

[jcalbert]

Relates to #63, #77 and #123.

[jcalbert]

From the travis-ci failures, it looks like the docstrings on _get_words_from_dataset() and basic_extractor() ask for more restricted training data that test_classifiers.py provides (many are not iterable). Separate issue?

[sloria]

Thanks @jcalbert for the PR! I'll take a look at this over the weekend.

[jcalbert]

Chased down those old bugs. TravisCI is complaining about a translation error, but I think that's Google's problem. Added that issue here: #161.

[iepathos]

Line 83 isn't quite python 3 compatible. Needs to use next() to work with python 3. I think github made it just bold so in case it isn't clear, its double underscore on both sides of next needed for Python 3 like you would see around an init function.

[komuher]

So guys?

[jcalbert]

As I mentioned above, the failed CI build is because of #161. #162 fixes that. Once it is merged then this (#136), #163, #167, and #170 should all pass CI after a rebase.

[sloria]

Thanks for the PR! And sorry for the delayed response.