在一个功能模型中组合两个 keras 顺序模型时出错：AttributeError: 'NoneType' object has no attribute '_inbound_nodes'

Question

我分别构建了两个keras序列模型，并使用keras函数api将两个模型结合起来。在它们之间，我应用了数据重新打包，调整了通过两个顺序模型的数据的大小。

当我进行组合时，存在错误。 但我不太明白哪个对象在“NoneType”中。

另一个问题是，我在进行数据张量重新打包时将所有零放在占位符内。当我运行 model.fit 时，它会被真实数据取代，还是会保持为零？ 将两个序列模型与张量调整之间的组合完全不适用吗？

# I use a keras sequential model to define the 2d model "base_model_SRCNN"
# I define the 3d model still using keras sequential model as "SRnet_3d_model"

ip = Input(shape = (TARGET_HEIGHT, TARGET_WIDTH, 1))
SRCNN_network = base_model_SRCNN(FILENAME, TARGET_HEIGHT, TARGET_WIDTH) (ip)
#----------------------------pack frame one by one---------------------------------------
...
# In this section, I re-pack the output tensor of SRCNN_network,
# the resulted new tensor is called "package_set_tensor"
# I've checked and ensure the type of "package_set_tensor" is a tensor, and the shape is (294, 5, 352, 288, 1)
#---------------------------build 3dSRnet model--------------------------------------------
SRnet_layer = SRnet_3d_model(AMOUNT, DEPTH, TARGET_HEIGHT, TARGET_WIDTH)(package_set_tensor)

#--------------------------test the result of combination----------------------------------
combined_model = Model(inputs = ip, outputs = SRnet_layer)

#I 've checked the type of SRnet_layer and ip are <class 'tensorflow.python.framework.ops.Tensor'>

编辑

#here's the lambda layer I defined:
def repacking(x):
    #----------------------------get shape from input tensor---------------------------------
    (AMOUNT, TARGET_HEIGHT, TARGET_WIDTH, tmp) = x.shape
    AMOUNT = keras.backend.shape(x)[0]#will return an empty tensor
    # or using 'list(x.shape)[0]' to return a Nontype object
    DEPTH = 5
    #AMOUNT = 26
    #----------------------------pack frame one by one---------------------------------------
    FIRST = True
    HALF_RANGE = math.floor(DEPTH/2)

    for i in range(AMOUNT):
        if (i - HALF_RANGE) < 0 or (i + HALF_RANGE) >= AMOUNT:
            AMOUNT = AMOUNT - 1
        else:
            if DEPTH%2 == 0:
                RANGE = range(i - HALF_RANGE, i + HALF_RANGE)
            else:
                RANGE = range(i - HALF_RANGE, i + HALF_RANGE + 1)

            for j in RANGE:
                frame = x[j, :, :, :]  #(352, 288, 1), type = tensor
                frame = tf.reshape(frame,(1, TARGET_HEIGHT, TARGET_WIDTH))

                if j == i - HALF_RANGE:
                    package = frame
                else:   
                    package = tf.concat([package, frame], 0)

            if FIRST == True:
                package_set = package
                FIRST = False
            else:
                package_set = tf.concat([package_set, package], 0)

    package_set = tf.reshape(package_set, (AMOUNT, DEPTH, TARGET_HEIGHT, TARGET_WIDTH, 1))   #(294, 5, 352, 288, 1)
    return package_set

但是batch_size（我命名为AMOUNT）信息不能作为for-loop索引。应该怎么做才能使用呢？

Traceback (most recent call last):
  File "main.py", line 71, in <module>
    model = combined(FILENAME, AMOUNT, DEPTH, TARGET_HEIGHT, TARGET_WIDTH)
  File "/home/user1/REUS/image-reconstruction/code/functional/model_build_up.py", line 132, in combined
    combined_model = Model(inputs = ip, outputs = SRnet_layer)
  File "/home/user1/.conda/envs/tf-cpu/lib/python3.7/site-packages/keras/legacy/interfaces.py", line 91, in wrapper
    return func(*args, **kwargs)
  File "/home/user1/.conda/envs/tf-cpu/lib/python3.7/site-packages/keras/engine/network.py", line 93, in __init__
    self._init_graph_network(*args, **kwargs)
  File "/home/user1/.conda/envs/tf-cpu/lib/python3.7/site-packages/keras/engine/network.py", line 231, in _init_graph_network
    self.inputs, self.outputs)
  File "/home/user1/.conda/envs/tf-cpu/lib/python3.7/site-packages/keras/engine/network.py", line 1366, in _map_graph_network
    tensor_index=tensor_index)
  File "/home/user1/.conda/envs/tf-cpu/lib/python3.7/site-packages/keras/engine/network.py", line 1353, in build_map
    node_index, tensor_index)
  File "/home/user1/.conda/envs/tf-cpu/lib/python3.7/site-packages/keras/engine/network.py", line 1325, in build_map
    node = layer._inbound_nodes[node_index]
AttributeError: 'NoneType' object has no attribute '_inbound_nodes'

Answer 1

Keras 层（模型表现为层）将 keras 层作为输入，而不是张量。

您需要将 Layer 参数传递给 SRnet_3d_model 而不是张量。

举个简单的例子：

inp = Input(shape=(2,))
out = Dense(1, activation='sigmoid')(inp)
model = Model(inp, out)

上面的变量inp 是一个输入层，它包含一个占位符张量，但还包含用于构建图形的附加信息。例如，当构建输出层 out 时，keras 会参考祖先层（恰好是 inp，这是 Dense.call 的参数）以确定该层的输入尺寸。

因此，在构建 keras 模型时，您需要始终根据层进行操作（模型是 Layer 的子类，可以视为复杂层）。

在某些情况下，现有层无法提供足够的功能。然后您有两个选择：Lambda 层或用户定义的层。 Lambda 层允许您以最小的开销在层内使用后端操作（或原始 tf 操作）。当您希望图层拥有自己的变量/权重时，需要自定义图层。

Lambda 层的一个简单示例是张量切片操作，例如：

Lambda(lambda x: x[:, 0])

请注意，在张量级别，您需要考虑批量维度。上面的 lambda 示例假设输入维度为 (batch_size, n_features)，并返回一个带有暗淡 (batch_size, 1) 的张量。