def __init__(self,
num_capsule,
dim_capsule,
routings=3,
kernel_size=(9, 1),
share_weights=True,
activation='default',
**kwargs):
super(Capsule_bojone, self).__init__(**kwargs)
self.num_capsule = num_capsule
self.dim_capsule = dim_capsule
self.routings = routings
self.kernel_size = kernel_size
self.share_weights = share_weights
if activation == 'default':
self.activation = squash_bojone
else:
self.activation = L.Activation(activation)
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)