def test_check_X_negative002(self):
""" If list of texts is specified as the NumPy array, then it must be a 1-D array. """
texts = np.array([[0, 1, 2], [3, 4, 5], [6, 7, 8], [9, 10, 11]])
true_err_msg = re.escape('`X` must be a 1-D array!')
try:
checking_method = self.assertRaisesRegex
except:
checking_method = self.assertRaisesRegexp
with checking_method(ValueError, true_err_msg):
Seq2SeqLSTM.check_X(texts, 'X')
def test_check_X_negative001(self):
""" All texts must be a string and have a `split` method. """
texts = ['123', 4, '567']
true_err_msg = re.escape('Sample 1 of `X` is wrong! This sample have not the `split` method.')
try:
checking_method = self.assertRaisesRegex
except:
checking_method = self.assertRaisesRegexp
with checking_method(ValueError, true_err_msg):
Seq2SeqLSTM.check_X(texts, 'X')
def test_creation(self):
seq2seq = Seq2SeqLSTM(batch_size=256,
epochs=200,
latent_dim=500,
validation_split=0.1,
grad_clipping=50.0,
lr=0.01,
rho=0.8,
epsilon=0.2,
lowercase=False,
verbose=True)
self.assertIsInstance(seq2seq, Seq2SeqLSTM)
self.assertTrue(hasattr(seq2seq, 'batch_size'))
self.assertEqual(seq2seq.batch_size, 256)
self.assertTrue(hasattr(seq2seq, 'epochs'))
self.assertEqual(seq2seq.epochs, 200)
self.assertTrue(hasattr(seq2seq, 'latent_dim'))
self.assertEqual(seq2seq.latent_dim, 500)
self.assertTrue(hasattr(seq2seq, 'validation_split'))
self.assertAlmostEqual(seq2seq.validation_split, 0.1)
self.assertTrue(hasattr(seq2seq, 'grad_clipping'))
self.assertAlmostEqual(seq2seq.grad_clipping, 50.0)
self.assertTrue(hasattr(seq2seq, 'lr'))
self.assertAlmostEqual(seq2seq.lr, 0.01)
self.assertTrue(hasattr(seq2seq, 'rho'))
self.assertAlmostEqual(seq2seq.rho, 0.8)
self.assertTrue(hasattr(seq2seq, 'lowercase'))
self.assertFalse(seq2seq.lowercase)
self.assertTrue(hasattr(seq2seq, 'verbose'))
self.assertTrue(seq2seq.verbose)
def test_serialize_untrained(self):
seq2seq = Seq2SeqLSTM(batch_size=256, epochs=200, latent_dim=500, validation_split=0.1,
grad_clipping=50.0, lr=0.01, weight_decay=0.0001, lowercase=False, verbose=True,
random_state=42)
with open(self.model_name, 'wb') as fp:
pickle.dump(seq2seq, fp)
with open(self.model_name, 'rb') as fp:
another_seq2seq = pickle.load(fp)
self.assertIsInstance(another_seq2seq, Seq2SeqLSTM)
self.assertTrue(hasattr(another_seq2seq, 'batch_size'))
self.assertEqual(another_seq2seq.batch_size, 256)
self.assertTrue(hasattr(another_seq2seq, 'epochs'))
self.assertEqual(another_seq2seq.epochs, 200)
self.assertTrue(hasattr(another_seq2seq, 'latent_dim'))
self.assertEqual(another_seq2seq.latent_dim, 500)
self.assertTrue(hasattr(another_seq2seq, 'validation_split'))
self.assertAlmostEqual(another_seq2seq.validation_split, 0.1)
self.assertTrue(hasattr(another_seq2seq, 'grad_clipping'))
self.assertAlmostEqual(another_seq2seq.grad_clipping, 50.0)
self.assertTrue(hasattr(another_seq2seq, 'lr'))
self.assertAlmostEqual(another_seq2seq.lr, 0.01)
self.assertTrue(hasattr(another_seq2seq, 'weight_decay'))
self.assertAlmostEqual(another_seq2seq.weight_decay, 0.0001)
self.assertTrue(hasattr(another_seq2seq, 'lowercase'))
self.assertFalse(another_seq2seq.lowercase)
self.assertTrue(hasattr(another_seq2seq, 'verbose'))
self.assertTrue(another_seq2seq.verbose)
self.assertTrue(hasattr(another_seq2seq, 'random_state'))
self.assertEqual(another_seq2seq.random_state, 42)
def test_serialize_trained(self):
input_texts, target_texts = self.load_text_pairs(self.data_set_name)
indices = list(range(len(input_texts)))
random.shuffle(indices)
n = int(round(0.2 * len(indices)))
input_texts_for_training = []
