The sklearn.feature_extraction.text.CountVectorizer and sklearn.feature_extraction.text.TfidfVectorizer classes suffer from a number of scalability issues that all stem from the internal usage of the vocabulary_ attribute (a Python dictionary) used to map the unicode string feature names to the integer feature indices.
The main scalability issues are:
vocabulary_ would be a shared state: complex synchronization and overheadvocabulary_ needs to be learned from the data: its size cannot be known before making one pass over the full datasetTo better understand the issue let's have a look at how the vocabulary_ attribute work. At fit time the tokens of the corpus are uniquely indentified by a integer index and this mapping stored in the vocabulary:
In [ ]:
from sklearn.feature_extraction.text import CountVectorizer
vectorizer = CountVectorizer(min_df=1)
vectorizer.fit([
"The cat sat on the mat.",
])
vectorizer.vocabulary_
The vocabulary is used at transform time to build the occurrence matrix:
In [ ]:
X = vectorizer.transform([
"The cat sat on the mat.",
"This cat is a nice cat.",
]).toarray()
print(len(vectorizer.vocabulary_))
print(vectorizer.get_feature_names())
print(X)
Let's refit with a slightly larger corpus:
In [ ]:
vectorizer = CountVectorizer(min_df=1)
vectorizer.fit([
"The cat sat on the mat.",
"The quick brown fox jumps over the lazy dog.",
])
vectorizer.vocabulary_
The vocabulary_ is the (logarithmically) growing with the size of the training corpus. Note that we could not have built the vocabularies in parallel on the 2 text documents as they share some words hence would require some kind of shared datastructure or synchronization barrier which is complicated to setup, especially if we want to distribute the processing on a cluster.
With this new vocabulary, the dimensionality of the output space is now larger:
In [ ]:
X = vectorizer.transform([
"The cat sat on the mat.",
"This cat is a nice cat.",
]).toarray()
print(len(vectorizer.vocabulary_))
print(vectorizer.get_feature_names())
print(X)
To illustrate the scalability issues of the vocabulary-based vectorizers, let's load a more realistic dataset for a classical text classification task: sentiment analysis on text documents. The goal is to tell apart negative from positive movie reviews from the Internet Movie Database (IMDb).
In the following sections, with a large subset of movie reviews from the IMDb that has been collected by Maas et al.
This dataset contains 50,000 movie reviews, which were split into 25,000 training samples and 25,000 test samples. The reviews are labeled as either negative (neg) or positive (pos). Moreover, positive means that a movie received >6 stars on IMDb; negative means that a movie received <5 stars, respectively.
Assuming that the ../fetch_data.py script was run successfully the following files should be available:
In [ ]:
import os
train_path = os.path.join('datasets', 'IMDb', 'aclImdb', 'train')
test_path = os.path.join('datasets', 'IMDb', 'aclImdb', 'test')
Now, let's load them into our active session via scikit-learn's load_files function
In [ ]:
from sklearn.datasets import load_files
train = load_files(container_path=(train_path),
categories=['pos', 'neg'])
test = load_files(container_path=(test_path),
categories=['pos', 'neg'])
The load_files function loaded the datasets into sklearn.datasets.base.Bunch objects, which are Python dictionaries:
In [ ]:
train.keys()
In particular, we are only interested in the data and target arrays.
In [ ]:
import numpy as np
for label, data in zip(('TRAINING', 'TEST'), (train, test)):
print('\n\n%s' % label)
print('Number of documents:', len(data['data']))
print('\n1st document:\n', data['data'][0])
print('\n1st label:', data['target'][0])
print('\nClass names:', data['target_names'])
print('Class count:',
np.unique(data['target']), ' -> ',
np.bincount(data['target']))
As we can see above the 'target' array consists of integers 0 and 1, where 0 stands for negative and 1 stands for positive.
Remember the bag of word representation using a vocabulary based vectorizer:
To workaround the limitations of the vocabulary-based vectorizers, one can use the hashing trick. Instead of building and storing an explicit mapping from the feature names to the feature indices in a Python dict, we can just use a hash function and a modulus operation:
More info and reference for the original papers on the Hashing Trick in the following site as well as a description specific to language here.
In [ ]:
from sklearn.utils.murmurhash import murmurhash3_bytes_u32
# encode for python 3 compatibility
for word in "the cat sat on the mat".encode("utf-8").split():
print("{0} => {1}".format(
word, murmurhash3_bytes_u32(word, 0) % 2 ** 20))
This mapping is completely stateless and the dimensionality of the output space is explicitly fixed in advance (here we use a modulo 2 ** 20 which means roughly 1M dimensions). The makes it possible to workaround the limitations of the vocabulary based vectorizer both for parallelizability and online / out-of-core learning.
The HashingVectorizer class is an alternative to the CountVectorizer (or TfidfVectorizer class with use_idf=False) that internally uses the murmurhash hash function:
In [ ]:
from sklearn.feature_extraction.text import HashingVectorizer
h_vectorizer = HashingVectorizer(encoding='latin-1')
h_vectorizer
It shares the same "preprocessor", "tokenizer" and "analyzer" infrastructure:
In [ ]:
analyzer = h_vectorizer.build_analyzer()
analyzer('This is a test sentence.')
