def test_training_summarizer_attention(self) -> None:
summarizer_attention = SummarizerAttention(lstm_size=10,
embedding_size=10)
summarizer_attention.init_model(preprocessor=self.preprocessor,
vectorizer=self.vectorizer,
embedding_weights_encoder=None,
embedding_weights_decoder=None)
loss_attention = 0
train_step = summarizer_attention.new_train_step(
loss_function=self.loss_func, batch_size=2)
for _ in range(10):
for source_seq, target_seq in self.dataset.take(-1):
loss_attention = train_step(source_seq, target_seq)
print(str(loss_attention))
self.assertAlmostEqual(1.577033519744873, float(loss_attention), 5)
output_attention = summarizer_attention.predict_vectors('a c', '')
expected_first_logits = np.array(
[-0.077805, 0.012667, 0.021359, -0.04872, 0.014989])
np.testing.assert_allclose(expected_first_logits,
output_attention['logits'][0],
atol=1e-6)
self.assertEqual('<start> a c <end>',
output_attention['preprocessed_text'][0])
self.assertEqual('d <end>', output_attention['predicted_text'])
def test_serde_happy_path(self) -> None:
preprocessor = Preprocessor()
tokenizer = KerasTokenizer(oov_token='<unk>')
tokenizer.fit([
'a b c {} {}'.format(preprocessor.start_token,
preprocessor.end_token)
])
vectorizer = Vectorizer(tokenizer, tokenizer)
summarizer = SummarizerAttention(lstm_size=10,
max_prediction_len=10,
embedding_size=10,
embedding_encoder_trainable=False)
summarizer.init_model(preprocessor=preprocessor, vectorizer=vectorizer)
# we need at least a train step to init the weights
train_step = summarizer.new_train_step(masked_crossentropy,
batch_size=1,
apply_gradients=True)
train_seq = tf.convert_to_tensor(np.array([[1, 1, 1]]), dtype=tf.int32)
train_step(train_seq, train_seq)
save_dir = os.path.join(self.temp_dir, 'summarizer_serde_happy_path')
summarizer.save(save_dir)
summarizer_loaded = SummarizerAttention.load(save_dir)
self.assertEqual(10, summarizer_loaded.lstm_size)
self.assertEqual(10, summarizer_loaded.max_prediction_len)
self.assertIsNotNone(summarizer_loaded.preprocessor)
self.assertIsNotNone(summarizer_loaded.vectorizer)
self.assertIsNotNone(summarizer_loaded.encoder)
self.assertIsNotNone(summarizer_loaded.decoder)
self.assertFalse(summarizer_loaded.encoder.embedding.trainable)
self.assertTrue(summarizer_loaded.decoder.embedding.trainable)
self.assertIsNotNone(summarizer_loaded.optimizer)
pred = summarizer.predict_vectors('a c', '')
pred_loaded = summarizer_loaded.predict_vectors('a c', '')
np.testing.assert_almost_equal(pred['logits'],
pred_loaded['logits'],
decimal=6)
import streamlit as st
from headliner.model.summarizer_transformer import SummarizerTransformer
from headliner.model.summarizer_attention import SummarizerAttention
summarizer_transformer = SummarizerTransformer.load('model/transformer')
summarizer_attention = SummarizerAttention.load('model/attention')
st.title('English-German Translator')
st.markdown('''
This is a demo showcasing our [Headliner package](). In particular, we trained
a simple seq2seq model on an English-German dataset. We didn't train it very long so
the model is not doing well as this was not our main goals anyway. For creating the app,
we use [Streamlit](https://streamlit.io/), a new open-source framework that lets users creating
apps for machine learning projects very easily.