target_texts_for_training = []
for ind in indices[:-n]:
input_texts_for_training.append(input_texts[ind])
target_texts_for_training.append(target_texts[ind])
input_texts_for_testing = []
target_texts_for_testing = []
for ind in indices[-n:]:
input_texts_for_testing.append(input_texts[ind])
target_texts_for_testing.append(target_texts[ind])
seq2seq = Seq2SeqLSTM(validation_split=None, epochs=10, lr=1e-3)
seq2seq.fit(input_texts_for_training, target_texts_for_training,
eval_set=(input_texts_for_testing, target_texts_for_testing))
predicted_texts_1 = seq2seq.predict(input_texts_for_testing)
with open(self.model_name, 'wb') as fp:
pickle.dump(seq2seq, fp)
del seq2seq
with open(self.model_name, 'rb') as fp:
another_seq2seq = pickle.load(fp)
predicted_texts_2 = another_seq2seq.predict(input_texts_for_testing)
self.assertEqual(predicted_texts_1, predicted_texts_2)
def test_fit_positive05(self):
""" Prepared evaluation set is used in the early stopping criterion. """
input_texts_for_training, target_texts_for_training = self.load_text_pairs(
self.data_set_name)
seq2seq = Seq2SeqLSTM(validation_split=None, lr=1e-2)
res = seq2seq.fit(input_texts_for_training[:-20],
target_texts_for_training[:-20],
eval_set=(input_texts_for_training[-20:],
target_texts_for_training[-20:]))
self.assertIsInstance(res, Seq2SeqLSTM)
self.assertTrue(hasattr(res, 'input_token_index_'))
self.assertIsInstance(res.input_token_index_, dict)
self.assertTrue(hasattr(res, 'target_token_index_'))
self.assertIsInstance(res.target_token_index_, dict)
self.assertTrue(hasattr(res, 'reverse_target_char_index_'))
self.assertIsInstance(res.reverse_target_char_index_, dict)
self.assertTrue(hasattr(res, 'max_encoder_seq_length_'))
self.assertIsInstance(res.max_encoder_seq_length_, int)
self.assertGreater(res.max_encoder_seq_length_, 0)
self.assertTrue(hasattr(res, 'max_decoder_seq_length_'))
self.assertIsInstance(res.max_decoder_seq_length_, int)
self.assertGreater(res.max_decoder_seq_length_, 0)
self.assertTrue(hasattr(res, 'encoder_model_'))
self.assertIsInstance(res.encoder_model_, Model)
self.assertTrue(hasattr(res, 'decoder_model_'))
self.assertIsInstance(res.decoder_model_, Model)
def test_fit_negative04(self):
""" Some parameter of the `Seq2SeqLSTM` object is wrong. """
input_texts_for_training, target_texts_for_training = self.load_text_pairs(self.data_set_name)
seq2seq = Seq2SeqLSTM(batch_size=0)
true_err_msg = re.escape('`batch_size` must be a positive number! 0 is not positive.')
try:
checking_method = self.assertRaisesRegex
except:
checking_method = self.assertRaisesRegexp
with checking_method(ValueError, true_err_msg):
seq2seq.fit(input_texts_for_training, target_texts_for_training)
def test_fit_negative03(self):
""" Number of input texts does not equal to number of target texts. """
input_texts_for_training, target_texts_for_training = self.load_text_pairs(self.data_set_name)
seq2seq = Seq2SeqLSTM()
true_err_msg = re.escape(f'`X` does not correspond to `y`! {len(input_texts_for_training)} != {len(target_texts_for_training) - 1}.')
try:
checking_method = self.assertRaisesRegex
except:
checking_method = self.assertRaisesRegexp
with checking_method(ValueError, true_err_msg):
seq2seq.fit(input_texts_for_training, target_texts_for_training[:-1])
def test_fit_negative02(self):
""" Object with target texts is not one of the basic sequence types. """
input_texts_for_training, target_texts_for_training = self.load_text_pairs(self.data_set_name)
seq2seq = Seq2SeqLSTM()
true_err_msg = re.escape(f'`{type({1, 2})}` is wrong type for `y`.')
