图形断开:无法获取张量张量的值(“input_2:0”，shape=(?，500,500，5，1) )

Question

def make_fcn_resnet(input_shape, nb_labels, use_pretraining, freeze_base):
    nb_rows, nb_cols, _ = input_shape
    input_tensor = Input(shape=input_shape)
    weights = 'imagenet' if use_pretraining else None

    model =ResnetBuilder.build(num_outputs=2,repetitions2=3, weights='present', input_shape=(1, 500, 500, 5))

    if freeze_base:
        for layer in model.layers:
            layer.trainable = False

    x32 = model.get_layer('act2').output
    print("x32", x32._keras_shape)
    x16 = model.get_layer('act3').output
    print("x16", x16._keras_shape)
    x8 = model.get_layer('act4').output
    print("x8", x8._keras_shape)

    c32 = Conv3D(nb_labels, (1, 1,5), name='conv_labels_32', padding='valid')(x32)
    c32=Reshape((500,500,2))(c32)
    print("c32", c32._keras_shape)
    c16 = Conv3D(nb_labels, (1, 1,5), name='conv_labels_16', padding='valid')(x16)
    c16=Reshape((250,250,2))(c16)
    print("c16", c16._keras_shape)
    c8 = Conv3D(nb_labels, (1, 1,5), name='conv_labels_8', padding='valid')(x8)
    c8=Reshape((125,125,2))(c8)
    print("c8", c8._keras_shape)

    def resize_bilinear(images):
        return tf.image.resize_bilinear(images, [nb_rows, nb_cols])

    r32 = Lambda(resize_bilinear, name='resize_labels_32')(c32)
    r16 = Lambda(resize_bilinear, name='resize_labels_16')(c16)
    r8 = Lambda(resize_bilinear, name='resize_labels_8')(c8)

    m = Add(name='merge_labels')([r32, r16, r8])

    x = Reshape((nb_rows * nb_cols, nb_labels))(m)
    x = Activation('softmax')(x)
    x = Reshape((nb_rows, nb_cols, nb_labels))(x)

    model = Model(inputs=input_tensor, outputs=x)

    #print model.summary()
    return model

在ResnetBuilder.build函数中，我使用conv3d层为Resnet编写了模型。在这里，我正在使用我自己的模型和设计新模型的Resnet层。当我调用新的模型时，我得到了错误，如图断开连接。

Answer 1

在keras(或tensorflow)中，张量之间的操作用数据流图表示。当使用数据流图时，模型(数据流图)的输入和输出之间应该有直接或间接的连接(链接)。在您的例子中，输入(input_tensor变量)和输出(X)之间没有任何联系。要解决这个问题，您应该将input_tensor与ResNet模型连接起来。Keras functional API具有调用张量上的模型的功能。

在您的情况下，请在冻结ReseNet models层之后进行以下更改。

model = model(input_tensor)
# Now model variable is just output tensor of resnet model.
x32 = model
print("x32", x32._keras_shape)
x16 = model
print("x16", x16._keras_shape)
x8 = model
print("x8", x8._keras_shape)