神经网络模型的训练输入得到错误的形状 ((预期形状=(None, 222, 222, 3), 发现形状=(None, None, 224, 224, 3) ))

Question

所以我正在尝试使用 Oxford 102 花数据集训练神经网络模型。 但由于某种未知的原因，我在训练时使用 model.fit 出现错误，以下是与代码相关的部分：

# Define the path to the dataset directory
# dataset_dir = "/content/drive/MyDrive/jpg"

# Define batch size and image size
batch_size = 32
img_size = (224, 224)

# Use image_dataset_from_directory to load the dataset
train_dataset = tf.keras.utils.image_dataset_from_directory(
    dataset_dir,
    labels="inferred",
    label_mode="categorical",  # Change to "binary" for binary classification
    image_size=img_size,
    batch_size=batch_size,
    validation_split=0.2,
    subset="training",
    seed=1337,
)

valid_dataset = tf.keras.utils.image_dataset_from_directory(
    dataset_dir,
    labels="inferred",
    label_mode="categorical",
    image_size=img_size,
    batch_size=batch_size,
    validation_split=0.2,
    subset="validation",
    seed=1337,
)


train_dataset = train_dataset.map(lambda x, y: (x / 255.0, y))  # Normalize pixel values for Train dataset
valid_dataset = valid_dataset.map(lambda x, y: (x / 255.0, y))  # Normalize pixel values for Validation dataset


# Batch the datasets
train_dataset = train_dataset.batch(batch_size)
valid_dataset = valid_dataset.batch(batch_size)

这是该单元格的输出：

Found 8189 files belonging to 1 classes.
Using 6552 files for training.
Found 8189 files belonging to 1 classes.
Using 1637 files for validation.

使用此模型：

# Model ( Sequential Model )
model = models.Sequential()

# Convolutional base ( Using Convolutional Neural network to extract information from images)
model.add(layers.Conv2D(32, (3, 3), activation='relu', input_shape=(224, 224, 3)))
model.add(layers.MaxPooling2D((2, 2)))
model.add(layers.Conv2D(64, (3, 3), activation='relu'))
model.add(layers.MaxPooling2D((2, 2)))
model.add(layers.Conv2D(128, (3, 3), activation='relu'))
model.add(layers.MaxPooling2D((2, 2)))

# Flatten the convolutional base
model.add(layers.Flatten())

# Dense layers to extract features from CNN output
# We use a Dropout of 50% to prevent the model from over-fitting and for regularization
model.add(layers.Dense(256, activation='relu'))
model.add(layers.Dropout(0.5))
model.add(layers.Dense(102, activation='softmax'))  # 102 classes for the flowers dataset

# Compile the model
model.compile(optimizer='adam', loss='categorical_crossentropy', metrics=['accuracy'])

# Display the model summary
model.summary()

模型总结输出：

_________________________________________________________________
 Layer (type)                Output Shape              Param #   
=================================================================
 conv2d_6 (Conv2D)           (None, 222, 222, 32)      896       
                                                                 
 max_pooling2d_6 (MaxPoolin  (None, 111, 111, 32)      0         
 g2D)                                                            
                                                                 
 conv2d_7 (Conv2D)           (None, 109, 109, 64)      18496     
                                                                 
 max_pooling2d_7 (MaxPoolin  (None, 54, 54, 64)        0         
 g2D)                                                            
                                                                 
 conv2d_8 (Conv2D)           (None, 52, 52, 128)       73856     
                                                                 
 max_pooling2d_8 (MaxPoolin  (None, 26, 26, 128)       0         
 g2D)                                                            
                                                                 
 flatten_2 (Flatten)         (None, 86528)             0         
                                                                 
 dense_4 (Dense)             (None, 256)               22151424  
                                                                 
 dropout_2 (Dropout)         (None, 256)               0         
                                                                 
 dense_5 (Dense)             (None, 102)               26214     
                                                                 
=================================================================
Total params: 22270886 (84.96 MB)
Trainable params: 22270886 (84.96 MB)
Non-trainable params: 0 (0.00 Byte)
_________________________________________________________________

这个列车单元将导致以下错误：

epochs = 10

Train = model.fit(
    train_dataset,
    epochs = epochs,
    validation_data=valid_dataset
)

运行训练单元后得到的错误：

ValueError: Input 0 of layer "sequential" is incompatible with the layer: expected shape=(None, 222, 222, 3), found shape=(None, None, 224, 224, 3) 我看过一些类似的帖子引用相同的问题，但没有一个得到正确或体面的答案，希望这个问题得到解决，以便人们稍后使用它。

Answer 1

你应该在第一个卷积层中

layers.Conv2D(..., padding = 'same')

。该层的默认值为“有效”。 （参考 -> 

https://keras.io/api/layers/convolution_layers/convolution2d/）

_________________________________________________________________
 Layer (type)                Output Shape              Param #   
=================================================================
 conv2d_22 (Conv2D)          (None, 224, 224, 32)      896       
                                                                 
 max_pooling2d_21 (MaxPooli  (None, 112, 112, 32)      0         
 ng2D)