features: List[Feature] = [manual_features, glove_features, elmo_features]
params = RNNModelParams(layers_size=LAYERS_SIZE,
spatial_dropout=SPATIAL_DROPOUT,
recurrent_dropout=RECURRENT_DROPOUT,
dropout_dense=DROPOUT_DENSE,
dense_encoder_size=DENSE)
model = RNNModel(inputs=[input, text_input],
features=features,
output=Dense(3, activation='softmax', name="output"),
params=params,
attention=ATTENTION)
model.build()
model.compile()
X = [
value for value in tqdm(
sem_eval_dataset.iterate_train_x(max_len=MAX_LEN, one_hot=True))
]
X_text = [
value for value in tqdm(
sem_eval_dataset.iterate_train_x(max_len=MAX_LEN, one_hot=False))
]
X_val = [
value for value in tqdm(
sem_eval_dataset_dev.iterate_x(max_len=MAX_LEN, one_hot=True))
]
#ext_features = [ext_features]
params = RNNModelParams(layers_size=LAYERS_SIZE,
spatial_dropout=SPATIAL_DROPOUT,
recurrent_dropout=RECURRENT_DROPOUT,
dropout_dense=DROPOUT_DENSE,
dense_encoder_size=DENSE)
word_encoder = RNNModel(inputs=[input],
features=features,
features_to_dense=[],
output=None,
params=params,
attention=True)
word_encoder.build()
word_encoder.model().summary()
input_model = Input(shape=(
3,
MAX_LEN,
), name='input_1')
review_word_enc = TimeDistributed(word_encoder.model())(input_model)
l_lstm_sent = Bidirectional(
LSTM(200, recurrent_dropout=0.2, return_sequences=True))(review_word_enc)
l_att_sent = Attention()(l_lstm_sent)
preds = Dense(2, activation='softmax', name='output_1')(l_att_sent)
model = Model(inputs=input_model, outputs=[preds])
model.compile(loss='categorical_crossentropy',
params = RNNModelParams(layers_size=LAYERS_SIZE,
spatial_dropout=SPATIAL_DROPOUT,
recurrent_dropout=RECURRENT_DROPOUT,
dropout_dense=DROPOUT_DENSE,
dense_encoder_size=DENSE)
word_encoder = RNNModel(inputs=[pretrained_elmo_input],
input_direct=pretrained_elmo_input,
features=[],
features_to_dense=[],
output=None,
params=params,
attention=True)
word_encoder.build()
word_encoder.model().summary()
input_model = Input(shape=(
3,
MAX_LEN,
), name='input_1')
review_word_enc = TimeDistributed(word_encoder.model())(input_model)
l_lstm_sent = Bidirectional(
LSTM(200, recurrent_dropout=0.2, return_sequences=True))(review_word_enc)
l_att_sent = Attention()(l_lstm_sent)
preds = Dense(2, activation='softmax', name='output_1')(l_att_sent)
model = Model(inputs=input_model, outputs=[preds])
model.compile(loss='categorical_crossentropy',
