ValueError: no signature found for builtin - cyfunction get_Model.locals.lambda

问题背景

使用cython编译keras工程，编译完成后运行时报错

问题原因

神经网络中存在keras.layers.Lambda层，对于构造参数function的lambda表示，cython无法读取为正常属性

问题代码及解决

# RNN layer
lstm_1 = LSTM(32, return_sequences = True, kernel_initializer = 'he_normal', name = 'lstm1')(inner)
lstm_1b = LSTM(32, return_sequences = True, go_backwards = True, kernel_initializer = 'he_normal',name = 'lstm1_b')(inner)  # (None, 64, 32)
# 此处为原代码，编译为pyd后执行报错
# reversed_lstm_1b = Lambda(function = lambda inputTensor: K.reverse(inputTensor, axes = 1))(lstm_1b)
def rev(inputTensor):
    return K.reverse(inputTensor, axes = 1)
reversed_lstm_1b = Lambda(function = rev)(lstm_1b)
lstm1_merged = concatenate([lstm_1, reversed_lstm_1b])  # (None, 64, 64)
lstm1_merged = BatchNormalization()(lstm1_merged)

代码及解决

我是编译为.so后，调用出错
原码

KL.Lambda(
        lambda t: tf.reshape(t, [tf.shape(t)[0], -1, 2]))(x)

改为

def rpn_class_lambda(t):
    return tf.reshape(t, [tf.shape(t)[0], -1, 2])

KL.Lambda(
        function = rpn_class_lambda)(x)