Keras多图像输入模型报错max_pooling2d期望4维输入收到5维如何解决

报错根因 #

MaxPooling2D层要求输入为4维张量，维度顺序为[batch_size, height, width, channels]，你传入的5维张量多出来的大小为4的维度，是4张输入图像的堆叠维度，导致维度匹配失败。

解决方案 #

方案1：调整输入数据格式（最快修复） #

你代码中定义了4个独立的输入层，训练/推理时不要将4张图像提前堆叠为[batch_size, 4, IMG_HEIGHT, IMG_WIDTH, IMG_CHANNELS]的5维张量传入，需要拆分为4个独立的4维张量，按模型定义的输入顺序传入即可：

# 假设原始堆叠数据集为all_inputs，形状为 [样本数, 4, H, W, C]
input_main = all_inputs[:,0,...] # 对应你代码中最后定义的inputs输入
input_1 = all_inputs[:,1,...] # 对应inputs1
input_2 = all_inputs[:,2,...] # 对应inputs2
input_3 = all_inputs[:,3,...] # 对应inputs3

# 训练时按顺序传入
model.fit([input_main, input_1, input_2, input_3], train_label, ......)

方案2：优化代码结构（可选，避免重复冗余） #

你当前代码中4个特征提取分支结构完全一致，重复代码较多易出错，可以封装为独立函数，同时支持直接接收5维堆叠输入，自动拆分处理：

# 封装通用特征提取分支
def build_feat_extractor():
    inputs = tf.keras.layers.Input((IMG_HEIGHT, IMG_WIDTH, IMG_CHANNELS))
    o = tf.keras.layers.Lambda(lambda x: x / 255)(inputs)
    c = tf.keras.layers.Conv2D(16, (3, 3), activation='relu', kernel_initializer='he_normal', padding='same')(o)
    c = tf.keras.layers.Dropout(0.1)(c)
    c = tf.keras.layers.Conv2D(16, (3, 3), activation='relu', kernel_initializer='he_normal', padding='same')(c)
    p = tf.keras.layers.MaxPooling2D((2, 2))(c)

    c = tf.keras.layers.Conv2D(32, (3, 3), activation='relu', kernel_initializer='he_normal', padding='same')(p)
    c = tf.keras.layers.Dropout(0.1)(c)
    c = tf.keras.layers.Conv2D(32, (3, 3), activation='relu', kernel_initializer='he_normal', padding='same')(c)
    p = tf.keras.layers.MaxPooling2D((2, 2))(c)

    c = tf.keras.layers.Conv2D(64, (3, 3), activation='relu', kernel_initializer='he_normal', padding='same')(p)
    c = tf.keras.layers.Dropout(0.2)(c)
    c = tf.keras.layers.Conv2D(64, (3, 3), activation='relu', kernel_initializer='he_normal', padding='same')(c)
    p = tf.keras.layers.MaxPooling2D((2, 2))(c)

    c = tf.keras.layers.Conv2D(128, (3, 3), activation='relu', kernel_initializer='he_normal', padding='same')(p)
    c = tf.keras.layers.Dropout(0.2)(c)
    c = tf.keras.layers.Conv2D(128, (3, 3), activation='relu', kernel_initializer='he_normal', padding='same')(c)
    p = tf.keras.layers.MaxPooling2D(pool_size=(2, 2))(c)
    return tf.keras.Model(inputs=inputs, outputs=p)

# 初始化4个特征提取分支，需要权值共享可以复用同一个实例
ext_main = build_feat_extractor()
ext1 = build_feat_extractor()
ext2 = build_feat_extractor()
ext3 = build_feat_extractor()

# 支持直接输入5维堆叠数据
stacked_input = tf.keras.layers.Input((4, IMG_HEIGHT, IMG_WIDTH, IMG_CHANNELS))
# 自动拆分4个输入
in_main = tf.keras.layers.Lambda(lambda x: x[:,0,...])(stacked_input)
in1 = tf.keras.layers.Lambda(lambda x: x[:,1,...])(stacked_input)
in2 = tf.keras.layers.Lambda(lambda x: x[:,2,...])(stacked_input)
in3 = tf.keras.layers.Lambda(lambda x: x[:,3,...])(stacked_input)

p_main = ext_main(in_main)
p1 = ext1(in1)
p2 = ext2(in2)
p3 = ext3(in3)

P = tf.keras.layers.concatenate([p1, p2, p3])
cp = tf.keras.layers.Conv2D(256, (3, 3), activation='relu', kernel_initializer='he_normal', padding='same')(P)
cp = tf.keras.layers.Dropout(0.3)(cp)
cp = tf.keras.layers.Conv2D(256, (3, 3), activation='relu', kernel_initializer='he_normal', padding='same')(cp)

one =  tf.keras.layers.concatenate([cp,p_main])
# 后续压缩、上采样层逻辑和你原有代码保持一致即可

额外注意 #

你现有代码中上采样部分有个笔误：tf.keras.layers.Conv2DTranspose(332, (2, 2), strides=(2, 2), padding='same')(d)，这里的通道数332大概率是写错了，应该为32，否则会导致后续卷积层通道匹配异常。

内容的提问来源于stack exchange，提问作者divine creator