Keras 1.2转2.9：UNet代码Concatenate层形状匹配错误求助

问题根源 #

Keras 1.2的merge函数会自动做尺寸对齐（裁剪/填充），但Keras 2.x的Concatenate层严格要求除拼接轴外的所有维度完全匹配。你遇到的[(None, 512, 2, 2), (None, 256, 1, 1)]输入形状（推测为channels-first格式），核心问题是特征图的高/宽维度（2,2 vs 1,1）不匹配，无法直接拼接。

修复方案 #

1. 统一数据格式（推荐） #

Keras 2.x默认使用channels-last格式（(None, H, W, C)），如果代码仍用channels-first，建议切换：

from keras import backend as K
K.set_image_data_format('channels_last')

后续拼接轴需对应：channels-last用axis=-1，channels-first用axis=1。

2. 用转置卷积保证尺寸匹配（首选） #

原版UNet用转置卷积（Conv2DTranspose）替代简单上采样，可直接控制输出尺寸，避免手动调整。以下是适配Keras 2.9的完整UNet代码：

from keras.models import Model
from keras.layers import Input, Conv2D, MaxPooling2D, Conv2DTranspose, Concatenate, Activation

def unet(input_size=(512, 512, 1)):  # channels-last格式
    inputs = Input(input_size)
    
    # 编码器（下采样）
    c1 = Conv2D(64, (3, 3), padding='same')(inputs)
    c1 = Activation('relu')(c1)
    c1 = Conv2D(64, (3, 3), padding='same')(c1)
    c1 = Activation('relu')(c1)
    p1 = MaxPooling2D((2, 2))(c1)
    
    c2 = Conv2D(128, (3, 3), padding='same')(p1)
    c2 = Activation('relu')(c2)
    c2 = Conv2D(128, (3, 3), padding='same')(c2)
    c2 = Activation('relu')(c2)
    p2 = MaxPooling2D((2, 2))(c2)
    
    c3 = Conv2D(256, (3, 3), padding='same')(p2)
    c3 = Activation('relu')(c3)
    c3 = Conv2D(256, (3, 3), padding='same')(c3)
    c3 = Activation('relu')(c3)
    p3 = MaxPooling2D((2, 2))(c3)
    
    c4 = Conv2D(512, (3, 3), padding='same')(p3)
    c4 = Activation('relu')(c4)
    c4 = Conv2D(512, (3, 3), padding='same')(c4)
    c4 = Activation('relu')(c4)
    p4 = MaxPooling2D((2, 2))(c4)
    
    # 瓶颈层
    c5 = Conv2D(1024, (3, 3), padding='same')(p4)
    c5 = Activation('relu')(c5)
    c5 = Conv2D(1024, (3, 3), padding='same')(c5)
    c5 = Activation('relu')(c5)
    
    # 解码器（上采样）
    u6 = Conv2DTranspose(512, (2, 2), strides=(2, 2), padding='same')(c5)
    concat6 = Concatenate(axis=-1)([c4, u6])
    c6 = Conv2D(512, (3, 3), padding='same')(concat6)
    c6 = Activation('relu')(c6)
    c6 = Conv2D(512, (3, 3), padding='same')(c6)
    c6 = Activation('relu')(c6)
    
    u7 = Conv2DTranspose(256, (2, 2), strides=(2, 2), padding='same')(c6)
    concat7 = Concatenate(axis=-1)([c3, u7])
    c7 = Conv2D(256, (3, 3), padding='same')(concat7)
    c7 = Activation('relu')(c7)
    c7 = Conv2D(256, (3, 3), padding='same')(c7)
    c7 = Activation('relu')(c7)
    
    u8 = Conv2DTranspose(128, (2, 2), strides=(2, 2), padding='same')(c7)
    concat8 = Concatenate(axis=-1)([c2, u8])
    c8 = Conv2D(128, (3, 3), padding='same')(concat8)
    c8 = Activation('relu')(c8)
    c8 = Conv2D(128, (3, 3), padding='same')(c8)
    c8 = Activation('relu')(c8)
    
    u9 = Conv2DTranspose(64, (2, 2), strides=(2, 2), padding='same')(c8)
    concat9 = Concatenate(axis=-1)([c1, u9])
    c9 = Conv2D(64, (3, 3), padding='same')(concat9)
    c9 = Activation('relu')(c9)
    c9 = Conv2D(64, (3, 3), padding='same')(c9)
    c9 = Activation('relu')(c9)
    
    outputs = Conv2D(1, (1, 1), activation='sigmoid')(c9)
    
    model = Model(inputs=[inputs], outputs=[outputs])
    return model

3. 手动裁剪/填充（保留UpSampling2D时用） #

如果坚持使用UpSampling2D，可手动调整特征图尺寸：

• 裁剪大尺寸特征图：

from keras.layers import Cropping2D
# 假设up6是(None,512,2,2)，conv4是(None,256,1,1)（channels-first）
cropped_up6 = Cropping2D(cropping=((0,1),(0,1)))(up6)  # 裁剪高/宽各1个像素
concat6 = Concatenate(axis=1)([conv4, cropped_up6])

• 填充小尺寸特征图：

from keras.layers import ZeroPadding2D
padded_conv4 = ZeroPadding2D(padding=((0,1),(0,1)))(conv4)  # 填充高/宽各1个像素
concat6 = Concatenate(axis=1)([padded_conv4, up6])

关键注意事项 #

• 编码器卷积层务必用padding='same'，避免特征图尺寸意外缩小；若用padding='valid'，需提前计算每一层的尺寸变化。
• 拼接轴参数必须和数据格式对应，channels-last用axis=-1，channels-first用axis=1。

内容的提问来源于stack exchange，提问作者Mufeed Patel