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 call(self, inputs):
x_input_pad = L.ZeroPadding1D(
(self.filter_size - 1, self.filter_size - 1))(inputs)
conv_1d = L.Conv1D(
filters=self.filter_num,
kernel_size=self.filter_size,
strides=1,
padding="VALID",
kernel_initializer="normal", # )(x_input_pad)
activation="tanh")(x_input_pad)
return conv_1d
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):
"""
build_model.
Args:
inputs: tensor, input of model
outputs: tensor, output of model
Returns:
None
"""
# LSTM or GRU
if self.rnn_type == "LSTM":
rnn_cell = L.LSTM
elif self.rnn_type == "CuDNNLSTM":
rnn_cell = L.CuDNNLSTM
elif self.rnn_type == "CuDNNGRU":
rnn_cell = L.CuDNNGRU
else:
rnn_cell = L.GRU
# CNN-LSTM, 提取n-gram特征和最大池化， 一般不用平均池化
conv_pools = []
for i in range(len(self.filters_size)):
conv = L.Conv1D(
name="conv-{0}-{1}".format(i, self.filters_size[i]),
kernel_size=self.filters_size[i],
activation=self.activate_mid,
filters=self.filters_num,
padding='same',
)(outputs)
conv_rnn = L.Bidirectional(
rnn_cell(
name="bi-lstm-{0}-{1}".format(i, self.filters_size[i]),
activation=self.activate_mid,
return_sequences=True,
units=self.rnn_unit,
))(conv)
x_dropout = L.Dropout(rate=self.dropout,
name="dropout-{0}-{1}".format(
i, self.filters_size[i]))(conv_rnn)
conv_pools.append(x_dropout)
# 拼接
x = L.Concatenate(axis=-1)(conv_pools)
x = L.Dropout(self.dropout)(x)
# CRF or Dense
if self.use_crf:
x = L.Dense(units=self.label, activation=self.activate_end)(x)
self.CRF = ConditionalRandomField(self.crf_lr_multiplier,
name="crf_bert4keras")
self.outputs = self.CRF(x)
self.trans = K.eval(self.CRF.trans).tolist()
self.loss = self.CRF.dense_loss if self.use_onehot else self.CRF.sparse_loss
self.metrics = [
self.CRF.dense_accuracy
if self.use_onehot else self.CRF.sparse_accuracy
]
else:
self.outputs = L.TimeDistributed(
L.Dense(units=self.label,
activation=self.activate_end,
name="dense-output"))(x)
self.model = M.Model(inputs, self.outputs)
self.model.summary(132)
def build_model(self, inputs, outputs):
"""
build_model.
Args:
inputs: tensor, input of model
outputs: tensor, output of model
Returns:
None
"""
# CNN, 提取n-gram特征和最大池化, DGCNN膨胀卷积(IDCNN)
conv_pools = []
for i in range(len(self.filters_size)):
conv = L.Conv1D(
name="conv-{0}-{1}".format(i, self.filters_size[i]),
dilation_rate=self.atrous_rates[0],
kernel_size=self.filters_size[i],
activation=self.activate_mid,
filters=self.filters_num,
padding="SAME",
)(outputs)
for j in range(len(self.atrous_rates) - 1):
conv = L.Conv1D(
name="conv-{0}-{1}-{2}".format(i, self.filters_size[i], j),
dilation_rate=self.atrous_rates[j],
kernel_size=self.filters_size[i],
activation=self.activate_mid,
filters=self.filters_num,
padding="SAME",
)(conv)
conv = L.Dropout(
name="dropout-{0}-{1}-{2}".format(i, self.filters_size[i],
j),
rate=self.dropout,
)(conv)
conv_pools.append(conv)
# 拼接
x = L.Concatenate(axis=-1)(conv_pools)
x = L.Dropout(self.dropout)(x)
# CRF or Dense
if self.use_crf:
x = L.Dense(units=self.label, activation=self.activate_end)(x)
self.CRF = ConditionalRandomField(self.crf_lr_multiplier,
name="crf_bert4keras")
self.outputs = self.CRF(x)
self.trans = K.eval(self.CRF.trans).tolist()
self.loss = self.CRF.dense_loss if self.use_onehot else self.CRF.sparse_loss
self.metrics = [
self.CRF.dense_accuracy
if self.use_onehot else self.CRF.sparse_accuracy
]
else:
x = L.Bidirectional(
L.GRU(
activation=self.activate_mid,
return_sequences=True,
units=self.rnn_unit,
name="bi-gru",
))(x)
self.outputs = L.TimeDistributed(
L.Dense(
activation=self.activate_end,
name="dense-output",
units=self.label,
))(x)
self.model = M.Model(inputs, self.outputs)
self.model.summary(132)