python - [自定义层]ValueError: tf.function-decorated 函数试图在非首次调用时创建变量

Question

问题

我正在尝试实现一个transformer encoder layer 我真正相信是导致上述错误的自定义层的自定义。除了上面的层，我还开发了 3 个更简单的自定义层，在这里我没有提到以保持整洁。

链接到代码

变压器层

class transformer(layers.Layer):
  
  def __init__(self, num_heads, transformer_layers, patch_size):
    super(transformer, self).__init__()
    self.num_heads = num_heads
    self.transformer_layers = transformer_layers
    self.patch_size = patch_size
  
  def call(self, encoded_patches):
    for _ in range(self.transformer_layers):
        x1 = layers.LayerNormalization(epsilon=1e-6)(encoded_patches)
        attention_output = layers.MultiHeadAttention(
            num_heads = self.num_heads, key_dim = projection_dim, dropout = 0.1 
        )(x1,x1)
        x2 = attention_output + encoded_patches
        x3 = layers.LayerNormalization(epsilon=1e-6)(x2)
        x3 = mlp(x3, transformer_units, 0.2)
        encoded_patches = x3 + x2

    representation = layers.LayerNormalization(epsilon=1e-6)(encoded_patches)
    resize_img = layers.Reshape([image_size // self.patch_size, image_size // self.patch_size, 64])(representation)
    
    return resize_img 

模型

def PVT():

  # Inputs
  input = layers.Input(shape=input_shape)
  augment = data_augmentation(input)

  # Stage 1
  patches_1 = Patch_1(patch_size_1)(augment)
  patches_1 = PatchEncoder(num_patches=(image_size // patch_size_1) ** 2, projection_dim=projection_dim)(patches_1)
  input_2 = transformer(num_heads, transformer_layers, patch_size_1)(patches_1) #Output 1

  # Stage 2
  patches_2 = Patch_2(patch_size_2)(input_2)
  patches_2 = PatchEncoder(num_patches=(image_size // patch_size_2) ** 2, projection_dim=projection_dim)(patches_2)
  input_3 = transformer(num_heads, transformer_layers, patch_size_2)(patches_2) #Output 2

  # Stage 3
  patches_3 = Patch_3(patch_size_3)(input_3)
  patches_3 = PatchEncoder(num_patches=(image_size // patch_size_3) ** 2, projection_dim=projection_dim)(patches_3)
  input_4 = transformer(num_heads, transformer_layers, patch_size_3)(patches_3) #Output 3

  # Stage 4
  patches_4 = Patch_4(patch_size_4)(input_4)
  patches_4 = PatchEncoder(num_patches=(image_size // patch_size_4) ** 2, projection_dim=projection_dim)(patches_4)
  input_5 = transformer(num_heads, transformer_layers, patch_size_4)(patches_4) #Output 4

  representation = layers.Flatten()(input_5)
  representation = layers.Dropout(0.5)(representation)
  # Classify outputs.
  logits = layers.Dense(num_classes)(representation)
  # Create the Keras model.  
  model = keras.Model(inputs=input, outputs=logits)

  return model

编译器和优化器

def run_experiment(model):

    optimizer = tfa.optimizers.AdamW(
        learning_rate=learning_rate, weight_decay=weight_decay, beta_1=0.9, beta_2=0.999
    )

    model.compile(
        optimizer=optimizer,
        loss=keras.losses.SparseCategoricalCrossentropy(from_logits=True),
        metrics=[
            keras.metrics.SparseCategoricalAccuracy(name="accuracy"),
            keras.metrics.SparseTopKCategoricalAccuracy(5, name="top-5-accuracy"),
        ],
    )


    history = model.fit(
        x=xtrain,
        y=ytrain,
        batch_size=batch_size,
        epochs=5,
        validation_split=0.1
    )

    model.save('model-5.h5')

    return history

pvt = PVT()
history = run_experiment(pvt)

笔记

我已经检查了几个关于这个的来源，但我仍然对理解这个错误感到困惑。在您将我引向其他来源之前，我向您保证，我已经全部检查过了。因此，我真诚地请求您在这里提供一个干净的解决方案。

链接到代码

模型摘要（仅供参考）

Model: "model_2"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
input_3 (InputLayer)         [(None, 32, 32, 3)]       0         
_________________________________________________________________
data_augmentation (Sequentia (None, 72, 72, 3)         7         
_________________________________________________________________
patch_1_2 (Patch_1)          (None, None, 48)          0         
_________________________________________________________________
patch_encoder_8 (PatchEncode (None, 324, 64)           23872     
_________________________________________________________________
transformer_8 (transformer)  (None, 18, 18, 64)        0         
_________________________________________________________________
patch_2_2 (Patch_2)          (None, None, 4096)        0         
_________________________________________________________________
patch_encoder_9 (PatchEncode (None, 81, 64)            267392    
_________________________________________________________________
transformer_9 (transformer)  (None, 9, 9, 64)          0         
_________________________________________________________________
patch_3_2 (Patch_3)          (None, None, 16384)       0         
_________________________________________________________________
patch_encoder_10 (PatchEncod (None, 16, 64)            1049664   
_________________________________________________________________
transformer_10 (transformer) (None, 4, 4, 64)          0         
_________________________________________________________________
patch_4_2 (Patch_4)          (None, None, 65536)       0         
_________________________________________________________________
patch_encoder_11 (PatchEncod (None, 4, 64)             4194624   
_________________________________________________________________
transformer_11 (transformer) (None, 2, 2, 64)          0         
=================================================================
Total params: 5,535,559
Trainable params: 5,535,552
Non-trainable params: 7

标签： pythontensorflowmachine-learningkerasdeep-learning