We can vectorize our datasets into a scipy sparse matrix exactly as we would have done with the CountVectorizer or TfidfVectorizer, except that we can directly call the transform method: there is no need to fit as HashingVectorizer is a stateless transformer:
In [ ]:
docs_train, y_train = train['data'], train['target']
docs_valid, y_valid = test['data'][:12500], test['target'][:12500]
docs_test, y_test = test['data'][12500:], test['target'][12500:]
The dimension of the output is fixed ahead of time to n_features=2 ** 20 by default (nearly 1M features) to minimize the rate of collision on most classification problem while having reasonably sized linear models (1M weights in the coef_ attribute):
In [ ]:
h_vectorizer.transform(docs_train)
Now, let's compare the computational efficiency of the HashingVectorizer to the CountVectorizer:
In [ ]:
h_vec = HashingVectorizer(encoding='latin-1')
%timeit -n 1 -r 3 h_vec.fit(docs_train, y_train)
In [ ]:
count_vec = CountVectorizer(encoding='latin-1')
%timeit -n 1 -r 3 count_vec.fit(docs_train, y_train)
As we can see, the HashingVectorizer is much faster than the Countvectorizer in this case.
Finally, let us train a LogisticRegression classifier on the IMDb training subset:
In [ ]:
from sklearn.linear_model import LogisticRegression
from sklearn.pipeline import Pipeline
h_pipeline = Pipeline([
('vec', HashingVectorizer(encoding='latin-1')),
('clf', LogisticRegression(random_state=1)),
])
h_pipeline.fit(docs_train, y_train)
In [ ]:
print('Train accuracy', h_pipeline.score(docs_train, y_train))
print('Validation accuracy', h_pipeline.score(docs_valid, y_valid))
In [ ]:
import gc
del count_vec
del h_pipeline
gc.collect()
Out-of-Core learning is the task of training a machine learning model on a dataset that does not fit into memory or RAM. This requires the following conditions:
partial_fit method in scikit-learn).In the following sections, we will set up a simple batch-training function to train an SGDClassifier iteratively.
But first, let us load the file names into a Python list:
In [ ]:
train_path = os.path.join('datasets', 'IMDb', 'aclImdb', 'train')
train_pos = os.path.join(train_path, 'pos')
train_neg = os.path.join(train_path, 'neg')
fnames = [os.path.join(train_pos, f) for f in os.listdir(train_pos)] +\
[os.path.join(train_neg, f) for f in os.listdir(train_neg)]
fnames[:3]
Next, let us create the target label array:
In [ ]:
y_train = np.zeros((len(fnames), ), dtype=int)
y_train[:12500] = 1
np.bincount(y_train)
Now, we implement the batch_train function as follows:
In [ ]:
from sklearn.base import clone
def batch_train(clf, fnames, labels, iterations=25, batchsize=1000, random_seed=1):
vec = HashingVectorizer(encoding='latin-1')
idx = np.arange(labels.shape[0])
c_clf = clone(clf)
rng = np.random.RandomState(seed=random_seed)
for i in range(iterations):
rnd_idx = rng.choice(idx, size=batchsize)
documents = []
for i in rnd_idx:
with open(fnames[i], 'r', encoding='latin-1') as f:
documents.append(f.read())
X_batch = vec.transform(documents)
batch_labels = labels[rnd_idx]
c_clf.partial_fit(X=X_batch,
y=batch_labels,
classes=[0, 1])
return c_clf
Note that we are not using LogisticRegression as in the previous section, but we will use a SGDClassifier with a logistic cost function instead. SGD stands for stochastic gradient descent, an optimization alrogithm that optimizes the weight coefficients iteratively sample by sample, which allows us to feed the data to the classifier chunk by chuck.
And we train the SGDClassifier; using the default settings of the batch_train function, it will train the classifier on 25*1000=25000 documents. (Depending on your machine, this may take >2 min)
In [ ]:
from sklearn.linear_model import SGDClassifier
sgd = SGDClassifier(loss='log', random_state=1)
sgd = batch_train(clf=sgd,
fnames=fnames,
labels=y_train)
Eventually, let us evaluate its performance:
In [ ]:
vec = HashingVectorizer(encoding='latin-1')
sgd.score(vec.transform(docs_test), y_test)
Using the Hashing Vectorizer makes it possible to implement streaming and parallel text classification but can also introduce some issues:
HashingVectorizer does not provide "Inverse Document Frequency" reweighting (lack of a use_idf=True option).The collision issues can be controlled by increasing the n_features parameters.
The IDF weighting might be reintroduced by appending a TfidfTransformer instance on the output of the vectorizer. However computing the idf_ statistic used for the feature reweighting will require to do at least one additional pass over the training set before being able to start training the classifier: this breaks the online learning scheme.
The lack of inverse mapping (the get_feature_names() method of TfidfVectorizer) is even harder to workaround. That would require extending the HashingVectorizer class to add a "trace" mode to record the mapping of the most important features to provide statistical debugging information.
In the mean time to debug feature extraction issues, it is recommended to use TfidfVectorizer(use_idf=False) on a small-ish subset of the dataset to simulate a HashingVectorizer() instance that have the get_feature_names() method and no collision issues.
In [ ]:
# %load solutions/23_batchtrain.py