def main(
corpus_path, project_name, features,
max_authors=0, external_corpus_path=None, max_iterations=100, explain=False):
"""
:param corpus_path: path to a dictionary with xml files
:param project_name: unique project code
:param features: list of feature names
:param max_authors: max. number of candidates
:param external_corpus_path: path to another corpus with documents used for feature tuning
:param max_iterations: max. number of a test repetition
:param explain: write to stdout some explain notes
:return: 2-tuple dictionary stats and int number of candidates
"""
total_stats = []
docs = [doc for doc in get_documents(corpus_path) if doc['author'] in ('asekol', 'resolar', 'independent')]
features_prepared = False
prepare = docs[:]
stats = defaultdict(dict)
f = features[:]
for ha in range(10):
features = random.sample(f, 5)
prepare_features(features, prepare, project_name)
for i, doc in enumerate(docs):
train = docs[:i] + docs[i + 1:]
test = [doc]
ml = UniversalModel(seed=ha, mode="one_layer")
x_train = get_features(train, project_name, features)[0]
y_train = [doc["author"] for doc in train]
x_test = get_features(test, project_name, features)[0]
y_test = [doc["author"] for doc in test]
ml.train(x_train, y_train)
for pos, (y, predictions) in enumerate(six.moves.zip(y_test, ml.test(x_test))):
winner = max(predictions.values())
for key, val in predictions.iteritems():
if val == winner:
guessed = key
stats[y].setdefault(guessed, 0)
stats[y][guessed] += 1
print(stats)
def main(
corpus_path, project_name, features,
max_authors=0, filter_authors_with_less_than_docs=10, docs_per_author_used=10,
external_corpus_path=None, max_iterations=100, explain=False):
"""
:param corpus_path: path to a dictionary with xml files
:param project_name: unique project code
:param features: list of feature names
:param max_authors: max. number of candidates
:param filter_authors_with_less_than_docs: ignore authors with less than X documents
:param docs_per_author_used: use only X documents from author for both training and testing
:param external_corpus_path: path to another corpus with documents used for feature tuning
:param max_iterations: max. number of a test repetition
:param explain: write to stdout some explain notes
:return: 2-tuple dictionary stats and int number of candidates
"""
total_stats = []
docs = [doc for doc in get_documents(corpus_path)]
if external_corpus_path:
external_documents = [doc for doc in get_documents(external_corpus_path)]
else:
external_documents = None
features_prepared = False
run_iterations = 0
train, test = np.array([]), np.array([])
while run_iterations < max_iterations:
for iteration, (prepare, train, test) in enumerate(
divide_documents(
docs, max_authors=max_authors, external_documents=external_documents,
min_per_author_abs=filter_authors_with_less_than_docs, doc_per_author=docs_per_author_used)):
if not features_prepared:
prepare_features(features, prepare, project_name)
features_prepared = True
ml = UniversalModel(seed=run_iterations, mode="one_layer")
x_train = get_features(train, project_name, features)[0]
y_train = [doc["author"] for doc in train]
x_test = get_features(test, project_name, features)[0]
y_test = [doc["author"] for doc in test]
ml.train(x_train, y_train)
iteration_stats = {True: 0, False: 0}
for pos, (y, predictions) in enumerate(six.moves.zip(y_test, ml.test(x_test))):
instance_result = predictions.get(y, 0.0) >= max(predictions.values())
iteration_stats[instance_result] += 1
if explain > 0 and run_iterations == 0:
if not instance_result:
explain -= 1
for key in ('author', 'text', 'morphology'):
print('%s: %s' % (key, test[pos][key]))
total_stats.append(iteration_stats)
run_iterations += 1
logger.info('Iteration %i done', run_iterations)
if iteration >= max_iterations:
break
task_size = len(author_stats(train + test))
accuracies = np.array([100.0 * stats[True] / (stats[True] + stats[False]) for stats in total_stats])
avg = np.average(accuracies)
delta = math.sqrt(np.average(np.square(accuracies - avg)))
# second round, remove irregular values
accuracies = [acc for acc in accuracies if acc <= avg + delta * 2]
avg = np.average(accuracies)
delta = math.sqrt(np.average(np.square(accuracies - avg)))
return {"accuracy": avg, "delta": delta, "task_size": task_size, "iterations": len(total_stats)}
