TensorFlow/Keras循环构建多模型时保持相同层名称的实现方法

I get it—when you're looping to build multiple copies of the same model, it's frustrating that Keras automatically increments layer and model names (like conv2d, conv2d_2, etc.) instead of keeping them identical across each model instance. Let's fix this with a couple of straightforward approaches.

Your Current Setup #

First, let's recap your code for context:

Model Building Function

def build_model():
    # 创建模型
    model = Sequential([
        InputLayer(input_shape = (28, 28, 1)),
        Conv2D(32, (3, 3)),
        Activation('relu'),
        MaxPooling2D((2, 2)),
        Conv2D(64, (3, 3)),
        Activation('relu'),
        MaxPooling2D((2, 2)),
        Flatten(),
        Dense(num_classes),
        Activation('softmax')
    ])
    # 编译模型
    model.compile(
        loss = 'categorical_crossentropy',
        optimizer = 'adam',
        metrics = [ 'accuracy' ]
    )
    # 返回模型
    return model

Loop to Build Models

for i in range(2):
    model = build_model()
    model.summary()

As you noticed, each new model gets incremented names like sequential_1, conv2d_2, etc. Here's how to stop that:

Solution 1: Manually Specify Layer Names (Most Reliable) #

The simplest and most maintainable way is to explicitly set the name parameter for each layer in your model. This overrides Keras's automatic name incrementing, ensuring every model instance uses the exact same layer names.

Modify your build_model function like this:

def build_model():
    # 创建模型，手动指定每个层的名称
    model = Sequential([
        InputLayer(input_shape=(28, 28, 1), name='input_layer'),
        Conv2D(32, (3, 3), name='conv2d'),
        Activation('relu', name='relu_1'),
        MaxPooling2D((2, 2), name='max_pool_1'),
        Conv2D(64, (3, 3), name='conv2d_1'),
        Activation('relu', name='relu_2'),
        MaxPooling2D((2, 2), name='max_pool_2'),
        Flatten(name='flatten'),
        Dense(num_classes, name='dense_output'),
        Activation('softmax', name='softmax_output')
    ])
    # 编译模型
    model.compile(
        loss='categorical_crossentropy',
        optimizer='adam',
        metrics=['accuracy']
    )
    # 返回模型
    return model

Now when you run your loop, every model's summary will show identical layer names (e.g., conv2d, conv2d_1) across all iterations.

Solution 2: Reset Keras's Name Counters (Less Recommended) #

If you don't want to manually name every layer, you can reset the internal counters Keras uses to generate automatic names. Note that this is more fragile (it depends on Keras's internal implementation) and can cause conflicts if you're running other model-building code alongside your loop.

For TensorFlow Keras, you can reset counters for specific layer types like this:

from tensorflow.keras.layers import Sequential, Conv2D, MaxPooling2D, Dense

def build_model():
    # 重置层的命名计数器
    Sequential._name_counter = 0
    Conv2D._name_counter = 0
    MaxPooling2D._name_counter = 0
    Dense._name_counter = 0
    # 剩下的模型构建代码和之前一样
    model = Sequential([
        InputLayer(input_shape = (28, 28, 1)),
        Conv2D(32, (3, 3)),
        Activation('relu'),
        MaxPooling2D((2, 2)),
        Conv2D(64, (3, 3)),
        Activation('relu'),
        MaxPooling2D((2, 2)),
        Flatten(),
        Dense(num_classes),
        Activation('softmax')
    ])
    model.compile(
        loss = 'categorical_crossentropy',
        optimizer = 'adam',
        metrics = [ 'accuracy' ]
    )
    return model

This will reset the counters each time you call build_model(), so Keras starts naming layers from the original base name again. However, if Keras changes how it tracks these counters in future versions, this method might break.

Why Does Keras Increment Names? #

Keras automatically increments layer names to ensure every layer in the global scope has a unique identifier. This prevents naming conflicts when multiple models are loaded or used in the same session. By manually setting names, you take control of that uniqueness (since each model is a separate instance, the same layer names won't conflict between them).

内容的提问来源于stack exchange，提问作者user14977424