''')
input = st.text_input(label='Type in some English words.',
value='I really like you.')
st.write('(transformer) {}'.format(summarizer_transformer.predict(input)))
st.write('(attention) {}'.format(summarizer_attention.predict(input)))
def test_training(self) -> None:
data = [('a b', 'c'), ('a b c', 'd')]
tokenizer_encoder = KerasTokenizer(lower=False, filters='')
tokenizer_decoder = KerasTokenizer(lower=False, filters='')
tokenizer_encoder.fit(['a b c <start> <end>'])
tokenizer_decoder.fit(['c d <start> <end>'])
vectorizer = Vectorizer(tokenizer_encoder=tokenizer_encoder,
tokenizer_decoder=tokenizer_decoder,
max_output_len=3)
preprocessor = Preprocessor()
batch_generator = DatasetGenerator(2)
data_prep = [preprocessor(d) for d in data]
data_vecs = [vectorizer(d) for d in data_prep]
dataset = batch_generator(lambda: data_vecs)
summarizer_transformer = SummarizerTransformer(num_heads=1,
num_layers=1,
feed_forward_dim=20,
embedding_size=10,
dropout_rate=0,
max_prediction_len=3)
summarizer_transformer.init_model(preprocessor=preprocessor,
vectorizer=vectorizer,
embedding_weights_encoder=None,
embedding_weights_decoder=None)
summarizer_attention = SummarizerAttention(lstm_size=10,
embedding_size=10)
summarizer_attention.init_model(preprocessor=preprocessor,
vectorizer=vectorizer,
embedding_weights_encoder=None,
embedding_weights_decoder=None)
summarizer = SummarizerBasic(lstm_size=10, embedding_size=10)
summarizer.init_model(preprocessor=preprocessor,
vectorizer=vectorizer,
embedding_weights_encoder=None,
embedding_weights_decoder=None)
loss_func = masked_crossentropy
loss_attention = 0
train_step = summarizer_attention.new_train_step(
loss_function=loss_func, batch_size=2)
for _ in range(10):
for source_seq, target_seq in dataset.take(-1):
loss_attention = train_step(source_seq, target_seq)
print(str(loss_attention))
self.assertAlmostEqual(1.5810251235961914, float(loss_attention), 10)
output_attention = summarizer_attention.predict_vectors('a c', '')
expected_first_logits = np.array(
[-0.069454, 0.00272, 0.007199, -0.039547, 0.014357])
np.testing.assert_allclose(expected_first_logits,
output_attention['logits'][0],
atol=1e-6)
self.assertEqual('a c', output_attention['preprocessed_text'][0])
self.assertEqual('<end>', output_attention['predicted_text'])
loss = 0
train_step = summarizer.new_train_step(loss_function=loss_func,
batch_size=2)
for e in range(0, 10):
for source_seq, target_seq in dataset.take(-1):
loss = train_step(source_seq, target_seq)
self.assertAlmostEqual(1.5771859884262085, float(loss), 10)
output = summarizer.predict_vectors('a c', '')
expected_first_logits = np.array(
[-0.03838864, 0.01226684, 0.01055636, -0.05209339, 0.02549592])
np.testing.assert_allclose(expected_first_logits,
output['logits'][0],
atol=1e-6)
self.assertEqual('a c', output['preprocessed_text'][0])
self.assertEqual('<end>', output['predicted_text'])
loss_transformer = 0
train_step = summarizer_transformer.new_train_step(
loss_function=loss_func, batch_size=2)
for e in range(0, 10):
for source_seq, target_seq in dataset.take(-1):
loss_transformer = train_step(source_seq, target_seq)
print(str(loss_transformer))
self.assertAlmostEqual(1.2841172218322754, float(loss_transformer), 10)
output_transformer = summarizer_transformer.predict_vectors('a c', '')
expected_first_logits = np.array(
[0.094787, 0.516092, 1.165521, 0.271338, 0.670318])
np.testing.assert_allclose(expected_first_logits,
output_transformer['logits'][0],
atol=1e-6)
