使用VGG16预训练模型训练CIFAR10时遇形状不匹配错误求助

错误原因 #

核心问题是损失函数与任务类型不匹配：

• CIFAR10是10分类任务，你已经通过np_utils.to_categorical将标签转换为One-Hot编码（形状为(None,10)），输出层也使用softmax激活函数输出多分类概率。
• 但你在model.compile中使用了binary_crossentropy（二分类专用损失函数），该函数默认期望标签形状为(None,1)，因此导致模型输出（(None,10)）与标签（(None,10)）在损失计算时出现形状不匹配的冲突。

解决步骤 #

只需将损失函数替换为多分类任务对应的categorical_crossentropy即可：

model.compile(loss='categorical_crossentropy',
              optimizer=optimizers.Adam(lr=LEARNING_RATE),
              metrics=['accuracy'])

额外性能优化提示（非当前错误，但影响模型效果） #

VGG16预训练模型基于224×224尺寸的ImageNet数据集训练，而CIFAR10图片尺寸仅为32×32。虽然代码中指定input_shape可以正常运行，但小尺寸输入会限制VGG16的特征提取能力。建议在输入模型前将图片resize到224×224：

from tensorflow.keras.preprocessing.image import img_to_array, array_to_img

x_train = np.array([img_to_array(array_to_img(img).resize((224, 224))) for img in x_train])
x_test = np.array([img_to_array(array_to_img(img).resize((224, 224))) for img in x_test])

修正后的完整代码 #

import numpy as np
from tensorflow import keras
from tensorflow.keras import layers, optimizers
from tensorflow.keras.applications import VGG16
from tensorflow.keras.models import Model
from tensorflow.keras.utils import to_categorical

(x_train, y_train), (x_test, y_test) = keras.datasets.cifar10.load_data()

# 可选：将图片resize到VGG16期望的224×224尺寸
from tensorflow.keras.preprocessing.image import img_to_array, array_to_img
x_train = np.array([img_to_array(array_to_img(img).resize((224, 224))) for img in x_train])
x_test = np.array([img_to_array(array_to_img(img).resize((224, 224))) for img in x_test])

height = x_train.shape[1]
width = x_train.shape[2]
channel = x_train.shape[3]

BATCH_SIZE = 32
EPOCHS = 2
NUM_CLASSES = 10
LEARNING_RATE = 1e-4

y_train = to_categorical(y_train, NUM_CLASSES)
y_test = to_categorical(y_test, NUM_CLASSES)

base_model_VGG16 = VGG16(weights='imagenet', include_top=False, input_shape=(height, width, channel))

for layer in base_model_VGG16.layers:
    layer.trainable = False

# 构建自定义分类头
x = base_model_VGG16.output
x = layers.GlobalAveragePooling2D()(x)
x = layers.BatchNormalization()(x)
x = layers.Dense(256, activation='relu')(x)
x = layers.Dense(256, activation='relu')(x)
x = layers.Dropout(0.6)(x)
predictions = layers.Dense(NUM_CLASSES, activation='softmax')(x)

model = Model(inputs=base_model_VGG16.input, outputs=predictions)

# 使用正确的多分类损失函数
model.compile(loss='categorical_crossentropy',
              optimizer=optimizers.Adam(learning_rate=LEARNING_RATE),
              metrics=['accuracy'])

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

model.fit(
    x_train, y_train,
    epochs=EPOCHS,
    callbacks=callbacks,
    batch_size=BATCH_SIZE
)

内容的提问来源于stack exchange，提问作者MD Rafsun Sheikh