ResNet50迁移学习加载权重时形状不匹配问题求助

问题说明 #

尝试用ResNet50构建自定义人脸识别模型，运行代码时出现权重形状不匹配错误，相关代码、报错信息如下：

原代码 #

def resnet50tl(input_shape, outclass, sigma='sigmoid'):
    base_model = None
    base_model = keras.applications.resnet50.ResNet50(weights='imagenet', include_top=False, input_shape=input_shape)
    base_model.load_weights(resnet50weight)
    for layer in base_model.layers:
        layer.trainable = False
    top_model = Sequential()
    top_model.add(Flatten(input_shape=base_model.output_shape[1:]))
    for i in range(2):
        top_model.add(Dense(4096, activation='relu'))
        top_model.add(Dropout(0.5))
    top_model.add(Dense(outclass, activation=sigma))
    model = Model(inputs=base_model.input, outputs=top_model(base_model.output))
    if resnet50weight is not None:
        model.load_weights(resnet50weight, by_name=True, skip_mismatch=True, reshape=True)
    return model
input_shape = (224, 224, 3)
numclasses = 6 
model = resnet50tl(input_shape, numclasses, 'softmax')
lr = 1e-5
decay = 1e-7
optimizer = RMSprop(lr=lr, decay=decay)
model.compile(loss='categorical_crossentropy', optimizer=optimizer, metrics=['accuracy'])

报错信息 #

File "c:\Users\mahmo\Downloads\test\VGG model.py", line 109, in <module>
    model = resnet50tl(input_shape, numclasses, 'softmax')
  File "c:\Users\mahmo\Downloads\test\VGG model.py", line 87, in resnet50tl
    base_model.load_weights(resnet50weight)
  File "C:\Users\mahmo\AppData\Local\Programs\Python\Python310\lib\site-packages\keras\src\utils\traceback_utils.py", line 70, in error_handler
    raise e.with_traceback(filtered_tb) from None
  File "C:\Users\mahmo\AppData\Local\Programs\Python\Python310\lib\site-packages\keras\src\backend.py", line 4361, in _assign_value_to_variable
    variable.assign(value)
ValueError: Cannot assign value to variable ' conv3_block1_0_conv/kernel:0': Shape mismatch.The variable shape (1, 1, 256, 512), and the assigned value shape (512, 128, 1, 1) are incompatible

错误原因及解决方案 #

核心问题 #

1. 重复加载权重：初始化base_model时已指定weights='imagenet'加载官方权重，紧接着又调用base_model.load_weights(resnet50weight)覆盖，两次加载的权重格式/结构不匹配。
2. 权重文件不兼容：resnet50weight对应的权重文件可能与Keras官方ResNet50的层形状、通道顺序（如通道在前/在后）不匹配，或是其他框架导出的权重。

修正代码 #

def resnet50tl(input_shape, outclass, sigma='sigmoid'):
    # 直接加载指定权重，无自定义权重时用imagenet
    base_model = keras.applications.resnet50.ResNet50(
        weights=resnet50weight if resnet50weight else 'imagenet',
        include_top=False,
        input_shape=input_shape
    )
    # 冻结基模型所有层
    for layer in base_model.layers:
        layer.trainable = False
    # 构建顶层分类器
    top_model = Sequential()
    top_model.add(Flatten(input_shape=base_model.output_shape[1:]))
    for i in range(2):
        top_model.add(Dense(4096, activation='relu'))
        top_model.add(Dropout(0.5))
    top_model.add(Dense(outclass, activation=sigma))
    # 拼接完整模型
    model = Model(inputs=base_model.input, outputs=top_model(base_model.output))
    return model

input_shape = (224, 224, 3)
numclasses = 6 
model = resnet50tl(input_shape, numclasses, 'softmax')
lr = 1e-5
decay = 1e-7
optimizer = RMSprop(lr=lr, decay=decay)
model.compile(loss='categorical_crossentropy', optimizer=optimizer, metrics=['accuracy'])

额外说明 #

• 若resnet50weight是自定义训练的权重，确保为Keras兼容的.h5格式，且与当前ResNet50结构完全匹配（输入形状、层名称一致）。
• 若为其他框架（如PyTorch）转换的权重，需调整通道顺序（Keras默认通道最后，PyTorch通道在前），或重新转换权重格式。
• 若必须分两次加载权重，需给base_model.load_weights()添加by_name=True, skip_mismatch=True参数跳过不匹配层，但不推荐此方式，易引发后续问题。

内容的提问来源于stack exchange，提问作者nariaa 99