self.assertEqual('a c', output_transformer['preprocessed_text'][0])
self.assertEqual('d <end>', output_transformer['predicted_text'])
def test_training(self) -> None:
data = [('a b', 'c'), ('a b c', 'd')]
tokenizer_encoder = KerasTokenizer(lower=False, filters='')
tokenizer_decoder = KerasTokenizer(lower=False, filters='')
tokenizer_encoder.fit(['a b c <start> <end>'])
tokenizer_decoder.fit(['c d <start> <end>'])
vectorizer = Vectorizer(tokenizer_encoder=tokenizer_encoder,
tokenizer_decoder=tokenizer_decoder,
max_output_len=3)
preprocessor = Preprocessor()
batch_generator = DatasetGenerator(2)
data_prep = [preprocessor(d) for d in data]
data_vecs = [vectorizer(d) for d in data_prep]
dataset = batch_generator(lambda: data_vecs)
summarizer_transformer = SummarizerTransformer(num_heads=1,
num_layers=1,
feed_forward_dim=20,
embedding_size=10,
dropout_rate=0,
max_prediction_len=3)
summarizer_transformer.init_model(preprocessor=preprocessor,
vectorizer=vectorizer,
embedding_weights_encoder=None,
embedding_weights_decoder=None)
summarizer_attention = SummarizerAttention(lstm_size=10,
embedding_size=10)
summarizer_attention.init_model(preprocessor=preprocessor,
vectorizer=vectorizer,
embedding_weights_encoder=None,
embedding_weights_decoder=None)
summarizer = SummarizerBasic(lstm_size=10, embedding_size=10)
summarizer.init_model(preprocessor=preprocessor,
vectorizer=vectorizer,
embedding_weights_encoder=None,
embedding_weights_decoder=None)
loss_func = masked_crossentropy
loss_attention = 0
train_step = summarizer_attention.new_train_step(
loss_function=loss_func, batch_size=2)
for _ in range(10):
for source_seq, target_seq in dataset.take(-1):
loss_attention = train_step(source_seq, target_seq)
print(str(loss_attention))
self.assertAlmostEqual(1.577033519744873, float(loss_attention), 5)
output_attention = summarizer_attention.predict_vectors('a c', '')
expected_first_logits = np.array(
[-0.077805, 0.012667, 0.021359, -0.04872, 0.014989])
np.testing.assert_allclose(expected_first_logits,
output_attention['logits'][0],
atol=1e-6)
self.assertEqual('<start> a c <end>',
output_attention['preprocessed_text'][0])
self.assertEqual('d <end>', output_attention['predicted_text'])
loss = 0
train_step = summarizer.new_train_step(loss_function=loss_func,
batch_size=2)
for e in range(0, 10):
for source_seq, target_seq in dataset.take(-1):
loss = train_step(source_seq, target_seq)
self.assertAlmostEqual(1.5713274478912354, float(loss), 5)
output = summarizer.predict_vectors('a c', '')
expected_first_logits = np.array(
[-0.051753, 0.013869, 0.010337, -0.073727, 0.033059])
np.testing.assert_allclose(expected_first_logits,
output['logits'][0],
atol=1e-6)
self.assertEqual('<start> a c <end>', output['preprocessed_text'][0])
self.assertEqual('<end>', output['predicted_text'])
loss_transformer = 0
train_step = summarizer_transformer.new_train_step(
loss_function=loss_func, batch_size=2)
for e in range(0, 10):
for source_seq, target_seq in dataset.take(-1):
loss_transformer = train_step(source_seq, target_seq)
print(str(loss_transformer))
self.assertAlmostEqual(1.175953984260559, float(loss_transformer), 5)
output_transformer = summarizer_transformer.predict_vectors('a c', '')
expected_first_logits = np.array(
[-0.197903, 0.884185, 1.147212, 0.318798, 0.97936])
np.testing.assert_allclose(expected_first_logits,
output_transformer['logits'][0],
atol=1e-6)
self.assertEqual('<start> a c <end>',
output_transformer['preprocessed_text'][0])
self.assertEqual('d <end>', output_transformer['predicted_text'])