使用Keras Tuner调参回归ANN时出现KeyError: 'val_mean_absolute_error'

问题重现 #

尝试用Keras Tuner对回归人工神经网络（ANN）做超参数调优时，执行超参数搜索收到错误：KeyError: 'val_mean_absolute_error'

代码 #

def build_model(hp):
    model = keras.Sequential()
    for i in range(hp.Int('num_layers', 2, 20)):
        model.add(layers.Dense(
            units=hp.Int('units_' + str(i), min_value=32, max_value=512, step=32),
            activation='relu'
        ))
        if hp.Boolean("dropout"):
            model.add(layers.Dropout(rate=0.5))
    model.add(layers.Dense(1, activation='linear'))
    model.compile(
        optimizer=keras.optimizers.Adam(hp.Choice('learning_rate', [1e-4, 1e-3, 1e-5])),
        loss='mean_absolute_error',
        metrics=['mean_absolute_error']
    )
    return model

tuner = kt.RandomSearch(
    build_model,
    objective='val_mean_absolute_error',
    max_trials=5,
    executions_per_trial=2,
    overwrite=True,
    directory="Local Files",
    project_name="Keras")

tuner.search(x_train, y_train, epochs=1)

报错回溯 #

Traceback (most recent call last):
  File "c:\Users\{user}\.venv\lib\site-packages\keras_tuner\engine\base_tuner.py", line 266, in _try_run_and_update_trial
    self._run_and_update_trial(trial, *fit_args, **fit_kwargs)
  File "c:\Users\{user}\.venv\lib\site-packages\keras_tuner\engine\base_tuner.py", line 257, in _run_and_update_trial
    tuner_utils.convert_to_metrics_dict(
  File "c:\Users\{user}\.venv\lib\site-packages\keras_tuner\engine\tuner_utils.py", line 270, in convert_to_metrics_dict
    [convert_to_metrics_dict(elem, objective) for elem in results]
  File "c:\Users\{user}\.venv\lib\site-packages\keras_tuner\engine\tuner_utils.py", line 270, in <listcomp>
    [convert_to_metrics_dict(elem, objective) for elem in results]
  File "c:\Users\{user}\.venv\lib\site-packages\keras_tuner\engine\tuner_utils.py", line 283, in convert_to_metrics_dict
    best_value, _ = _get_best_value_and_best_epoch_from_history(
  File "c:\Users\{user}\.venv\lib\site-packages\keras_tuner\engine\tuner_utils.py", line 254, in _get_best_value_and_best_epoch_from_history
    objective_value = objective.get_value(metrics)
  File "c:\Users\{user}\.venv\lib\site-packages\keras_tuner\engine\objective.py", line 57, in get_value
    return logs[self.name]
KeyError: 'val_mean_absolute_error'

原因分析 #

你指定的优化目标是val_mean_absolute_error（验证集平均绝对误差），但调用tuner.search()时既没有传入单独的验证集数据，也没有设置validation_split参数划分训练集的一部分作为验证集。Keras Tuner无法计算验证集指标，因此找不到val_mean_absolute_error这个键，触发KeyError。

解决方法 #

方法1：提供验证集数据 #

在tuner.search()中添加验证集相关参数，两种方式任选其一：

• 用validation_split从训练集中划分比例作为验证集：

tuner.search(x_train, y_train, epochs=1, validation_split=0.2)  # 划分20%训练数据为验证集

• 传入提前准备好的验证集：

tuner.search(x_train, y_train, epochs=1, validation_data=(x_val, y_val))

方法2：修改优化目标为训练集指标 #

如果不需要验证集评估，直接把目标改为训练集的mean_absolute_error：

tuner = kt.RandomSearch(
    build_model,
    objective='mean_absolute_error',  # 替换为训练集指标
    max_trials=5,
    executions_per_trial=2,
    overwrite=True,
    directory="Local Files",
    project_name="Keras")

内容的提问来源于stack exchange，提问作者Tim