Keras/Tensorflow - 无法使 model.fit() 工作

Question

我正在尝试制作一个 CNN 网络来对蘑菇图像进行预测。

遗憾的是，我什至无法开始训练我的模型，fit() 方法总是给我错误。

有 10 个类，tf 数据集根据它们的子文件夹正确找到了它们的名称。

使用我当前的代码，它说：

InvalidArgumentError:  logits and labels must have the same first
dimension, got logits shape [12800,10] and labels shape [32]

模型总结：

 Layer (type)                Output Shape              Param #   
=================================================================  
input_5 (InputLayer)        [(None, 64, 64, 3)]       0         
                                                                           
conv2d_4 (Conv2D)           (None, 62, 62, 32)        896       
                                                                           
max_pooling2d_2 (MaxPooling  (None, 20, 20, 32)       0              
2D) 
    
                                                                       
re_lu_2 (ReLU)              (None, 20, 20, 32)        0         
                                                                       
dense_2 (Dense)             (None, 20, 20, 10)        330 

  
                                                                 
=================================================================

这是我的代码：

#Data loading
train_set = keras.preprocessing.image_dataset_from_directory(
      data_path, 
      labels="inferred",
      label_mode="int",
      batch_size=32, 
      image_size=(64, 64),
      shuffle=True,
      seed=1446,
      validation_split = 0.2,
      subset="training")

validation_set = keras.preprocessing.image_dataset_from_directory(
  data_path, 
  labels="inferred",
  label_mode="int",
  batch_size=32, 
  image_size=(64, 64),
  shuffle=True,
  seed=1446,
  validation_split = 0.2,
  subset="validation")

#Constructing layers
input_layer = keras.Input(shape=(64, 64, 3))
x = layers.Conv2D(filters=32, kernel_size=(3, 3), activation="relu")(input_layer)
x = layers.MaxPooling2D(pool_size=(3, 3))(x)
x = keras.layers.ReLU()(x)
output = layers.Dense(10, activation="softmax")(x)

#Making and fitting the model
model = keras.Model(inputs=input_layer, outputs=output)
model.compile(optimizer="adam", loss="sparse_categorical_crossentropy", metrics=['accuracy'])
model.fit(train_set, epochs=5, validation_data=validation_set)

Answer 1

我认为您需要在传递到Dense 层之前进行展平

input_layer = keras.Input(shape=(64, 64, 3))
x = layers.Conv2D(filters=32, kernel_size=(3, 3), activation="relu")(input_layer)
x = layers.MaxPooling2D(pool_size=(3, 3))(x)
x = keras.layers.ReLU()(x)
x = keras.layers.Flatten()(x) # try adding this
output = layers.Dense(10, activation="softmax")(x)

Answer 2

你需要做的是在你的模型中的 ReLU 层和输出层之间添加一个 flatten 层。

input_layer = keras.Input(shape=(64, 64, 3))
x = layers.Conv2D(filters=32, kernel_size=(3, 3), activation="relu")(input_layer)
x = layers.MaxPooling2D(pool_size=(3, 3))(x)
x = keras.layers.ReLU()(x)
x = keras.layers.Flatten()(x)
output = layers.Dense(10, activation="softmax")(x)

当您看到 model.fit 由于 logits 和标签的差异而引发错误时，最好打印出模型摘要

print(model.summary())

通过查看摘要通常有助于找出问题所在。