ValueError：模型的输入张量必须来自`keras.layers.input`缺少元数据 - 尝试在keras tensorflow中添加lambda层以重新连接

Question

我有一个对我有用的重新连接模型。我试图在其前面添加几层lambda层。它在抱怨以下错误：

“ ValueError：模型的输入张量必须来自keras.layers.input。接收：tensor：tensor（“ shape：0”，shape =（？，），dtype = int32）（缺少上一层元数据）。

而且我不确定这意味着什么或如何解决。

我在lambda层的以下两个块中添加了代码，这引起了问题。我知道我可以在重新连接之前做一个Numpy FFT。以前已经这样做了，但是正在努力将FFT层作为带有固定系数的重新系统的前层。如果还有其他方法可以实现这一目标，例如创建另一个具有FFT lambda层并将其与Resnet进行顺序的网络，以便可以通过两个网络将错误反向传播，这是可以的。

主要目标是将FFT功能放在重新系统层的前面，并能够以最简单的/最佳方式将重新网络分类器的误差倒入FFT层之前。

def resnode(x_in,filter_count,kernel_size=3,downsample=False):
    x = x_in
    x = keras_contrib.layers.InstanceNormalization()(x)
    x = keras.layers.Lambda(lambda v: tf.spectral.fft(tf.cast(v,tf.complex64)))(x) 
    x = keras.layers.Lambda(lambda v: tf.abs(tf.cast(v,tf.complex64)))(x)    
    x = keras.layers.ReLU()(x)
    x = keras.layers.Conv1D(filter_count,kernel_size,padding='same')(x)
    x = keras.layers.Dropout(0.1)(x)
    x = keras_contrib.layers.InstanceNormalization()(x)
    x = keras.layers.ReLU()(x)
    x = keras.layers.Conv1D(filter_count,kernel_size,padding='same')(x)

    if downsample:
        x_in = keras.layers.AveragePooling1D()(x_in)
        x = keras.layers.AveragePooling1D()(x)

if x_in.shape[-1] != x.shape[-1]:
    print('convolving x_in',x_in.shape)
    x_in = keras.layers.Conv1D(filter_count,1,padding='same')(x_in)
    print('Newshape x_in',x_in.shape)
    
x = keras.layers.Add()([x_in,x])

return x

def createresnet(outputs=4):  
    n_features = N    
    x_input = keras.layers.Input(shape=(None,1))

    x_input = tf.shape(tf.squeeze(x_input))
    x = keras.layers.Lambda(lambda v: tf.spectral.fft(tf.cast(v,tf.complex64)))(x_input) # 
    x = keras.layers.Lambda(lambda v: tf.abs(tf.cast(v,tf.complex64)))(x)
    x = tf.cast(x, tf.float32)
    x = tf.expand_dims(x, axis=0)
    x = tf.expand_dims(x, axis=2)

    filter_size = 32 # 
    x = resnode(x,filter_size,kernel_size=7)  # FFT layer

    filter_size = 32
    x = resnode(x,filter_size,kernel_size=5,downsample=True)
    filter_size = 64
    x = resnode(x,filter_size,kernel_size=5,downsample=True)
    x = resnode(x,filter_size,kernel_size=5,downsample=True)
    filter_size = 128
    x = resnode(x,filter_size,downsample=True)
    x = resnode(x,filter_size,downsample=True)
    filter_size = 128
    x = resnode(x,filter_size,downsample=True)
    x = resnode(x,filter_size,downsample=True)
    filter_size = 128
    x = resnode(x,filter_size,downsample=True)
    x = resnode(x,filter_size,downsample=True)
    filter_size = 128
    x = resnode(x,filter_size,downsample=True)
    x = resnode(x,filter_size,downsample=True)
    filter_size = 64
    x = resnode(x,filter_size,downsample=True)

    if False:
        filter_size = 64
        x = resnode(x,filter_size)
        x = resnode(x,filter_size,downsample=True)
        filter_size = 64
        x = resnode(x,filter_size)
        x = resnode(x,filter_size,downsample=True)
        filter_size = 128
        x = resnode(x,filter_size)
        x = resnode(x,filter_size,downsample=True)
        filter_size = 128
        x = resnode(x,filter_size)
        x = resnode(x,filter_size,downsample=True)
        filter_size = 128
        x = resnode(x,filter_size)
        x = resnode(x,filter_size,downsample=True)
        filter_size = 128
        x = resnode(x,filter_size)
        x = resnode(x,filter_size,downsample=True)

    filter_size = outputs
    x = resnode(x,filter_size,downsample=True)
    x = keras.layers.GlobalAveragePooling1D()(x)
    x = keras.layers.Activation('softmax')(x)

    model = keras.Model(inputs=x_input,outputs=x)
    print(model.layers[0].output)
    print(model.layers[1].output)    

    return model

感谢任何可以提供帮助的人！

Answer 1

正如评论中的@frightera所说，Squeeze（）删除了大小1的所有维度。您可能正在寻找Flatten（），这将使您的频道减少到适用1D傅立叶变换的FFT（）中的1D张量。

x_input = tf.flatten(x_input)
x = keras.layers.Lambda(lambda v: tf.spectral.fft(tf.cast(v,tf.complex64)))(x_input)