python - Keras 将功能模型转换为模型子类

Question

我尝试使用 Keras 模型子类化来重写函数模型，但在新的模型子类中，摘要生成不起作用。

作为引用，这里是功能模型及其输出。

filters = 32

# placeholder for inputs
inputs = Input(shape=[16, 16, 16, 12])  

# L-hand side of UNet
conv1 = DoubleConv3D(filters*1)(inputs)
pool1 = MaxPooling3D()(conv1)
...

# middle bottleneck
conv5 = DoubleConv3D(filters*5)(pool4)

# R-hand side of UNet
rsdc6 = ConcatConv3D(filters*4)(conv5, conv4)
conv6 = DoubleConv3D(filters*4)(rsdc6)
...

# sigmoid activation
outputs = Conv3D(1, (1, 1, 1), activation='sigmoid')(conv9)

model = Model(inputs=[inputs], outputs=[outputs])
model.summary()

Layer (type)                    Output Shape         Param #     Connected to                     
==================================================================================================
input_29 (InputLayer)           (None, 16, 16, 16, 1 0                                            
__________________________________________________________________________________________________
conv3d_111 (Conv3D)             (None, 16, 16, 16, 3 10400       input_29[0][0]                   
__________________________________________________________________________________________________
...

模型子类看起来像:

class UNet3D(Model):
    def __init__(self, **kwargs):
        super(UNet3D, self).__init__(name="UNet3D", **kwargs)        
        self.filters = 32

    def __call__(self, inputs):

        # L-hand side of UNet
        conv1 = DoubleConv3D(self.filters*1)(inputs)
        pool1 = MaxPooling3D()(conv1)
        ...

        # middle bottleneck
        conv5 = DoubleConv3D(self.filters*5)(pool4)

        # R-hand side of UNet
        rsdc6 = ConcatConv3D(self.filters*4)(conv5, conv4)
        conv6 = DoubleConv3D(self.filters*4)(rsdc6)
        ...

        # sigmoid activation
        outputs = Conv3D(1, (1, 1, 1), activation='sigmoid')(conv9)
        return outputs

unet3d = UNet3D()
unet3d.build(Input(shape=[None, None, None, 1]))
unet3d.summary()

但是，摘要没有输出层数和参数数量，而是给出了

_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
Total params: 0
Trainable params: 0
Non-trainable params: 0
_________________________________________________________________

最初，我认为这是在调用摘要之前未调用 build 的错误，并尝试显式调用该函数并在第一个卷积层之前添加 InputLayer ，如本related answer中所述。但是，这两种解决方案都有助于修复模型子类上的摘要生成。

Answer 1

通过查看以下内容，我找到了此模型子类化问题的解决方案 example 。功劳应归于该存储库的作者。

将 Keras 函数创建转换为模型子类的一种方法是创建并调用一个复制模型初始化的函数，例如模型(输入=[输入]，输出=[输出])。在这里，我们使用 _build 函数来完成此操作。

class UNet3D(Model):
    def __init__(self, **kwargs):

        # Initialize model parameters.
        self.filters = 32
        ...

        # Initialize model.
        self._build(**kwargs)

    def __call__(self, inputs):

        # L-hand side of UNet
        conv1 = DoubleConv3D(self.filters*1)(inputs)
        pool1 = MaxPooling3D()(conv1)
        ...

        # middle bottleneck
        conv5 = DoubleConv3D(self.filters*5)(pool4)

        # R-hand side of UNet
        rsdc6 = ConcatConv3D(self.filters*4)(conv5, conv4)
        conv6 = DoubleConv3D(self.filters*4)(rsdc6)
        ...

        # sigmoid activation
        outputs = Conv3D(1, (1, 1, 1), activation='sigmoid')(conv9)
        return outputs

    def _build(self, **kwargs):
        """
        Replicates Model(inputs=[inputs], outputs=[outputs]) of functional model.
        """
        # Replace with shape=[None, None, None, 1] if input_shape is unknown.
        inputs  = Input(shape=[16, 16, 16, 12])
        outputs = self.__call__(inputs)
        super(UNet3D, self).__init__(name="UNet3D", inputs=inputs, outputs=outputs, **kwargs) 

unet3d = UNet3D()
unet3d.summary()