Xgboost Amazon Sagemaker网格搜索替代方案

Question

我正在使用Amazon Sagemaker运行一个xgboost模型，以便押注最好的超参数组合。我必须使用sagemaker实现，而不是notebook替代方案来测试它是否比gridsearch运行得更快。我的问题是如何让它在循环中工作。有什么想法吗？我的理解是，我必须用不同的组合来编写大量的工作。我试着这样做作为测试：

for i in range (1,3):
    for j in range (13,15):
        job_name = 'regression' + '-'+str(i) +"-"+str(j)+"-" +strftime("%Y-%m-%d-%H-%M-%S", gmtime())

        job_name_params = copy.deepcopy(parameters_xgboost)
        job_name_params['TrainingJobName'] = job_name
        job_name_params['OutputDataConfig']['S3OutputPath']= "....."
        job_name_params['HyperParameters']['objective'] = "reg:linear"
        job_name_params['HyperParameters']['silent'] = "0"
        job_name_params['HyperParameters']['max_depth'] = str(i)
        job_name_params['HyperParameters']['min_child_weight'] = str(j)
        job_name_params['HyperParameters']['eta'] = "0.01"
        job_name_params['HyperParameters']['num_round'] = "1000"
        job_name_params['HyperParameters']['subsample'] = "0.5"
        job_name_params['HyperParameters']['colsample_bytree'] = "0.5"

        sm = boto3.Session().client('.....')


        sm.create_training_job(**job_name_params)
        sm.get_waiter('training_job_completed_or_stopped').wait(TrainingJobName=job_name)
        status = sm.describe_training_job(TrainingJobName=job_name)['TrainingJobStatus']
        print("Training job ended with status: " + status)

parameters_xgboost是Sagemaker读取基本信息和超参数列表的方式。

好消息是它是有效的。糟糕的是，这样一次只能训练一个模型。我希望所有这些组合都能在同一时间运行。我该怎么做呢？

Answer 1

亚马逊SageMaker正在为您提供自动运行超参数优化(HPO)的调整服务。该服务现在已普遍可用。有关详细信息，请参阅the documentation。

​