def attention(self, embed, query):
"""
注意力机制
:param embed:
:param query:
:return:
"""
with tf.name_scope("attention"):
w = tf.get_variable(name="attention_w",
shape=[2 * hp.num_units, hp.attention_size],
dtype=tf.float32)
b = tf.get_variable(name="attention_b",
shape=[hp.attention_size],
dtype=tf.float32)
u = tf.get_variable(name="attention_u",
shape=[hp.attention_size, 1],
dtype=tf.float32)
value = tf.concat([embed, query], axis=-1)
value = tf.reshape(value, [-1, 2 * hp.num_units])
attention = tf.matmul(tf.tanh(tf.matmul(value, w) + b), u)
attention = tf.reshape(attention, shape=[-1, self.max_len])
attention = tf.nn.softmax(attention, axis=-1)
attention = tf.tile(tf.expand_dims(attention, axis=-1),
multiples=[1, 1, hp.num_units])
output = tf.reduce_sum(attention * query, axis=1)
output = layer_normalize(output)
return output
def multi_dense_layer(inputs):
"""
多层感知机 T*T*channel -> dense_size ->2
:param inputs: batch T T channel
:return:
"""
_, width, height, channel = inputs.get_shape().as_list()
size = width * height * channel
inputs = tf.reshape(inputs, shape=[-1, size])
with tf.variable_scope("dense_layer"):
w = tf.get_variable(name='w', dtype=tf.float32, shape=[size, hp.dense_size])
b = tf.get_variable(name='b', dtype=tf.float32, shape=[hp.dense_size])
outputs = layer_normalize(tf.matmul(inputs, w) + b, )
with tf.variable_scope("logit_layer"):
w = tf.get_variable(name='w', dtype=tf.float32, shape=[hp.dense_size, 2])
b = tf.get_variable(name='b', dtype=tf.float32, shape=[2])
outputs = tf.nn.softmax(tf.matmul(outputs, w) + b, axis=-1)
pre_y = tf.cast(tf.argmax(outputs, axis=-1), dtype=tf.int32)
return outputs, pre_y