Keras Tuner与Functional API调参时层名称重复错误排查

问题场景 #

尝试使用Keras Tuner优化多输入模型的超参数，代码如下：

def criar_modelo(hp):
    lstm_input = Input(shape=(x_train_lstm.shape[1], 1), name='LSTM_Input_Layer')
    static_input = Input(shape=(x_train_static.shape[1], ), name='Static_Input_Layer')
    
    # LSTM 1
    lstm_layer_1 = LSTM(units=hp.Int('units_lstm_layer_1', min_value=128, max_value=256, step=64), activation='tanh', return_sequences=False, name='1_LSTM_Layer')(lstm_input)
    
    # Static 1 
    static_layer_1 = Dense(units=hp.Int('units_static_layer_1', min_value=64, max_value=192, step=64), activation=hp.Choice('activation', ['relu', 'tanh']), name='1_Static_Layer')(static_input)

    # Static 2 e/ou 3
    for i in range(hp.Int('num_static_layers', 1, 3)):
        static_layer = Dense(units=hp.Int(f'static_units_{i}', 128, 192, step=32), activation=hp.Choice('activation', ['relu', 'tanh']), name=f'{i+1}_Static_Layer')(static_layer_1)
        static_layer_1 = static_layer
    
    concatenar = Concatenate(axis=1, name='Concatenate')([lstm_layer_1, static_layer_1])
    dense_1 = Dense(units=4*len(np.unique(y_train)), activation='relu', name='1_Dense_Layer')(concatenar)
    dense_2 = Dense(units=2*len(np.unique(y_train)), activation='relu', name='2_Dense_Layer')(dense_1)
    saida = Dense(units=len(np.unique(y_train)), activation='softmax', name='Output_Layer')(dense_2)
    model = Model(inputs=[lstm_input, static_input], outputs=[saida])
    model.compile(optimizer='adam', loss='sparse_categorical_crossentropy', metrics=['accuracy'])
    return model

运行时触发以下错误：

Traceback (most recent call last):
  Cell In[11], line 27
    tuner = keras_tuner.GridSearch(
  File ~\miniconda3\envs\tf-gpu\lib\site-packages\keras_tuner\src\tuners\gridsearch.py:420 in __init__
    super().__init__(oracle, hypermodel, **kwargs)
  File ~\miniconda3\envs\tf-gpu\lib\site-packages\keras_tuner\src\engine\tuner.py:122 in __init__
    super().__init__(
  File ~\miniconda3\envs\tf-gpu\lib\site-packages\keras_tuner\src\engine\base_tuner.py:132 in __init__
    self._populate_initial_space()
  File ~\miniconda3\envs\tf-gpu\lib\site-packages\keras_tuner\src\engine\base_tuner.py:192 in _populate_initial_space
    self._activate_all_conditions()
  File ~\miniconda3\envs\tf-gpu\lib\site-packages\keras_tuner\src\engine\base_tuner.py:149 in _activate_all_conditions
    self.hypermodel.build(hp)
  Cell In[11], line 23 in criar_modelo
    model = Model(inputs=[lstm_input, static_input], outputs=[saida])
  File ~\miniconda3\envs\tf-gpu\lib\site-packages\tensorflow\python\training\tracking\base.py:629 in _method_wrapper
    result = method(self, *args, **kwargs)
  File ~\miniconda3\envs\tf-gpu\lib\site-packages\keras\engine\functional.py:146 in __init__
    self._init_graph_network(inputs, outputs)
  File ~\miniconda3\envs\tf-gpu\lib\site-packages\tensorflow\python\training\tracking\base.py:629 in _method_wrapper
    result = method(self, *args, **kwargs)
  File ~\miniconda3\envs\tf-gpu\lib\site-packages\keras\engine\functional.py:229 in _init_graph_network
    nodes, nodes_by_depth, layers, _ = _map_graph_network(
  File ~\miniconda3\envs\tf-gpu\lib\site-packages\keras\engine\functional.py:1049 in _map_graph_network
    raise ValueError(
ValueError: The name "1_Static_Layer" is used 2 times in the model. All layer names should be unique.

用户已确认循环变量i值为0，可通过修改层命名规则绕过错误，但希望了解根本原因。

根本原因 #

1. 命名逻辑直接冲突：
   • 手动创建的第一个静态层已命名为1_Static_Layer
   • 循环中超参数num_static_layers的最小值为1，对应range(1)生成i=0，此时循环内的层名计算为f'{i+1}_Static_Layer'，即1_Static_Layer，与手动层完全重名
2. Keras核心约束：
   Keras要求模型内所有层的名称必须唯一，检测到重复名称会直接抛出ValueError
3. Tuner的预检查机制：
   Keras Tuner初始化时会调用_activate_all_conditions()，遍历所有可能的超参数组合构建模型以验证空间有效性，因此即使当前运行单一组参数，Tuner也会检测所有参数组合的合法性，触发了命名冲突的检测。

解决方案 #

方案1：修改手动层命名 #

将手动创建的第一个静态层改为不冲突的名称，比如：

static_layer_1 = Dense(..., name='0_Static_Layer')(static_input)

方案2：调整循环层命名规则 #

保持手动层命名不变，让循环层起始编号从2开始：

name=f'{i+2}_Static_Layer'

方案3：调整超参数取值范围 #

将num_static_layers最小值改为0，配合调整循环层编号逻辑，避免重复：

for i in range(hp.Int('num_static_layers', 0, 2)):
    static_layer = Dense(..., name=f'{i+2}_Static_Layer')(static_layer_1)

内容的提问来源于stack exchange，提问作者Murilo