无法在 Tensorflow 中使用序列模型答案

【问题标题】：Can't use Sequential Model in Tensorflow无法在 Tensorflow 中使用序列模型
【发布时间】：2022-01-12 00:20:13
【问题描述】：

这是代码：

def point_wise_feed_forward_network(d_model, dff):
  return tf.keras.Sequential([
      tf.keras.layers.Dense(dff, activation='relu'),  # (batch_size, seq_len, dff)
      tf.keras.layers.Dense(d_model)  # (batch_size, seq_len, d_model)
  ])

我在一个 phew 类中使用它，将其初始化为：

class Foo(tf.keras.layers.Layer):
   def __init__(self, d_model, dff):
      super().__init__()
      self.net = point_wise_feed_forward_network(d_model, dff)
   ...
   
   def call(self, args):
      ... # getting prev_layer (which is a tf.keras.layers.LayerNormalization() layer)
      var = self.net(prev_layer)
      ...

主要输出错误是：

ValueError: Weights for model decoder_sequential have not yet been created. Weights are created when the Model is first called on inputs or `build()` is called with an `input_shape`

File "<ipython-input-314-94b9d1a33527>", line 25, in train_step  *
        gradients = tape.gradient(loss, transformer.trainable_variables)
    File "C:\Users\User\anaconda3\envs\tfm2\lib\site-packages\keras\engine\base_layer.py", line 2308, in trainable_variables
        return self.trainable_weights
    File "C:\Users\User\anaconda3\envs\tfm2\lib\site-packages\keras\engine\training.py", line 2104, in trainable_weights
        trainable_variables += trackable_obj.trainable_variables
    File "C:\Users\User\anaconda3\envs\tfm2\lib\site-packages\keras\engine\base_layer.py", line 2308, in trainable_variables
        return self.trainable_weights
    File "C:\Users\User\anaconda3\envs\tfm2\lib\site-packages\keras\engine\base_layer.py", line 1357, in trainable_weights
        children_weights = self._gather_children_attribute('trainable_variables')
    File "C:\Users\User\anaconda3\envs\tfm2\lib\site-packages\keras\engine\base_layer.py", line 2915, in _gather_children_attribute
        return list(
    File "C:\Users\User\anaconda3\envs\tfm2\lib\site-packages\keras\engine\base_layer.py", line 2917, in <genexpr>
        getattr(layer, attribute) for layer in nested_layers))
    File "C:\Users\User\anaconda3\envs\tfm2\lib\site-packages\keras\engine\base_layer.py", line 2308, in trainable_variables
        return self.trainable_weights
    File "C:\Users\User\anaconda3\envs\tfm2\lib\site-packages\keras\engine\base_layer.py", line 1357, in trainable_weights
        children_weights = self._gather_children_attribute('trainable_variables')
    File "C:\Users\User\anaconda3\envs\tfm2\lib\site-packages\keras\engine\base_layer.py", line 2915, in _gather_children_attribute
        return list(
    File "C:\Users\User\anaconda3\envs\tfm2\lib\site-packages\keras\engine\base_layer.py", line 2917, in <genexpr>
        getattr(layer, attribute) for layer in nested_layers))
    File "C:\Users\User\anaconda3\envs\tfm2\lib\site-packages\keras\engine\base_layer.py", line 2308, in trainable_variables
        return self.trainable_weights
    File "C:\Users\User\anaconda3\envs\tfm2\lib\site-packages\keras\engine\training.py", line 2099, in trainable_weights
        self._assert_weights_created()
    File "C:\Users\User\anaconda3\envs\tfm2\lib\site-packages\keras\engine\sequential.py", line 471, in _assert_weights_created
        super(functional.Functional, self)._assert_weights_created()  # pylint: disable=bad-super-call
    File "C:\Users\User\anaconda3\envs\tfm2\lib\site-packages\keras\engine\training.py", line 2736, in _assert_weights_created
        raise ValueError(f'Weights for model {self.name} have not yet been '

所以，我在每个使用它的类中都对其进行了初始化。为什么说我没有创建模型？

PD：这个错误只有在我使用tf.GradientTape()时才会出现

PDD：I'm following this Tensorflow tutorial

【问题讨论】：

第一次密集调用不需要传递 input_shape 吗？
@MarkLavin 不，但更早的时候这已经解决了。请看我正在做的教程
不要指向教程，而是包含您自己的代码来重现问题，因为您显然在做不同的事情。
@Dr.Snoopy 很好，现在可以了。我什么都没碰，我刚刚重新启动了我的电脑。似乎这是依赖关系之间的“停电”；

标签： tensorflow tensorflow2.0

【解决方案1】：

对于模型中的第一层，您需要传入输入形状，例如 (224,224,3)

tf.keras.layers.Dense(dff, activation='relu', input_shape=(224,224,3)
```
you also need to compile your model

【讨论】：

你好，格里，谢谢你的回答。我刚刚重新启动了我的电脑，现在它可以工作了，所以它必须与依赖关系有关。