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):
# 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)
def squash_bojone(x, axis=-1):
"""
activation of squash
:param x: vector
:param axis: int
:return: vector
"""
s_squared_norm = K.sum(K.square(x), axis, keepdims=True)
scale = K.sqrt(s_squared_norm + K.epsilon())
return x / scale
def build_model(self, inputs, outputs):
x = L.Dense(units=self.label, activation=self.activate_mid)(outputs)
self.CRF = ConditionalRandomField(self.crf_lr_multiplier,
name="crf_bert4keras")
self.outputs = self.CRF(x)
self.model = M.Model(inputs, self.outputs)
self.model.summary(132)
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
]
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
# Bi-LSTM-LAN
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)
x_att = SelfAttention(K.int_shape(x)[-1])(x)
outputs = L.Concatenate()([x, x_att])
outputs = L.Dropout(self.dropout)(outputs)
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 call(self, x):
WQ = K.dot(x, self.kernel[0])
WK = K.dot(x, self.kernel[1])
WV = K.dot(x, self.kernel[2])
# print("WQ.shape",WQ.shape)
# print("K.permute_dimensions(WK, [0, 2, 1]).shape",K.permute_dimensions(WK, [0, 2, 1]).shape)
QK = K.batch_dot(WQ, K.permute_dimensions(WK, [0, 2, 1]))
QK = QK / (64**0.5)
QK = K.softmax(QK)
# print("QK.shape",QK.shape)
V = K.batch_dot(QK, WV)
return V
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):
"""
build_model.
Args:
inputs: tensor, input of model
outputs: tensor, output of model
Returns:
None
"""
# CuDNNGRU or GRU
x = None
if self.rnn_type.upper() == "CUDNNGRU":
rnn_cell = L.CuDNNGRU
else:
rnn_cell = L.GRU
# Bi-GRU
for nrl in range(self.num_rnn_layers):
x = L.Bidirectional(
rnn_cell(
units=self.rnn_unit,
return_sequences=True,
activation=self.activate_mid,
))(outputs)
x = L.Dropout(self.dropout)(x)
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))(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
"""
# 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 call(self, u_vecs):
if self.share_weights:
u_hat_vecs = K.conv1d(u_vecs, self.W)
else:
u_hat_vecs = K.local_conv1d(u_vecs, self.W, [1], [1])
batch_size = K.shape(u_vecs)[0]
input_num_capsule = K.shape(u_vecs)[1]
u_hat_vecs = K.reshape(u_hat_vecs,
(batch_size, input_num_capsule,
self.num_capsule, self.dim_capsule))
u_hat_vecs = K.permute_dimensions(u_hat_vecs, (0, 2, 1, 3))
# final u_hat_vecs.shape = [None, num_capsule, input_num_capsule, dim_capsule]
b = K.zeros_like(
u_hat_vecs[:, :, :,
0]) # shape = [None, num_capsule, input_num_capsule]
outputs = None
for i in range(self.routings):
b = K.permute_dimensions(
b, (0, 2, 1)) # shape = [None, input_num_capsule, num_capsule]
c = K.softmax(b)
c = K.permute_dimensions(c, (0, 2, 1))
b = K.permute_dimensions(b, (0, 2, 1))
outputs = self.activation(K.batch_dot(c, u_hat_vecs, [2, 2]))
if i < self.routings - 1:
b = K.batch_dot(outputs, u_hat_vecs, [2, 3])
return outputs
def call(self, x, mask=None):
# computes a probability distribution over the timesteps
# uses "max trick" for numerical stability
# reshape is done to avoid issue with Tensorflow
# and 1-dimensional weights
logits = K.dot(x, self.W)
x_shape = K.shape(x)
logits = K.reshape(logits, (x_shape[0], x_shape[1]))
ai = K.exp(logits - K.max(logits, axis=-1, keepdims=True))
# masked timesteps have zero weight
if mask is not None:
mask = K.cast(mask, K.floatx())
ai = ai * mask
att_weights = ai / (K.sum(ai, axis=1, keepdims=True) + K.epsilon())
weighted_input = x * K.expand_dims(att_weights)
result = K.sum(weighted_input, axis=1)
if self.return_attention:
return [result, att_weights]
return result
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)