def build_model(self, inputs, outputs):
outputs_spati = L.SpatialDropout1D(self.dropout_spatial)(outputs)
conv_pools = []
for filter in self.filters_size:
x = L.Conv1D(
filters=self.filters_num,
kernel_size=filter,
padding="valid",
kernel_initializer="normal",
activation="relu",
)(outputs_spati)
capsule = Capsule_bojone(num_capsule=self.num_capsule,
dim_capsule=self.dim_capsule,
routings=self.routings,
kernel_size=(filter, 1),
share_weights=True)(x)
conv_pools.append(capsule)
capsule = L.Concatenate(axis=-1)(conv_pools)
x = L.Flatten()(capsule)
x = L.Dropout(self.dropout)(x)
# dense-mid
x = L.Dense(units=min(max(self.label, 64), self.embed_size),
activation=self.activate_mid)(x)
x = L.Dropout(self.dropout)(x)
# dense-end, 最后一层, dense到label
self.outputs = L.Dense(units=self.label,
activation=self.activate_end)(x)
self.model = M.Model(inputs=inputs, outputs=self.outputs)
self.model.summary(132)
def build_model(self, inputs, outputs):
x = L.SpatialDropout1D(self.dropout_spatial)(outputs)
x = SelfAttention(K.int_shape(outputs)[-1])(x)
x_max = L.GlobalMaxPooling1D()(x)
x_avg = L.GlobalAveragePooling1D()(x)
x = L.Concatenate()([x_max, x_avg])
x = L.Dropout(self.dropout)(x)
x = L.Flatten()(x)
# dense-mid
x = L.Dense(units=min(max(self.label, 64), self.embed_size),
activation=self.activate_mid)(x)
x = L.Dropout(self.dropout)(x)
# dense-end, 最后一层, dense到label
self.outputs = L.Dense(units=self.label,
activation=self.activate_end)(x)
self.model = M.Model(inputs=inputs, outputs=self.outputs)
self.model.summary(132)
def build_model(self, inputs, outputs):
# rnn type, RNN的类型
if self.rnn_unit == "LSTM":
layer_cell = L.LSTM
elif self.rnn_unit == "CuDNNLSTM":
layer_cell = L.CuDNNLSTM
elif self.rnn_unit == "CuDNNGRU":
layer_cell = L.CuDNNGRU
else:
layer_cell = L.GRU
# embedding遮挡
embedding_output_spatial = L.SpatialDropout1D(
self.dropout_spatial)(outputs)
# CNN
convs = []
for kernel_size in self.filters_size:
conv = L.Conv1D(
self.filters_num,
kernel_size=kernel_size,
strides=1,
padding='SAME',
kernel_regularizer=keras.regularizers.l2(self.l2),
bias_regularizer=keras.regularizers.l2(self.l2),
)(embedding_output_spatial)
convs.append(conv)
x = L.Concatenate(axis=1)(convs)
# Bi-LSTM, 论文中使用的是LSTM
x = L.Bidirectional(
layer_cell(units=self.rnn_unit,
return_sequences=True,
activation='relu',
kernel_regularizer=keras.regularizers.l2(self.l2),
recurrent_regularizer=keras.regularizers.l2(
self.l2)))(x)
x = L.Dropout(self.dropout)(x)
x = L.Flatten()(x)
# dense-mid
x = L.Dense(units=min(max(self.label, 64), self.embed_size),
activation=self.activate_mid)(x)
x = L.Dropout(self.dropout)(x)
# dense-end, 最后一层, dense到label
self.outputs = L.Dense(units=self.label,
activation=self.activate_end)(x)
self.model = M.Model(inputs=inputs, outputs=self.outputs)
self.model.summary(132)
def build_model(self, inputs, outputs):
# rnn type, RNN的类型
if self.rnn_unit == "LSTM":
layer_cell = L.LSTM
elif self.rnn_unit == "CuDNNLSTM":
layer_cell = L.CuDNNLSTM
elif self.rnn_unit == "CuDNNGRU":
layer_cell = L.CuDNNGRU
else:
layer_cell = L.GRU
x = L.Activation(self.activate_mid)(outputs)
# embedding遮挡
x = L.SpatialDropout1D(self.dropout_spatial)(x)
lstm_0_output = L.Bidirectional(layer_cell(
units=self.rnn_unit,
return_sequences=True,
activation='relu',
kernel_regularizer=keras.regularizers.l2(self.l2),
recurrent_regularizer=keras.regularizers.l2(self.l2)),
name="bi_lstm_0")(x)
lstm_1_output = L.Bidirectional(layer_cell(
units=self.rnn_unit,
return_sequences=True,
activation='relu',
kernel_regularizer=keras.regularizers.l2(self.l2),
recurrent_regularizer=keras.regularizers.l2(self.l2)),
name="bi_lstm_1")(lstm_0_output)
x = L.Concatenate()([lstm_1_output, lstm_0_output, x])
x = AttentionWeightedAverage(name='attlayer',
return_attention=False)(x)
x = L.Dropout(self.dropout)(x)
x = L.Flatten()(x)
# dense-mid
x = L.Dense(units=min(max(self.label, 64), self.embed_size),
activation=self.activate_mid)(x)
x = L.Dropout(self.dropout)(x)
# dense-end, 最后一层, dense到label
self.outputs = L.Dense(units=self.label,
activation=self.activate_end)(x)
self.model = M.Model(inputs=inputs, outputs=self.outputs)
self.model.summary(132)