在 Tensorflow 2.0 中使用层子类化在我的自定义层中出现此错误“必须始终传递 `Layer.call` 的第一个参数。”

Question

I've made this custom layer using layer using Tensorflow 2.0 layer subclassing. I'm trying to make a layer of residual block. But when I add this custom layer in my model through sequential API I'm getting the below error.

class ResidualBlock(Layer):

    def __init__(self, **kwargs):
        super(ResidualBlock, self).__init__(**kwargs)

    def build(self, input_shape):
        """
        This method should build the layers according to the above specification. Make sure 
        to use the input_shape argument to get the correct number of filters, and to set the
        input_shape of the first layer in the block.
        """
        self.bn_1 = BatchNormalization(input_shape=input_shape)
        self.conv_1 = Conv2D( input_shape[0],(3,3), padding='SAME')
        self.bn_2 = BatchNormalization()
        self.conv_2 = Conv2D( input_shape[0],(3,3), padding='SAME')






    def call(self, inputs, training=False):
        """
        This method should contain the code for calling the layer according to the above
        specification, using the layer objects set up in the build method.
        """
        h = self.bn_1(training=True)(inputs)
        h = tf.nn.relu(h)
        h = self.conv_1(h)
        h = self.bn_2(training=True)(h)
        h = tf.nn.relu(h)
        h = self.conv_2(h)
        return Add(inputs, h)

But when I initialize this layer I'm getting the error.

test_model = tf.keras.Sequential([ResidualBlock(input_shape=(28, 28, 1), name="residual_block")])
test_model.summary()

My error logs:

---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
<ipython-input-13-991ed1d78e4b> in <module>()
      1 # Test your custom layer - the following should create a model using your layer
      2 
----> 3 test_model = tf.keras.Sequential([ResidualBlock(input_shape=(28, 28, 1), name="residual_block")])
      4 test_model.summary()

5 frames
/usr/local/lib/python3.6/dist-packages/tensorflow/python/autograph/impl/api.py in wrapper(*args, **kwargs)
    263       except Exception as e:  # pylint:disable=broad-except
    264         if hasattr(e, 'ag_error_metadata'):
--> 265           raise e.ag_error_metadata.to_exception(e)
    266         else:
    267           raise

ValueError: in user code:

    <ipython-input-12-3beea3ca10b0>:32 call  *
        h = self.bn_1(training=True)(inputs)
    /usr/local/lib/python3.6/dist-packages/tensorflow/python/keras/engine/base_layer.py:800 __call__  **
        'The first argument to `Layer.call` must always be passed.')

    ValueError: The first argument to `Layer.call` must always be passed.

Answer 1

During the call method change the forward pass of batch norms to h=self.bn_1(inputs) . Since you are passing training=True for the whole layer tensorflow will automatically take care of maintaining the same flag for all of it's sublayers and you don't need to pass it explicitly for each of them. But if your application is such that you want to control batch norm differently compared to other layers use h=self.bn_1(inputs, training=True) . And your final return statement is not in correct format and should be like Add()([inputs, h])

class ResidualBlock(Layer):

    def __init__(self, **kwargs):
        super(ResidualBlock, self).__init__(**kwargs)

    def build(self, input_shape):
        """
        This method should build the layers according to the above specification. Make sure 
        to use the input_shape argument to get the correct number of filters, and to set the
        input_shape of the first layer in the block.
        """
        self.bn_1 = BatchNormalization(input_shape=input_shape)
        self.conv_1 = Conv2D(input_shape[3],(3,3), padding='SAME')
        self.bn_2 = BatchNormalization()
        self.conv_2 = Conv2D(input_shape[3],(3,3), padding='SAME')

    def call(self, inputs, training=False):
        """
        This method should contain the code for calling the layer according to the above
        specification, using the layer objects set up in the build method.
        """
        h = self.bn_1(inputs)
        h = tf.nn.relu(h)
        h = self.conv_1(h)
        h = self.bn_2(h)
        h = tf.nn.relu(h)
        h = self.conv_2(h)
        return Add()([inputs, h])

pk = ResidualBlock()
model = tf.keras.Sequential([pk])
model(tf.ones((1, 28, 28, 3)))

So once your model is called with example input of tf.ones(), build will be called to create the batch norm and convolution layers. For conv layer you are using the number of filters same as the input by indexing into the last dimension