def build_model(self, inputs, outputs):
"""
build_model.
Args:
inputs: tensor, input of model
outputs: tensor, output of model
Returns:
None
"""
embed_char = outputs[0]
embed_word = outputs[1]
if self.wclstm_embed_type == "ATTNENTION":
x_word = L.TimeDistributed(
SelfAttention(K.int_shape(embed_word)[-1]))(embed_word)
x_word_shape = K.int_shape(x_word)
x_word = L.Reshape(target_shape=(x_word_shape[:2],
x_word_shape[2] *
x_word_shape[3]))
x_word = L.Dense(self.embed_size,
activation=self.activate_mid)(x_word)
# elif self.wclstm_embed_type == "SHORT":
else:
x_word = L.Lambda(lambda x: x[:, :, 0, :])(embed_word)
outputs_concat = L.Concatenate(axis=-1)([embed_char, x_word])
# 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
# Bi-LSTM-CRF
for nrl in range(self.num_rnn_layers):
x = L.Bidirectional(
rnn_cell(
units=self.rnn_unit * (nrl + 1),
return_sequences=True,
activation=self.activate_mid,
))(outputs_concat)
outputs = L.Dropout(self.dropout)(x)
if self.use_crf:
x = L.Dense(units=self.label,
activation=self.activate_end)(outputs)
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))(outputs)
self.model = M.Model(inputs, self.outputs)
self.model.summary(132)
def build_model(self, inputs, outputs):
if self.embed_type in ["xlnet"]:
# x = L.Lambda(lambda x: x[:, -2:-1, :])(outputs) # xlnet获取CLS
x = L.Lambda(lambda x: x[:, -1], name="Token-CLS")(outputs)
else:
# x = L.Lambda(lambda x: x[:, 0:1, :])(outputs) # bert-like获取CLS
x = L.Lambda(lambda x: x[:, 0], name="Token-CLS")(outputs)
# x = L.Flatten()(x)
# 最后就是softmax
self.outputs = L.Dense(
self.label,
activation=self.activate_end,
kernel_initializer=keras.initializers.TruncatedNormal(
stddev=0.02))(x)
self.model = M.Model(inputs, self.outputs)
self.model.summary(132)
def build_model(self, inputs, outputs):
# rnn type, RNN的类型
if self.rnn_type == "LSTM":
layer_cell = L.LSTM
else:
layer_cell = L.GRU
# backword, 反向
x_backwords = layer_cell(
units=self.rnn_unit,
return_sequences=True,
kernel_regularizer=keras.regularizers.l2(0.32 * 0.1),
recurrent_regularizer=keras.regularizers.l2(0.32),
go_backwards=True)(outputs)
x_backwords_reverse = L.Lambda(lambda x: K.reverse(x, axes=1))(
x_backwords)
# fordword, 前向
x_fordwords = layer_cell(
units=self.rnn_unit,
return_sequences=True,
kernel_regularizer=keras.regularizers.l2(0.32 * 0.1),
recurrent_regularizer=keras.regularizers.l2(0.32),
go_backwards=False)(outputs)
# concatenate, 拼接
x_feb = L.Concatenate(axis=2)(
[x_fordwords, outputs, x_backwords_reverse])
# dropout, 随机失活
x_feb = L.Dropout(self.dropout)(x_feb)
# Concatenate, 拼接后的embedding_size
dim_2 = K.int_shape(x_feb)[2]
x_feb_reshape = L.Reshape((self.length_max, dim_2, 1))(x_feb)
# n-gram, conv, maxpool, 使用n-gram进行卷积和池化
conv_pools = []
for filter in self.filters_size:
conv = L.Conv2D(
filters=self.filters_num,
kernel_size=(filter, dim_2),
padding='valid',
kernel_initializer='normal',
activation='relu',
)(x_feb_reshape)
pooled = L.MaxPooling2D(
pool_size=(self.length_max - filter + 1, 1),
strides=(1, 1),
padding='valid',
)(conv)
conv_pools.append(pooled)
# concatenate, 拼接TextCNN
x = L.Concatenate()(conv_pools)
x = L.Dropout(self.dropout)(x)
# dense-mid, 中间全连接到中间的隐藏元
x = L.Flatten()(x)
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)
dev_data = txt_read(path_dev)
len_train_rate = int(len(train_data) * rate)
len_dev_rate = int(len(dev_data) * rate)
train_data = train_data[:len_train_rate]
dev_data = dev_data[:len_dev_rate]
logger.info("训练/验证语料读取完成")
# 数据预处理类初始化
preprocess_xy = ListPrerocessXY(embed,
train_data,
path_dir=path_model_dir,
length_max=length_max)
x = L.Lambda(lambda x: x[:, 0], name="Token-CLS")(embed.model.output)
# 最后就是softmax
outputs = L.Dense(
len(preprocess_xy.l2i),
activation="softmax",
kernel_initializer=keras.initializers.TruncatedNormal(stddev=0.02))(x)
model = M.Model(embed.model.input, outputs)
model.summary(132)
model.compile(optimizer=O.Adam(lr=1e-5),
loss="categorical_crossentropy",
metrics=["accuracy"])
len_train_data = len(train_data)
lg_train = ListGenerator(train_data,