def test(self):
stats = dict.fromkeys(self.classes)
success = 0.0
errors = 0.0
for k in stats:
stats[k] = {'tp': 0, 'fp': 0, 'fn': 0, 'tn': 0}
start = time.time()
for doc in self.testing_docs.find():
if not database.doc_has_data(doc):
continue
cls = self.classify(doc)
if doc['field'] == cls:
success += 1
stats[cls]['tp'] += 1
log.debug('Success (%s/%s)', doc['author'], doc['field'])
else:
errors += 1
# doc['field'] is the correct label, cls is wrong.
# Account a false positive at the correct label: doc['field']
stats[doc['field']]['fp'] += 1
# Account a false negative: cls
stats[cls]['fn'] += 1
log.debug('Error (%s/%s)', doc['author'], doc['field'])
log.info("Took %.3f secs to classify", time.time() - start)
info_stats = {}
for cls, stat in stats.iteritems():
log.info('Class %s (tp=%d,fp=%d,fn=%d,tn=%d)', cls, stat['tp'],
stat['fp'], stat['fn'], stat['tn'])
precision = stat['tp'] / float(stat['tp'] + stat['fp'])
try:
recall = stat['tp'] / float(stat['tp'] + stat['fn'])
f1 = 2.0 * precision * recall / (precision + recall)
except ZeroDivisionError:
f1 = recall = 'nan'
log.debug('\tPrecision=%.3f, Recall=%.3f, F1=%.3f', precision,
recall, f1)
info_stats[cls] = {}
info_stats[cls]['precision'] = precision
info_stats[cls]['recall'] = recall
info_stats[cls]['f1'] = f1
info_stats[cls]['features'] = len(self.features)
global_accuracy = success / float(success + errors)
log.info('Global accuracy: %.2f %%', global_accuracy * 100.0)
info_stats['global_accuracy'] = global_accuracy
return info_stats
def test(self):
stats = dict.fromkeys(self.classes)
success = 0.0
errors = 0.0
for k in stats:
stats[k] = {'tp': 0 , 'fp': 0, 'fn': 0, 'tn': 0}
start = time.time()
for doc in self.testing_docs.find():
if not database.doc_has_data(doc):
continue
cls = self.classify(doc)
if doc['field'] == cls:
success += 1
stats[cls]['tp'] += 1
log.debug('Success (%s/%s)', doc['author'], doc['field'])
else:
errors += 1
# doc['field'] is the correct label, cls is wrong.
# Account a false positive at the correct label: doc['field']
stats[doc['field']]['fp'] += 1
# Account a false negative: cls
stats[cls]['fn'] += 1
log.debug('Error (%s/%s)', doc['author'], doc['field'])
log.info("Took %.3f secs to classify", time.time() - start)
info_stats = {}
for cls, stat in stats.iteritems():
log.info('Class %s (tp=%d,fp=%d,fn=%d,tn=%d)', cls, stat['tp'], stat['fp'], stat['fn'], stat['tn'])
precision = stat['tp']/float(stat['tp'] + stat['fp'])
try:
recall = stat['tp']/float(stat['tp'] + stat['fn'])
f1 = 2.0 * precision * recall / (precision + recall)
except ZeroDivisionError:
f1 = recall = 'nan'
log.debug('\tPrecision=%.3f, Recall=%.3f, F1=%.3f', precision, recall, f1)
info_stats[cls] = {}
info_stats[cls]['precision'] = precision
info_stats[cls]['recall'] = recall
info_stats[cls]['f1'] = f1
info_stats[cls]['features'] = len(self.features)
global_accuracy = success/float(success + errors)
log.info('Global accuracy: %.2f %%', global_accuracy * 100.0)
info_stats['global_accuracy'] = global_accuracy
return info_stats
def train(self):
""" Performs a naive-bayes on the given features. Fills out
dictionaries prior and cond with prior probabilities
P(c) of a class 'c' and conditional P(term|c).
