使用Conv2D构建6分类图像识别模型遇形状不兼容错误求助

我尝试用Conv2D构建6类图像识别模型，多次测试都出现错误：

ValueError: Shapes (None, None) and (None, 296, 296, 6) are incompatible

我的代码实现如下：

from sklearn.preprocessing import normalize
im_shape = (300, 300)

val_data_generator = ImageDataGenerator(preprocessing_function=normalize,validation_split=0.2)

train_generator = data_generator.flow_from_directory(TRAINING_DIR, target_size=im_shape, shuffle=True,
                                                class_mode='categorical', batch_size=BATCH_SIZE, subset="training")

validation_generator = val_data_generator.flow_from_directory(TRAINING_DIR, target_size=im_shape, shuffle=False,
                                                class_mode='categorical', batch_size=BATCH_SIZE, subset="validation")


test_generator = ImageDataGenerator(preprocessing_function=normalize)
test_generator = test_generator.flow_from_directory(TEST_DIR, target_size=im_shape, shuffle=False,
                                                class_mode='categorical', batch_size=BATCH_SIZE)
nb_train_samples = train_generator.samples
nb_validation_samples = validation_generator.samples
nb_test_samples = test_generator.samples
num_classes  = 6

model = models.Sequential()
model.add(layers.Conv2D(32, (3, 3),activation="elu", kernel_initializer="glorot_uniform", input_shape=(300, 300, 3)))
model.add(layers.Dropout(0.2))
model.add(layers.BatchNormalization())
model.add(layers.Conv2D(64, (3, 3), activation="elu", kernel_initializer="glorot_uniform"))
model.add(layers.Dropout(0.2))
model.add(layers.BatchNormalization())
model.add(layers.Dense(100, activation="elu", kernel_initializer="he_normal"))
model.add(layers.Dropout(0.2))
model.add(layers.BatchNormalization())
model.add(layers.Dense(100, activation="elu", kernel_initializer="he_normal"))
model.add(layers.Dropout(0.2))
model.add(layers.BatchNormalization())
model.add(layers.Dense(6, activation="softmax", kernel_initializer="glorot_uniform"))

model.compile(optimizer=Adam(), loss='categorical_crossentropy',metrics=['accuracy'])
history = model.fit(
train_generator,
steps_per_epoch=nb_train_samples // BATCH_SIZE,
epochs=150,
callbacks = [keras.callbacks.EarlyStopping(monitor='val_loss', patience=30,verbose=1)],
validation_data=validation_generator,
verbose = 1,
validation_steps=nb_validation_samples // BATCH_SIZE)

核心问题是Conv2D层输出的4D特征图直接连接Dense层，未做展平处理：

• Conv2D层输出的是(批量大小, 高, 宽, 通道数)的4D张量，比如第一个Conv2D后输出(None, 296, 296, 32)，第二个Conv2D后是(None, 294, 294, 64)（3x3卷积无填充时，尺寸会减2）。
• Dense层默认接收(批量大小, 特征数)的2D张量，两者维度不匹配，导致形状冲突报错。

此外还有几个次要问题：

1. train_generator使用了未定义的data_generator变量；
2. 用sklearn.preprocessing.normalize做图像预处理不合适，该函数默认按样本归一化，会破坏图像空间结构；
3. 缺少池化层，特征图尺寸过大，计算量极高且容易过拟合；
4. 未导入Adam优化器和keras相关模块。

关键修复点： #

• 在Conv2D层和第一个Dense层之间添加Flatten()层，将4D特征图展平为2D向量；
• 每个Conv2D层后添加MaxPooling2D()层，缩小特征图尺寸，降低计算量；
• 替换图像预处理方式，改用rescale=1./255做归一化；
• 补全所有缺失的导入，修正未定义变量问题。

修改后的完整代码： #

from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import Conv2D, Dropout, BatchNormalization, Dense, Flatten, MaxPooling2D
from tensorflow.keras.optimizers import Adam
from tensorflow.keras.preprocessing.image import ImageDataGenerator
from tensorflow.keras.callbacks import EarlyStopping

# 定义参数
im_shape = (300, 300)
BATCH_SIZE = 32  # 根据硬件配置调整
TRAINING_DIR = "你的训练集路径"
TEST_DIR = "你的测试集路径"

# 数据生成器
data_generator = ImageDataGenerator(rescale=1./255, validation_split=0.2)

train_generator = data_generator.flow_from_directory(
    TRAINING_DIR,
    target_size=im_shape,
    shuffle=True,
    class_mode='categorical',
    batch_size=BATCH_SIZE,
    subset="training"
)

validation_generator = data_generator.flow_from_directory(
    TRAINING_DIR,
    target_size=im_shape,
    shuffle=False,
    class_mode='categorical',
    batch_size=BATCH_SIZE,
    subset="validation"
)

test_generator = ImageDataGenerator(rescale=1./255)
test_generator = test_generator.flow_from_directory(
    TEST_DIR,
    target_size=im_shape,
    shuffle=False,
    class_mode='categorical',
    batch_size=BATCH_SIZE
)

nb_train_samples = train_generator.samples
nb_validation_samples = validation_generator.samples
nb_test_samples = test_generator.samples
num_classes = 6

# 构建模型
model = Sequential()
model.add(Conv2D(32, (3, 3), activation="elu", kernel_initializer="glorot_uniform", input_shape=(300, 300, 3)))
model.add(MaxPooling2D(pool_size=(2, 2)))  # 添加池化层
model.add(Dropout(0.2))
model.add(BatchNormalization())

model.add(Conv2D(64, (3, 3), activation="elu", kernel_initializer="glorot_uniform"))
model.add(MaxPooling2D(pool_size=(2, 2)))  # 添加池化层
model.add(Dropout(0.2))
model.add(BatchNormalization())

model.add(Flatten())  # 关键：展平4D特征图为2D向量
model.add(Dense(100, activation="elu", kernel_initializer="he_normal"))
model.add(Dropout(0.2))
model.add(BatchNormalization())

model.add(Dense(100, activation="elu", kernel_initializer="he_normal"))
model.add(Dropout(0.2))
model.add(BatchNormalization())

model.add(Dense(num_classes, activation="softmax", kernel_initializer="glorot_uniform"))

# 编译模型
model.compile(optimizer=Adam(learning_rate=0.001), loss='categorical_crossentropy', metrics=['accuracy'])

# 训练模型
history = model.fit(
    train_generator,
    steps_per_epoch=nb_train_samples // BATCH_SIZE,
    epochs=150,
    callbacks=[EarlyStopping(monitor='val_loss', patience=30, verbose=1)],
    validation_data=validation_generator,
    verbose=1,
    validation_steps=nb_validation_samples // BATCH_SIZE
)

内容的提问来源于stack exchange，提问作者I'mStuckOnLine911