为什么tensorflow的image.adjust_contrast方法声称它缺少位置输入

Question

我构建了一个模型的迷你版本，我想要训练该模型，以便使用此函数进行调试：

def build_mini_model(contrast_factor=1.25):
    # Define input layer with variable input shape
    inputs = tf.keras.Input(shape=(None, None, 3))  # Accepts images of any size with 3 channels (RGB)

    # Convert RGB to grayscale
    x= tf.image.rgb_to_grayscale(inputs)

    # Enhance contrast
    x = tf.image.adjust_contrast(x, contrast_factor)

    # Create a functional model
    mini_model = Model(inputs=inputs, outputs=x)
    return mini_model

然后我创建并编译它：

mini_model = build_mini_model()
mini_model.compile(optimizer=Adam(), loss='sparse_categorical_crossentropy', 
              metrics=['accuracy'])
mini_model.summary()

并查看以下摘要：

Model: "model_3"
_________________________________________________________________
 Layer (type)                Output Shape              Param #   
=================================================================
 input_5 (InputLayer)        [(None, None, None, 3)]   0         
                                                                 
 tf.image.rgb_to_grayscale_  (None, None, None, 1)     0         
 4 (TFOpLambda)                                                  
                                                                 
 tf.image.adjust_contrast_4  (None, None, None, 1)     0         
  (TFOpLambda)                                                   
                                                                 
=================================================================
Total params: 0 (0.00 Byte)
Trainable params: 0 (0.00 Byte)
Non-trainable params: 0 (0.00 Byte)

但是，当我尝试像这样测试通过它使用图像时：

l_in = tf.random.normal((1, 32, 32, 3))
for x in range(len(mini_model.layers)):
    print(x,mini_model.layers[x].name, l_in.shape)
    l_out = mini_model.layers[x](l_in)  
    print("              ", l_out.shape)
    l_in = l_out

我得到以下输出：

0 input_5 (1, 32, 32, 3)
               (1, 32, 32, 3)
1 tf.image.rgb_to_grayscale_4 (1, 32, 32, 3)
               (1, 32, 32, 1)
2 tf.image.adjust_contrast_4 (1, 32, 32, 1)


---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
Cell In[68], line 4
      2 for x in range(len(mini_model.layers)):
      3     print(x,mini_model.layers[x].name, l_in.shape)
----> 4     l_out = mini_model.layers[x](l_in) 
      5     print("              ", l_out.shape)
      6     l_in = l_out

File ~/anaconda3/envs/IK_env/lib/python3.11/site-packages/keras/src/utils/traceback_utils.py:70, in filter_traceback.<locals>.error_handler(*args, **kwargs)
     67     filtered_tb = _process_traceback_frames(e.__traceback__)
     68     # To get the full stack trace, call:
     69     # `tf.debugging.disable_traceback_filtering()`
---> 70     raise e.with_traceback(filtered_tb) from None
     71 finally:
     72     del filtered_tb

File ~/anaconda3/envs/IK_env/lib/python3.11/site-packages/tensorflow/python/util/dispatch.py:1170, in add_dispatch_support.<locals>.decorator.<locals>.op_dispatch_handler(*args, **kwargs)
   1168 if iterable_params is not None:
   1169   args, kwargs = replace_iterable_params(args, kwargs, iterable_params)
-> 1170 result = api_dispatcher.Dispatch(args, kwargs)
   1171 if result is not NotImplemented:
   1172   return result

TypeError: Missing required positional argument

当我查看 TF_Docs 的文档时，我发现

tf.image.adjust_contrast

仅需要 2 个参数：

tf.image.adjust_contrast(
    images, contrast_factor
)

我已经给了它。

如果我逐层分解并明确调用

tf.image.adjust_contrast

，一切都很好：

l_in = tf.random.normal((1, 32, 32, 3))
l_out = model.layers[0](l_in)
print(0,model.layers[0].name, l_in.shape)
print("    ",l_out.shape)

l_in = l_out
l_out = model.layers[1](l_in)
print(1,model.layers[1].name, l_in.shape)
print("    ",l_out.shape)

l_in = l_out
print(x,model.layers[2].name, l_in.shape)
l_out = tf.image.adjust_contrast(l_in, 1.25)
print("    ",l_out.shape)

给出这个输出：

0 input_3 (1, 32, 32, 3)
     (1, 32, 32, 3)
1 tf.image.rgb_to_grayscale_2 (1, 32, 32, 3)
     (1, 32, 32, 1)
2 tf.image.adjust_contrast_2 (1, 32, 32, 1)
     (1, 32, 32, 1)

我在这里做错了什么？

Answer 1

看起来（？）你正在通过

mini_model.layers[x](l_in)

单独调用每个图层的调用方法，所以对于调整对比度它应该是

mini_model.layers[x](l_in, constrast_factor)

。

解决这个问题最简单的方法是通过子类化创建一个新层：

class AdjustContrastLayer(tf.keras.layers.Layer):
    def __init__(self, contrast_factor, **kwargs):
        super(AdjustContrastLayer, self).__init__(**kwargs)
        self.contrast_factor = contrast_factor

    def call(self, inputs):
        return tf.image.adjust_contrast(inputs, self.contrast_factor)

    def get_config(self):
        config = super(AdjustContrastLayer, self).get_config()
        config.update({"contrast_factor": self.contrast_factor})
        return config

模型结构如下所示：

def build_mini_model(contrast_factor=1.25):
    inputs = tf.keras.Input(shape=(None, None, 3))

    x= tf.image.rgb_to_grayscale(inputs)

    x = AdjustContrastLayer(contrast_factor)(x)

    mini_model = Model(inputs=inputs, outputs=x)
    return mini_model

之后你的 for 循环也应该正常执行。