"""
docs = self.training_docs
n = docs.count()
prior = {}
cond = {}
memo = {}
log.debug('starting training on %d documents..', n)
start_time = time.time()
for cls in self.classes:
# Compute prior probabilities
nc = docs.find({'field': cls}).count()
# Maximum Likelihood Estimate (MLE)
prior[cls] = numpy.log(nc / float(n))
# Join all documents for faster counting
textfile = StringIO.StringIO()
for doc in docs.find({'field': cls}):
if not database.doc_has_data(doc):
continue
textfile.write((' '.join(tokenize(doc['data'])) + ' ').encode(
'utf-8', 'replace'))
# Count vocabulary occurences on joined documents
nterm = dict.fromkeys(self.features, 0)
for term in textfile.getvalue().split():
if term in self.features:
nterm[term] += 1
# Precompute denominator for conditional prob. estimator
base = float(sum(nterm.values()) + len(nterm))
# Compute conditional probabilities
for term in self.features:
if term not in cond:
cond[term] = {}
val = (nterm[term] + 1) / base
if val not in memo:
memo[val] = numpy.log(val)
cond[term][cls] = memo[val]
log.info('finished training (took %.3f secs)',
time.time() - start_time)
self.prior = prior
self.cond = cond
def train(self):
""" Performs a naive-bayes on the given features. Fills out
dictionaries prior and cond with prior probabilities
P(c) of a class 'c' and conditional P(term|c).
"""
docs = self.training_docs
n = docs.count()
prior = {}
cond = {}
memo = {}
log.debug('starting training on %d documents..', n)
start_time = time.time()
for cls in self.classes:
# Compute prior probabilities
nc = docs.find({'field': cls}).count()
# Maximum Likelihood Estimate (MLE)
prior[cls] = numpy.log(nc/float(n))
# Join all documents for faster counting
textfile = StringIO.StringIO()
for doc in docs.find({'field': cls}):
if not database.doc_has_data(doc):
continue
textfile.write((' '.join(tokenize(doc['data'])) + ' ').encode('utf-8', 'replace'))
# Count vocabulary occurences on joined documents
nterm = dict.fromkeys(self.features, 0)
for term in textfile.getvalue().split():
if term in self.features:
nterm[term] += 1
# Precompute denominator for conditional prob. estimator
base = float(sum(nterm.values()) + len(nterm))
# Compute conditional probabilities
for term in self.features:
if term not in cond:
cond[term] = {}
val = (nterm[term] + 1)/base
if val not in memo:
memo[val] = numpy.log(val)
cond[term][cls] = memo[val]
log.info('finished training (took %.3f secs)', time.time() - start_time)
self.prior = prior
self.cond = cond
def select_features(self):
""" Select most frequent terms.
"""
start_time = time.time()
log.info('selecting features..')
f = {}
for doc in self.training_docs.find():
if not database.doc_has_data(doc):
continue
for term in bag.bag_of_words(doc['data']):
f.setdefault(term, 1)
f[term] += 1
cut = (numpy.max(f.values()) + numpy.mean(f.values()))/self.size_divider
higher = dict(filter(lambda n: n[1] >= cut, f.iteritems()))
log.info('selected %d terms (took %.3f secs)', len(higher), time.time() - start_time)
return higher
def select_features(self):
""" Select most frequent terms.
"""
start_time = time.time()
log.info('selecting features..')
f = {}
for doc in self.training_docs.find():
if not database.doc_has_data(doc):
continue
for term in bag.bag_of_words(doc['data']):
f.setdefault(term, 1)
f[term] += 1
cut = (numpy.max(f.values()) + numpy.mean(f.values())) / 8.0
higher = dict(filter(lambda n: n[1] >= cut, f.iteritems()))
log.info('selected %d terms (took %.3f secs)', len(higher),
time.time() - start_time)
return higher