def bilstm_self_mult_attention_model(self, params):
"""
Method builds uncompiled model of BiLSTM with self multiplicative attention
Args:
params: disctionary of parameters for NN
Returns:
Uncompiled model
"""
inp = Input(shape=(params['text_size'], params['embedding_size']))
output = Bidirectional(LSTM(params['units_lstm'], activation='tanh',
return_sequences=True,
kernel_regularizer=l2(params['coef_reg_lstm']),
dropout=params['dropout_rate'],
recurrent_dropout=params['rec_dropout_rate']))(inp)
output = MaxPooling1D(pool_size=2, strides=3)(output)
output = multiplicative_self_attention(output, n_hidden=params.get("self_att_hid"),
n_output_features=params.get("self_att_out"))
output = GlobalMaxPooling1D()(output)
output = Dropout(rate=params['dropout_rate'])(output)
output = Dense(params['dense_size'], activation=None,
kernel_regularizer=l2(params['coef_reg_den']))(output)
output = Activation('relu')(output)
output = Dropout(rate=params['dropout_rate'])(output)
output = Dense(self.n_classes, activation=None,
kernel_regularizer=l2(params['coef_reg_den']))(output)
act_output = Activation(params.get("last_layer_activation", "sigmoid"))(output)
model = Model(inputs=inp, outputs=act_output)
return model
def bilstm_self_mult_attention_model(self,
units_lstm: int,
dense_size: int,
self_att_hid: int,
self_att_out: int,
coef_reg_lstm: float = 0.,
coef_reg_den: float = 0.,
dropout_rate: float = 0.,
rec_dropout_rate: float = 0.,
**kwargs) -> Model:
"""
Method builds uncompiled model of BiLSTM with self multiplicative attention.
Args:
units_lstm: number of units for LSTM.
self_att_hid: number of hidden units in self-attention
self_att_out: number of output units in self-attention
dense_size: number of units for dense layer.
coef_reg_lstm: l2-regularization coefficient for LSTM. Default: ``0.0``.
coef_reg_den: l2-regularization coefficient for dense layers. Default: ``0.0``.
dropout_rate: dropout rate to be used after BiLSTM and between dense layers. Default: ``0.0``.
rec_dropout_rate: dropout rate for LSTM. Default: ``0.0``.
kwargs: other non-used parameters
Returns:
keras.models.Model: uncompiled instance of Keras Model
"""
inp = Input(shape=(self.opt['text_size'], self.opt['embedding_size']))
output = Bidirectional(
LSTM(units_lstm,
activation='tanh',
return_sequences=True,
kernel_regularizer=l2(coef_reg_lstm),
dropout=dropout_rate,
recurrent_dropout=rec_dropout_rate))(inp)
output = MaxPooling1D(pool_size=2, strides=3)(output)
output = multiplicative_self_attention(output,
n_hidden=self_att_hid,
n_output_features=self_att_out)
output = GlobalMaxPooling1D()(output)
output = Dropout(rate=dropout_rate)(output)
output = Dense(dense_size,
activation=None,
kernel_regularizer=l2(coef_reg_den))(output)
output = Activation('relu')(output)
output = Dropout(rate=dropout_rate)(output)
output = Dense(self.n_classes,
activation=None,
kernel_regularizer=l2(coef_reg_den))(output)
act_output = Activation(
self.opt.get("last_layer_activation", "sigmoid"))(output)
model = Model(inputs=inp, outputs=act_output)
return model
def bilstm_self_mult_attention_model(self, units_lstm: int, dense_size: int, self_att_hid: int, self_att_out: int,
coef_reg_lstm: float = 0., coef_reg_den: float = 0.,
dropout_rate: float = 0., rec_dropout_rate: float = 0.,
input_projection_size: Optional[int] = None, **kwargs) -> Model:
"""
Method builds uncompiled model of BiLSTM with self multiplicative attention.
Args:
units_lstm: number of units for LSTM.
self_att_hid: number of hidden units in self-attention
self_att_out: number of output units in self-attention
dense_size: number of units for dense layer.
coef_reg_lstm: l2-regularization coefficient for LSTM. Default: ``0.0``.
coef_reg_den: l2-regularization coefficient for dense layers. Default: ``0.0``.
dropout_rate: dropout rate to be used after BiLSTM and between dense layers. Default: ``0.0``.
rec_dropout_rate: dropout rate for LSTM. Default: ``0.0``.
input_projection_size: if not None, adds Dense layer (with ``relu`` activation)
right after input layer to the size ``input_projection_size``.
Useful for input dimentionaliry recuction. Default: ``None``.
kwargs: other non-used parameters
Returns:
keras.models.Model: uncompiled instance of Keras Model
"""
inp = Input(shape=(self.opt['text_size'], self.opt['embedding_size']))
output = inp
if input_projection_size is not None:
output = Dense(input_projection_size, activation='relu')(output)
output = Bidirectional(LSTM(units_lstm, activation='tanh',
return_sequences=True,
kernel_regularizer=l2(coef_reg_lstm),
dropout=dropout_rate,
recurrent_dropout=rec_dropout_rate))(output)
output = MaxPooling1D(pool_size=2, strides=3)(output)
output = multiplicative_self_attention(output, n_hidden=self_att_hid,
n_output_features=self_att_out)
output = GlobalMaxPooling1D()(output)
output = Dropout(rate=dropout_rate)(output)
output = Dense(dense_size, activation=None,
kernel_regularizer=l2(coef_reg_den))(output)
output = Activation('relu')(output)
output = Dropout(rate=dropout_rate)(output)
output = Dense(self.n_classes, activation=None,
kernel_regularizer=l2(coef_reg_den))(output)
act_output = Activation(self.opt.get("last_layer_activation", "sigmoid"))(output)
model = Model(inputs=inp, outputs=act_output)
return model