try:
checking_method = self.assertRaisesRegex
except:
checking_method = self.assertRaisesRegexp
with checking_method(ValueError, true_err_msg):
seq2seq.fit(input_texts_for_training, set(target_texts_for_training))
def test_predict_negative002(self):
""" Input texts for prediction are wrong. """
input_texts_for_testing, target_texts_for_testing = self.load_text_pairs(self.data_set_name)
seq2seq = Seq2SeqLSTM(validation_split=None, epochs=20)
seq2seq.fit(input_texts_for_testing, target_texts_for_testing)
true_err_msg = re.escape(f'`{type({1, 2})}` is wrong type for `X`.')
try:
checking_method = self.assertRaisesRegex
except:
checking_method = self.assertRaisesRegexp
with checking_method(ValueError, true_err_msg):
_ = seq2seq.predict(set(input_texts_for_testing))
def test_fit_negative09(self):
""" Number of input texts does not equal to number of target texts in the special evaluation set. """
input_texts_for_training, target_texts_for_training = self.load_text_pairs(self.data_set_name)
seq2seq = Seq2SeqLSTM()
true_err_msg = re.escape('`X_eval_set` does not correspond to `y_eval_set`! 20 != 19.')
try:
checking_method = self.assertRaisesRegex
except:
checking_method = self.assertRaisesRegexp
with checking_method(ValueError, true_err_msg):
seq2seq.fit(input_texts_for_training[:-20], target_texts_for_training[:-20],
eval_set=(input_texts_for_training[-20:], target_texts_for_training[-19:]))
def test_fit_negative07(self):
""" Object with input texts in the special evaluation set is not one of the basic sequence types. """
input_texts_for_training, target_texts_for_training = self.load_text_pairs(self.data_set_name)
seq2seq = Seq2SeqLSTM()
true_err_msg = re.escape(f'`{type({1, 2})}` is wrong type for `X_eval_set`.')
try:
checking_method = self.assertRaisesRegex
except:
checking_method = self.assertRaisesRegexp
with checking_method(ValueError, true_err_msg):
seq2seq.fit(input_texts_for_training[:-20], target_texts_for_training[:-20],
eval_set=(set(input_texts_for_training[-20:]), target_texts_for_training[-20:]))
def test_fit_negative06(self):
""" Special evaluation set is not a two-element tuple. """
input_texts_for_training, target_texts_for_training = self.load_text_pairs(self.data_set_name)
seq2seq = Seq2SeqLSTM(validation_split=None)
true_err_msg = re.escape('`eval_set` must be a two-element sequence! 3 != 2')
try:
checking_method = self.assertRaisesRegex
except:
checking_method = self.assertRaisesRegexp
with checking_method(ValueError, true_err_msg):
seq2seq.fit(input_texts_for_training[:-20], target_texts_for_training[:-20],
eval_set=(input_texts_for_training[-20:], target_texts_for_training[-20:], [3, 4]))
def test_fit_negative05(self):
""" Special evaluation set is neither list nor tuple. """
input_texts_for_training, target_texts_for_training = self.load_text_pairs(self.data_set_name)
seq2seq = Seq2SeqLSTM(validation_split=None)
true_err_msg = re.escape(f'`eval_set` must be `{type((1, 2))}` or `{type([1, 2])}`, not `{type({1: "a", 2: "b"})}`!')
try:
checking_method = self.assertRaisesRegex
except:
checking_method = self.assertRaisesRegexp
with checking_method(ValueError, true_err_msg):
seq2seq.fit(input_texts_for_training[:-20], target_texts_for_training[:-20],
eval_set={'X': input_texts_for_training[-20:], 'y': target_texts_for_training[-20:]})
def test_predict_positive001(self):
""" Part of correctly predicted texts must be greater than 0.1. """
input_texts, target_texts = self.load_text_pairs(self.data_set_name)
seq2seq = Seq2SeqLSTM(validation_split=None, epochs=200, lr=1e-2, verbose=True, lowercase=False)
predicted_texts = seq2seq.fit_predict(input_texts, target_texts)
self.assertIsInstance(predicted_texts, list)
self.assertEqual(len(predicted_texts), len(input_texts))
indices = list(range(len(predicted_texts)))
random.shuffle(indices)
print('')
print('Some predicted texts:')
for ind in range(min(5, len(predicted_texts))):
