keras - 加载最佳试用版时，Keras 调谐器中是否存在错误？

Question

我正在使用 Keras 来调整模型的超参数。搜索过程很顺利，但是，当我使用“tuner.results_summary()”或“tuner.oracle.get_best_trials(num_trials=1)[0]”查看得分最高的轨迹时，我发现了一些奇怪的东西。如果我运行“tuner.results_summary()”几次，搜索结束时，前 10 个试验是不同的。这是为什么？

project_name='InputNormN2_1'
Batch_size=128
h_tuner = Hyperband(
    build_model,
    max_epochs=81,
    factor=3,
    objective=kt.Objective("val_root_mean_squared_error", direction="min"),
    executions_per_trial=1,
    directory='Paper_Results',
    project_name=project_name)
stop_early = tf.keras.callbacks.EarlyStopping(monitor='val_root_mean_squared_error', patience=50)

h_tuner.search(x=x_train,y=y_train,validation_data=(x_val, y_val), 
            batch_size=Batch_size,
            shuffle=True,
            callbacks=[stop_early])

如果我再次加载调谐器并查看结果摘要：

h_tuner = Hyperband(
    build_model,
    max_epochs=81,
    factor=3,
    objective=kt.Objective("val_root_mean_squared_error", direction="min"),
    executions_per_trial=1,
    directory='Paper_Results',
    project_name=project_name)

print(h_tuner.results_summary())

Trial_n=h_tuner.oracle.get_best_trials(num_trials=1)[0].hyperparameters.values

print(Trial_n)