""" ====================================================== Out-of-core classification of text documents ====================================================== This is an example showing how scikit-learn can be used for classification using an out-of-core approach. Out-of-core learning means that we can learn from data that would not fit into the computer main memory. To achieve this goal we make use of an online classifier (i.e. that supports the `partial_fit` method) that will be fed with batches of examples. Moreover, to guarantee that the features space remains the same over time we leverage the `HashingVectorizer` class that will project each example into the same input space. This is especially useful in the case of text classification where new features (e.g. words) are discovered on the fly. The dataset used in this example is Reuters-21578 as provided by the UCI ML repository. It will be automatically downloaded and uncompressed in the current directory on first run. The plot represents is the learning curve of the classifier i.e. the evolution of classification accuracy with the number of mini-batches fed to the classifier. `ReutersParser` and `ReutersStreamReader` classes are utility classes to parse and stream examples to the main learning loop. To limit the amount of consumed memory at any time we enqueue examples up to a fixed amount before calling the features transformation and learning routines. We then clear the examples queue and proceed with enqueuing again and so on. To study the performance of the method we sample the first 1000 examples of the dataset and hold them out as separate testing data. We then use it to estimate accuracy after each mini-batch. """ # Author: Eustache Diemert # License: BSD 3 clause from __future__ import print_function import time import re import os.path import fnmatch import sgmllib import urllib import tarfile import numpy as np import pylab as pl from sklearn.feature_extraction.text import HashingVectorizer from sklearn.linear_model.stochastic_gradient import SGDClassifier import itertools ############################################################################### # Reuters Dataset related routines ############################################################################### def _not_in_sphinx(): # Hack to detect whether we are running by the sphinx builder return '__file__' in globals() class ReutersParser(sgmllib.SGMLParser): """Utility class to parse a SGML file and yield documents one at a time.""" def __init__(self, verbose=0): sgmllib.SGMLParser.__init__(self, verbose) self._reset() def _reset(self): self.in_title = 0 self.in_body = 0 self.in_topics = 0 self.in_topic_d = 0 self.title = "" self.body = "" self.topics = [] self.topic_d = "" def parse(self, fd): self.docs = [] for chunk in fd: self.feed(chunk) for doc in self.docs: yield doc self.docs = [] self.close() def handle_data(self, data): if self.in_body: self.body += data elif self.in_title: self.title += data elif self.in_topic_d: self.topic_d += data def start_reuters(self, attributes): pass def end_reuters(self): self.body = re.sub(r'\s+', r' ', self.body) self.docs.append({'title': self.title, 'body': self.body, 'topics': self.topics}) self._reset() def start_title(self, attributes): self.in_title = 1 def end_title(self): self.in_title = 0 def start_body(self, attributes): self.in_body = 1 def end_body(self): self.in_body = 0 def start_topics(self, attributes): self.in_topics = 1 def end_topics(self): self.in_topics = 0 def start_d(self, attributes): self.in_topic_d = 1 def end_d(self): self.in_topic_d = 0 self.topics.append(self.topic_d) self.topic_d = "" class ReutersStreamReader(): """Iterate over documents of the Reuters dataset. The Reuters archive will automatically be downloaded and uncompressed if the `data_path` directory does not exist. Documents are represented as dictionaries with 'body' (str), 'title' (str), 'topics' (list(str)) keys. """ DOWNLOAD_URL = ('http://archive.ics.uci.edu/ml/machine-learning-databases/' 'reuters21578-mld/reuters21578.tar.gz') ARCHIVE_FILENAME = 'reuters21578.tar.gz' def __init__(self, data_path): self.data_path = data_path if not os.path.exists(self.data_path): self.download_dataset() def download_dataset(self): """Download the dataset.""" print("downloading dataset (once and for all) into %s" % self.data_path) os.mkdir(self.data_path) def progress(blocknum, bs, size): total_sz_mb = '%.2f MB' % (size / 1e6) current_sz_mb = '%.2f MB' % ((blocknum * bs) / 1e6) if _not_in_sphinx(): print('\rdownloaded %s / %s' % (current_sz_mb, total_sz_mb), end='') urllib.urlretrieve(self.DOWNLOAD_URL, filename=os.path.join(self.data_path, self.ARCHIVE_FILENAME), reporthook=progress) if _not_in_sphinx(): print('\r', end='') print("untaring data ...") tfile = tarfile.open(os.path.join(self.data_path, self.ARCHIVE_FILENAME), 'r:gz') tfile.extractall(self.data_path) print("done !") def iterdocs(self): """Iterate doc by doc, yield a dict.""" for root, _dirnames, filenames in os.walk(self.data_path): for filename in fnmatch.filter(filenames, '*.sgm'): path = os.path.join(root, filename) parser = ReutersParser() for doc in parser.parse(open(path)): yield doc ############################################################################### # Main ############################################################################### # Create the hasher and limit the number of features to a reasonable maximum hasher = HashingVectorizer(decode_error='ignore', n_features=2 ** 18) # Create an online classifier i.e. supporting `partial_fit()` classifier = SGDClassifier() # Create the data_streamer that parses Reuters SGML files and iterates on # documents as a stream data_streamer = ReutersStreamReader('reuters').iterdocs() # Here we propose to learn a binary classification between the positive class # and all other documents.""" all_classes = np.array([0, 1]) # NB: the 'acq' class was chosen as it is more or less evenly distributed in # the Reuters files. For other datasets, one should take care of creating a # test set with a realistic portion of positive instances. positive_class = 'acq' def get_minibatch(doc_iter, size, transformer=hasher, pos_class=positive_class): """Extract a minibatch of examples, return a tuple X, y. Note: size is before excluding invalid docs with no topics assigned. """ data = [('{title}\n\n{body}'.format(**doc), pos_class in doc['topics']) for doc in itertools.islice(doc_iter, size) if doc['topics']] if not len(data): return np.asarray([], dtype=int), np.asarray([], dtype=int) X, y = zip(*data) return transformer.transform(X), np.asarray(y, dtype=int) def iter_minibatchs(doc_iter, minibatch_size): """Generator of minibatchs.""" X, y = get_minibatch(doc_iter, minibatch_size) while X.shape[0]: yield X, y X, y = get_minibatch(doc_iter, minibatch_size) # structure to track accuracy history stats = {'n_train': 0, 'n_test': 0, 'n_train_pos': 0, 'n_test_pos': 0, 'accuracy': 0.0, 'accuracy_history': [(0, 0)], 't0': time.time(), 'runtime_history': [(0, 0)]} # First we hold out a number of examples to estimate accuracy n_test_documents = 1000 X_test, y_test = get_minibatch(data_streamer, 1000) stats['n_test'] += len(y_test) stats['n_test_pos'] += sum(y_test) print("Test set is %d documents (%d positive)" % (len(y_test), sum(y_test))) def progress(stats): """Report progress information, return a string.""" duration = time.time() - stats['t0'] s = "%(n_train)6d train docs (%(n_train_pos)6d positive) " % stats s += "%(n_test)6d test docs (%(n_test_pos)6d positive) " % stats s += "accuracy: %(accuracy).3f " % stats s += "in %.2fs (%5d docs/s)" % (duration, stats['n_train'] / duration) return s # We will feed the classifier with mini-batches of 100 documents; this means # we have at most 100 docs in memory at any time. minibatch_size = 100 # Main loop : iterate on mini-batchs of examples minibatch_iterators = iter_minibatchs(data_streamer, minibatch_size) for i, (X_train, y_train) in enumerate(minibatch_iterators): # update estimator with examples in the current mini-batch classifier.partial_fit(X_train, y_train, classes=all_classes) # accumulate test accuracy stats stats['n_train'] += X_train.shape[0] stats['n_train_pos'] += sum(y_train) stats['accuracy'] = classifier.score(X_test, y_test) stats['accuracy_history'].append((stats['accuracy'], stats['n_train'])) stats['runtime_history'].append((stats['accuracy'], time.time() - stats['t0'])) if i % 10 == 0: print(progress(stats)) ############################################################################### # Plot results ############################################################################### def plot_accuracy(x, y, plot_placement, x_legend): """Plot accuracy as a function of x.""" x = np.array(x) y = np.array(y) pl.subplots_adjust(hspace=0.5) pl.subplot(plot_placement) pl.title('Classification accuracy as a function of %s' % x_legend) pl.xlabel('%s' % x_legend) pl.ylabel('Accuracy') pl.grid(True) pl.plot(x, y) pl.figure(1) # Plot accuracy evolution with #examples accuracy, n_examples = zip(*stats['accuracy_history']) plot_accuracy(n_examples, accuracy, 211, "training examples (#)") # Plot accuracy evolution with runtime accuracy, runtime = zip(*stats['runtime_history']) plot_accuracy(runtime, accuracy, 212, 'runtime (s)') pl.show()