print(' True: ' + self.detokenize_text(target_texts[indices[ind]]) +
'\t Predicted: ' + self.detokenize_text(predicted_texts[indices[ind]]))
self.assertGreater(self.estimate(predicted_texts, target_texts), 0.0001)
def test_fit_positive02(self):
""" Input and target texts for training are the 1-D numpy arrays. """
input_texts_for_training, target_texts_for_training = self.load_text_pairs(self.data_set_name)
seq2seq = Seq2SeqLSTM(lr=1e-2)
res = seq2seq.fit(np.array(input_texts_for_training), np.array(target_texts_for_training))
self.assertIsInstance(res, Seq2SeqLSTM)
self.assertTrue(hasattr(res, 'input_token_index_'))
self.assertIsInstance(res.input_token_index_, dict)
self.assertTrue(hasattr(res, 'target_token_index_'))
self.assertIsInstance(res.target_token_index_, dict)
self.assertTrue(hasattr(res, 'reverse_target_char_index_'))
self.assertIsInstance(res.reverse_target_char_index_, dict)
self.assertTrue(hasattr(res, 'max_encoder_seq_length_'))
self.assertIsInstance(res.max_encoder_seq_length_, int)
self.assertGreater(res.max_encoder_seq_length_, 0)
self.assertTrue(hasattr(res, 'max_decoder_seq_length_'))
self.assertIsInstance(res.max_decoder_seq_length_, int)
self.assertGreater(res.max_decoder_seq_length_, 0)
self.assertTrue(hasattr(res, 'encoder_model_'))
self.assertIsInstance(res.encoder_model_, Model)
self.assertTrue(hasattr(res, 'decoder_model_'))
self.assertIsInstance(res.decoder_model_, Model)
def test_fit_positive04(self):
""" Early stopping is not used in the training process. """
input_texts_for_training, target_texts_for_training = self.load_text_pairs(self.data_set_name)
seq2seq = Seq2SeqLSTM(validation_split=None, lr=1e-2)
res = seq2seq.fit(input_texts_for_training, target_texts_for_training)
self.assertIsInstance(res, Seq2SeqLSTM)
self.assertTrue(hasattr(res, 'input_token_index_'))
self.assertIsInstance(res.input_token_index_, dict)
self.assertTrue(hasattr(res, 'target_token_index_'))
self.assertIsInstance(res.target_token_index_, dict)
self.assertTrue(hasattr(res, 'reverse_target_char_index_'))
self.assertIsInstance(res.reverse_target_char_index_, dict)
self.assertTrue(hasattr(res, 'max_encoder_seq_length_'))
self.assertIsInstance(res.max_encoder_seq_length_, int)
self.assertGreater(res.max_encoder_seq_length_, 0)
self.assertTrue(hasattr(res, 'max_decoder_seq_length_'))
self.assertIsInstance(res.max_decoder_seq_length_, int)
self.assertGreater(res.max_decoder_seq_length_, 0)
self.assertTrue(hasattr(res, 'encoder_model_'))
self.assertIsInstance(res.encoder_model_, Model)
self.assertTrue(hasattr(res, 'decoder_model_'))
self.assertIsInstance(res.decoder_model_, Model)
def test_creation(self):
seq2seq = Seq2SeqLSTM(batch_size=256, epochs=200, latent_dim=500, validation_split=0.1,
grad_clipping=50.0, lr=0.01, weight_decay=0.0001, lowercase=False, verbose=True)
self.assertIsInstance(seq2seq, Seq2SeqLSTM)
self.assertTrue(hasattr(seq2seq, 'batch_size'))
self.assertEqual(seq2seq.batch_size, 256)
self.assertTrue(hasattr(seq2seq, 'epochs'))
self.assertEqual(seq2seq.epochs, 200)
self.assertTrue(hasattr(seq2seq, 'latent_dim'))
self.assertEqual(seq2seq.latent_dim, 500)
self.assertTrue(hasattr(seq2seq, 'validation_split'))
self.assertAlmostEqual(seq2seq.validation_split, 0.1)
self.assertTrue(hasattr(seq2seq, 'grad_clipping'))
self.assertAlmostEqual(seq2seq.grad_clipping, 50.0)
self.assertTrue(hasattr(seq2seq, 'lr'))
self.assertAlmostEqual(seq2seq.lr, 0.01)
self.assertTrue(hasattr(seq2seq, 'weight_decay'))
self.assertAlmostEqual(seq2seq.weight_decay, 0.0001)
self.assertTrue(hasattr(seq2seq, 'lowercase'))
self.assertFalse(seq2seq.lowercase)
self.assertTrue(hasattr(seq2seq, 'verbose'))
self.assertTrue(seq2seq.verbose)
self.assertTrue(hasattr(seq2seq, 'random_state'))
self.assertIsNone(seq2seq.random_state)
def test_tokenize_text_positive02(self):
""" Tokenization with bringing the resulting tokens to lowercase. """
