def _test_neural_network_methods (ut) :
# Neural network connection weights are initiated pseudorandomly, this introduces determinism to the randomness.
random.seed(0)
nn = NeuralNetwork('abc', 3, 'def', name='foo')
training_patterns = [('ab', 'f'),
('ac', 'f'),
('b', 'd'),
('c', 'e')]
error = round(sum(nn.train(inputs, outputs) for _ in range(20) for inputs, outputs in training_patterns), 5)
ut.assert_equal(error, 10.51832)
correct_outputs = [('f', 'd', 'e'),
('d', 'e', 'f'),
('e', 'd', 'f'),
('f', 'd', 'e'),
('f', 'e', 'd'),
('d', 'e', 'f')]
for inputs, correct in zip(['a', 'b', 'c', 'ab', 'ac', 'bc'], correct_outputs) :
ut.assert_equal(tuple(object for object, score in sorted(nn.predict(inputs), reverse=True)), correct)
def tests (nn) :
ut.assert_true('a' in nn)
ut.assert_true('d' in nn)
ut.assert_true('z' not in nn)
ut.assert_true(173 not in nn)
ut.assert_equal(list(nn.inputs()), list('abc'))
ut.assert_equal(list(nn.outputs()), list('def'))
ut.assert_equal(list(nn), list('abcdef'))
ut.assert_equal([(object, '%.5f' % score) for object, score in nn.scores()],
[('d', '0.68689'), ('e', '0.60615'), ('f', '-0.05136')])
tests(nn)
db = Database()
with db as session :
session.add(nn)
tests(nn)
with db as session :
tests(session.access.nn('foo'))
tests(nn)
def __demo__ (ut) :
from nlplib.data import builtin_db
from nlplib.core.model import Word, NeuralNetwork
from nlplib.exterior.train import usable
from nlplib.exterior.util import plot
top = 10
with builtin_db() as session :
top_words = list(session.access.most_common(Word, top=top)) + [None]
nn = NeuralNetwork(top_words, int(len(top_words)*2), top_words)
with builtin_db() as session :
patterns = usable(nn, list(session.access.all_documents())[:], gram_size=2)
errors = [nn.train(input_words, output_words, rate=0.1) for input_words, output_words in patterns]
from nlplib.general.math import avg
print(avg(errors[:]))
plot(errors,
sample_size=100)
def ask (string) :
with builtin_db() as session :
words = session.access.words(string)
scored = Scored(nn.predict(words)).sorted()
print('\n'.join(repr(score) for score in list(scored)[:12]))
return ask
