Question

我想在Keras中编写一个图层，但出现错误：

AttributeError：“ NoneType”对象没有属性“ _inbound_nodes”

后端功能可能是问题所在。

但是，我搜索了几个小时，发现的大多数答案是“在Lambda中翘曲”，但是当我在Lambda中包装所有后端函数（我使用tensorflow）时，我仍然会出错。

我什至在keras github＃12672中打开了一个问题，但没有人提出申请，帮助。 :-( https://github.com/keras-team/keras/issues/12672

from keras.engine.topology import Layer
import keras.backend as K
from keras.layers import Lambda
import numpy as np

class Attention(Layer):

    def __init__(self, **kwargs):
        super(Attention, self).__init__(**kwargs)

    def build(self, input_shape):
        # three weight
        # Wh: att_size, att_size -->  previous hidden state
        # Wq: query_dim, att_size --> target hidden state
        # V:  att_size, 1 --> tanh
        # score(previous, target) = Vt * tanh(Wh * previous + target * Wq + b???) --> (1, 1)

        # the dimension of previous hidden state
        self.att_size = input_shape[0][-1]
        # the dimension of target hidden state
        self.query_dim = input_shape[1][-1]


        self.Wq = self.add_weight(name='kernal_query_features', shape=(self.query_dim, self.att_size),
                                  initializer='glorot_normal', trainable=True)

        self.Wh = self.add_weight(name='kernal_hidden_features', shape=(self.att_size, self.att_size),
                                  initializer='glorot_normal', trainable=True)

        self.v = self.add_weight(name='query_vector', shape=(self.att_size, 1),
                                 initializer='zeros', trainable=True)

        super(Attention, self).build(input_shape)


    def call(self, inputs, mask=None):


        # score(previous, target) = Vt * tanh(Wh * memory + target * Wq)

        memory, query = inputs[0], inputs[1]
        hidden = K.dot(memory, self.Wh) + K.expand_dims(K.dot(query, self.Wq), 1)
        hidden = K.tanh(hidden)
        # remove the dimension whose shape is 1
        s = K.squeeze(K.dot(hidden, self.v), -1)

        #s= K.reshape(K.dot(hidden, self.v), (-1, self.att_size))
        # compute the weight use soft_max
        s = K.softmax(s)

        return K.sum(memory * K.expand_dims(s), axis=1)


    def compute_output_shape(self, input_shape):
        att_size = input_shape[0][-1]
        batch = input_shape[0][0]
        return batch, att_size

在我的主要代码中

lstm_global_seq = LSTM(units=2 * neighbor_slide_len * neighbor_slide_len,
                           return_sequences=True, dropout=0.1,
                           recurrent_dropout=0.1, name="att_global")(global_data)

    att_global = Attention()([lstm_global_seq[:-1], lstm_global_seq[-1]])

    lstm_global = merge.Concatenate(axis=-1)([att_global, lstm_global_seq[-1]])

我希望这个自定义图层可以正确运行

当我在Keras中编写Layer时，出现此错误。 AttributeError：“ NoneType”对象没有属性“ _inbound_nodes”

0 个答案: