def build_model_bilstm_single(self):
if args.use_lstm:
if args.use_cudnn_cell:
layer_cell = CuDNNLSTM
else:
layer_cell = LSTM
else:
if args.use_cudnn_cell:
layer_cell = CuDNNGRU
else:
layer_cell = GRU
# bert embedding
bert_inputs, bert_output = KerasBertEmbedding().bert_encode()
# Bi-LSTM
x = Bidirectional(
layer_cell(units=args.units,
return_sequences=args.return_sequences,
kernel_regularizer=regularizers.l2(args.l2 * 0.1),
recurrent_regularizer=regularizers.l2(
args.l2)))(bert_output)
x = Dropout(args.keep_prob)(x)
x = Flatten()(x)
# 最后就是softmax
dense_layer = Dense(args.label, activation=args.activation)(x)
output_layers = [dense_layer]
self.model = Model(bert_inputs, output_layers)
def build_model_r_cnn(self):
######### RCNN #########
# bert embedding
bert_inputs, bert_output = KerasBertEmbedding().bert_encode()
# rcnn
bert_output_emmbed = SpatialDropout1D(rate=self.keep_prob)(bert_output)
if args.use_lstm:
if args.use_cudnn_cell:
layer_cell = CuDNNLSTM
else:
layer_cell = LSTM
else:
if args.use_cudnn_cell:
layer_cell = CuDNNGRU
else:
layer_cell = GRU
x = Bidirectional(
layer_cell(units=args.units,
return_sequences=args.return_sequences,
kernel_regularizer=regularizers.l2(args.l2 * 0.1),
recurrent_regularizer=regularizers.l2(
args.l2)))(bert_output_emmbed)
x = Dropout(args.keep_prob)(x)
x = Conv1D(filters=int(self.embedding_dim / 2),
kernel_size=2,
padding='valid',
kernel_initializer='normal',
activation='relu')(x)
x = GlobalMaxPooling1D()(x)
x = Dropout(args.keep_prob)(x)
# 最后就是softmax
dense_layer = Dense(self.label, activation=self.activation)(x)
output_layers = [dense_layer]
self.model = Model(bert_inputs, output_layers)
def build_model_bilstm_layers(self):
if args.use_lstm:
if args.use_cudnn_cell:
layer_cell = CuDNNLSTM
else:
layer_cell = LSTM
else:
if args.use_cudnn_cell:
layer_cell = CuDNNGRU
else:
layer_cell = GRU
# bert embedding
bert_inputs, bert_output = KerasBertEmbedding().bert_encode()
# bert_output = bert_output[:0:]
# layer_get_cls = Lambda(lambda x: x[:, 0:1, :])
# bert_output = layer_get_cls(bert_output)
# print("layer_get_cls:")
# print(bert_output.shape)
# Bi-LSTM
x = Bidirectional(
layer_cell(units=args.units,
return_sequences=args.return_sequences,
kernel_regularizer=regularizers.l2(args.l2 * 0.1),
recurrent_regularizer=regularizers.l2(
args.l2)))(bert_output)
# blstm_layer = TimeDistributed(Dropout(args.keep_prob))(blstm_layer) 这个用不了，好像是输入不对, dims<3吧
x = Dropout(args.keep_prob)(x)
x = Bidirectional(
layer_cell(units=args.units,
return_sequences=args.return_sequences,
kernel_regularizer=regularizers.l2(args.l2 * 0.1),
recurrent_regularizer=regularizers.l2(args.l2)))(x)
x = Dropout(args.keep_prob)(x)
x = Bidirectional(
layer_cell(units=args.units,
return_sequences=args.return_sequences,
kernel_regularizer=regularizers.l2(args.l2 * 0.1),
recurrent_regularizer=regularizers.l2(args.l2)))(x)
x = Dropout(args.keep_prob)(x)
# 平均池化、最大池化拼接
avg_pool = GlobalAvgPool1D()(x)
max_pool = GlobalMaxPool1D()(x)
print(max_pool.shape)
print(avg_pool.shape)
concat = concatenate([avg_pool, max_pool])
x = Dense(int(args.units / 4), activation="relu")(concat)
x = Dropout(args.keep_prob)(x)
# 最后就是softmax
dense_layer = Dense(args.label, activation=args.activation)(x)
output_layers = [dense_layer]
self.model = Model(bert_inputs, output_layers)
def build_model_text_cnn(self):
######### text-cnn #########
# bert embedding
bert_inputs, bert_output = KerasBertEmbedding().bert_encode()
# text cnn
bert_output_emmbed = SpatialDropout1D(rate=self.keep_prob)(bert_output)
concat_out = []
for index, filter_size in enumerate(self.filters):
x = Conv1D(name='TextCNN_Conv1D_{}'.format(index),
filters=int(self.embedding_dim / 2),
kernel_size=self.filters[index],
padding='valid',
kernel_initializer='normal',
activation='relu')(bert_output_emmbed)
x = GlobalMaxPooling1D(
name='TextCNN_MaxPool1D_{}'.format(index))(x)
concat_out.append(x)
x = Concatenate(axis=1)(concat_out)
x = Dropout(self.keep_prob)(x)
# 最后就是softmax
dense_layer = Dense(self.label, activation=self.activation)(x)
output_layers = [dense_layer]
self.model = Model(bert_inputs, output_layers)
def build_model_avt_cnn(self):
#########text-cnn#########
# bert embedding
bert_inputs, bert_output = KerasBertEmbedding().bert_encode()
# text cnn
bert_output_emmbed = SpatialDropout1D(rate=self.keep_prob)(bert_output)
concat_x = []
concat_y = []
concat_z = []
for index, filter_size in enumerate(self.filters):
conv = Conv1D(name='TextCNN_Conv1D_{}'.format(index),
filters=int(self.embedding_dim / 2),
kernel_size=self.filters[index],
padding='valid',
kernel_initializer='normal',
activation='relu')(bert_output_emmbed)
x = GlobalMaxPooling1D(
name='TextCNN_MaxPooling1D_{}'.format(index))(conv)
y = GlobalAveragePooling1D(
name='TextCNN_AveragePooling1D_{}'.format(index))(conv)
z = AttentionWeightedAverage(
name='TextCNN_Annention_{}'.format(index))(conv)
concat_x.append(x)
concat_y.append(y)
concat_z.append(z)
merge_x = Concatenate(axis=1)(concat_x)
merge_y = Concatenate(axis=1)(concat_y)
merge_z = Concatenate(axis=1)(concat_z)
merge_xyz = Concatenate(axis=1)([merge_x, merge_y, merge_z])
x = Dropout(self.keep_prob)(merge_xyz)
# 最后就是softmax
dense_layer = Dense(self.label, activation=self.activation)(x)
output_layers = [dense_layer]
self.model = Model(bert_inputs, output_layers)