Keras 图像分类 - “logits 和标签必须具有相同的第一维”错误

Question

我正在学习和使用 Keras 进行图像分类项目。 运行环境是Google Colab，最近报错：

File "/usr/local/lib/python3.10/dist-packages/keras/backend.py", line 5660, in sparse_categorical_crossentropy
      res = tf.nn.sparse_softmax_cross_entropy_with_logits(
Node: 'sparse_categorical_crossentropy/SparseSoftmaxCrossEntropyWithLogits/SparseSoftmaxCrossEntropyWithLogits'
logits and labels must have the same first dimension, got logits shape [32,2048] and labels shape [1248]
     [[{{node sparse_categorical_crossentropy/SparseSoftmaxCrossEntropyWithLogits/SparseSoftmaxCrossEntropyWithLogits}}]] [Op:__inference_train_function_26557]

到目前为止，我的项目中的代码行是（最后一行是给我错误的那一行）：

import os
from pydrive.auth import GoogleAuth
from pydrive.drive import GoogleDrive
from google.colab import auth
from oauth2client.client import GoogleCredentials
from tensorflow import keras

auth.authenticate_user()
gauth = GoogleAuth()
gauth.credentials = GoogleCredentials.get_application_default()
drive = GoogleDrive(gauth)

file = drive.CreateFile({'id':'1D8Of3b0WdQHQnOrCWPqoHn-lSDZxkNYS'})
file.GetContentFile('Augmented_PlantVillage_Dataset.zip')

!unzip Augmented_PlantVillage_Dataset.zip

keras.preprocessing.image.load_img('Augmented_PlantVillage_Dataset/train/Apple___Apple_scab/image (4).JPG')

LABELS = 'inferred'
LABEL_MODE = 'categorical'
BATCH_SIZE = 32
IMAGE_SIZE = (256, 256)

training_dataset = keras.utils.image_dataset_from_directory(
    directory = 'Augmented_PlantVillage_Dataset/train',
    labels = LABELS,
    label_mode = LABEL_MODE,
    batch_size = BATCH_SIZE,
    image_size = IMAGE_SIZE
)

validation_dataset = keras.utils.image_dataset_from_directory(
    directory = 'Augmented_PlantVillage_Dataset/val',
    labels = LABELS,
    label_mode = LABEL_MODE,
    batch_size = BATCH_SIZE,
    image_size = IMAGE_SIZE
)

test_dataset = keras.utils.image_dataset_from_directory(
    directory = 'Augmented_PlantVillage_Dataset/test',
    labels = LABELS,
    label_mode = LABEL_MODE,
    batch_size = BATCH_SIZE,
    image_size = IMAGE_SIZE
)

training_dataset.class_names

training_dataset

WEIGHTS = 'imagenet'
INPUT_SHAPE = (256, 256, 3)
POOLING = 'max'
CLASSES = 39

model = keras.applications.ResNet50(
    include_top = False,
    weights = WEIGHTS,
    input_shape = INPUT_SHAPE,
    pooling = POOLING,
    classes = CLASSES
)

model.summary()

model

keras.utils.plot_model(model, 'model.png', show_shapes = True)

LOSS = keras.losses.sparse_categorical_crossentropy
OPTIMIZER = keras.optimizers.Adam()
METRICS = ['accuracy']

model.compile(
    loss = LOSS,
    optimizer = OPTIMIZER,
    metrics = METRICS
)

EPOCHS = 10
CALLBACKS = [
    keras.callbacks.ModelCheckpoint("save_at_{epoch}.keras"),
]

model.fit(
    training_dataset,
    epochs = EPOCHS,
    callbacks = CALLBACKS,
    validation_data = validation_dataset
)

我一直试图了解变量的形状有什么问题，但我找不到错误。

此外，我希望那些在以下主题方面有经验的人：迁移学习、微调和图像分类，向我推荐他们认为有用的任何内容和资源，并为我提供有关如何学习和理解我在 Keras 中使用的功能的提示。我意识到这是一个很大的话题。

我添加了 model 的 plot 作为请求

Answer 1

ResNet50 是预训练的 model 用于特征提取。 keras.applications.ResNet50这个 model 的 output 是 (None,2048)。

您需要从此基础 model ResNet50 中获取特征，然后应用GlobalAveragePooling2D()

base_model = tf.keras.applications.ResNet50(weights='imagenet', input_shape=(256,256,3), include_top=False)

base_model.trainable = True

model.add(layers.GlobalAveragePooling2D())
model.add(layers.Dropout(0.5))

model.add(layers.Dense(2048, activation='relu'))
model.add(layers.Dense(512, activation='relu', kernel_regularizer=tf.keras.regularizers.L2(0.001)))
model.add(layers.Dropout(0.25))
model.add(layers.Dense(512, activation='relu', kernel_regularizer=tf.keras.regularizers.L2(0.001)))
model.add(layers.Dropout(0.25))
model.add(layers.Dense(CLASSES, activation='softmax'))

model.summary()

注意：如果通道 1 为黑白图像，则无法使用 imag.net weight。

base_model = tf.keras.applications.ResNet50(weights=None, input_shape=(256,256,1), include_top=False)