src = 'a\t B c Мама мыла \n\r раму 1\n'
dst_true = ['a', 'b', 'c', 'мама', 'мыла', 'раму', '1']
dst_predicted = Seq2SeqLSTM.tokenize_text(src, lowercase=True)
self.assertEqual(dst_predicted, dst_true)
def test_tokenize_text_positive01(self):
""" Tokenization with saving of the characters register. """
src = 'a\t B c Мама мыла \n\r раму 1\n'
dst_true = ['a', 'B', 'c', 'Мама', 'мыла', 'раму', '1']
dst_predicted = Seq2SeqLSTM.tokenize_text(src, lowercase=False)
self.assertEqual(dst_predicted, dst_true)
def main():
if len(sys.argv) > 1:
model_name = os.path.normpath(sys.argv[1].strip())
if len(model_name) == 0:
model_name = None
else:
model_dir_name = os.path.dirname(model_name)
if len(model_dir_name) > 0:
assert os.path.isdir(
model_dir_name), u'Directory "{0}" does not exist!'.format(
model_dir_name)
else:
model_name = None
input_texts_for_training, target_texts_for_training = shuffle_text_pairs(
*load_text_pairs(
os.path.join(os.path.dirname(__file__), 'data',
'eng_rus_for_training.txt')))
print(u'')
print(u'There are {0} text pairs in the training data.'.format(
len(input_texts_for_training)))
print(u'Some samples of these text pairs:')
for ind in range(10):
input_text = input_texts_for_training[ind]
target_text = target_texts_for_training[ind]
print(u' ' + detokenize_text(input_text) + u'\t' +
detokenize_text(target_text))
print(u'')
input_texts_for_testing, target_texts_for_testing = load_text_pairs(
os.path.join(os.path.dirname(__file__), 'data',
'eng_rus_for_testing.txt'))
print(u'There are {0} text pairs in the testing data.'.format(
len(input_texts_for_testing)))
print(u'Some samples of these text pairs:')
indices = list(range(len(input_texts_for_testing)))
random.shuffle(indices)
for ind in indices[:10]:
input_text = input_texts_for_testing[ind]
target_text = target_texts_for_testing[ind]
print(u' ' + detokenize_text(input_text) + u'\t' +
detokenize_text(target_text))
print(u'')
if (model_name is not None) and os.path.isfile(model_name):
with open(model_name, 'rb') as fp:
seq2seq = pickle.load(fp)
assert isinstance(seq2seq, Seq2SeqLSTM), \
u'A sequence-to-sequence neural model cannot be loaded from file "{0}".'.format(model_name)
print(u'')
print(u'Model has been successfully loaded from file "{0}".'.format(
model_name))
else:
seq2seq = Seq2SeqLSTM(latent_dim=256,
validation_split=0.1,
epochs=1,
lr=1e-3,
verbose=True,
lowercase=False,
batch_size=64)
seq2seq.fit(input_texts_for_training, target_texts_for_training)
print(u'')
print(u'Training has been successfully finished.')
if model_name is not None:
with open(model_name, 'wb') as fp:
pickle.dump(seq2seq, fp, protocol=2)
print(u'Model has been successfully saved into file "{0}".'.format(
model_name))
start_time = time.time()
predicted_texts = seq2seq.predict(input_texts_for_testing)
end_time = time.time()
sentence_correct, word_correct, character_correct = estimate(
predicted_texts, target_texts_for_testing)
print(u'')
print(u'{0} texts have been predicted.'.format(len(predicted_texts)))
print(u'Some samples of predicted text pairs:')
for ind in indices[:10]:
input_text = input_texts_for_testing[ind]
target_text = predicted_texts[ind]
print(u' ' + detokenize_text(input_text) + u'\t' +
detokenize_text(target_text))
print(u'')
print(u'Total sentence correct is {0:.2%}.'.format(sentence_correct))
print(u'Total word correct is {0:.2%}.'.format(word_correct))
print(u'Total character correct is {0:.2%}.'.format(character_correct))
print(u'')
print(u'Mean time of sentence prediction is {0:.3} sec.'.format(
(end_time - start_time) / len(predicted_texts)))
def test_predict_negative001(self):
""" Usage of the seq2seq model for prediction without training. """
input_texts_for_testing, _ = self.load_text_pairs(self.data_set_name)
seq2seq = Seq2SeqLSTM(validation_split=None, epochs=20)
with self.assertRaises(NotFittedError):
_ = seq2seq.predict(input_texts_for_testing)