将数据移入 ML 管道和在 ML 管道之间移动数据的步骤 (Python)Moving data into and between ML pipeline steps (Python)

• 08/20/2020
• • l

本文提供了用于在 Azure 机器学习管道中的步骤之间导入、转换和移动数据的代码。This article provides code for importing, transforming, and moving data between steps in an Azure Machine Learning pipeline. 有关数据在 Azure 机器学习中的工作原理的概述，请参阅访问 Azure 存储服务中的数据。For an overview of how data works in Azure Machine Learning, see Access data in Azure storage services. 有关 Azure 机器学习管道的优点和结构，请参阅什么是 Azure 机器学习管道？。For the benefits and structure of Azure Machine Learning pipelines, see What are Azure Machine Learning pipelines?.

本文将向你介绍如何：This article will show you how to:

• 将 Dataset 对象用于预先存在的数据Use Dataset objects for pre-existing data
• 在步骤中访问数据Access data within your steps
• 将 Dataset 数据拆分为子集，例如训练子集和验证子集Split Dataset data into subsets, such as training and validation subsets
• 创建 PipelineData 对象来将数据传输到下一管道步骤Create PipelineData objects to transfer data to the next pipeline step
• 使用 PipelineData 对象作为管道步骤的输入Use PipelineData objects as input to pipeline steps
• 基于 PipelineData 创建你要持久保存的新 Dataset 对象Create new Dataset objects from PipelineData you wish to persist

先决条件Prerequisites

需要：You'll need:

• Azure 订阅。An Azure subscription. 如果没有 Azure 订阅，请在开始前创建一个试用帐户。If you don't have an Azure subscription, create a trial account before you begin. 试用试用版 Azure 机器学习。Try the trial version of Azure Machine Learning.

• 适用于 Python 的 Azure 机器学习 SDK，或 Azure 机器学习工作室的访问权限。The Azure Machine Learning SDK for Python, or access to Azure Machine Learning studio.

• Azure 机器学习工作区。An Azure Machine Learning workspace.

  创建 Azure 机器学习工作区或通过 Python SDK 使用现有工作区。Either create an Azure Machine Learning workspace or use an existing one via the Python SDK. 导入 Workspace 和 Datastore 类，并使用函数 from_config() 从文件 config.json 中加载订阅信息。Import the Workspace and Datastore class, and load your subscription information from the file config.json using the function from_config(). 默认情况下，此函数会查找当前目录中的 JSON 文件，但你也可以使用 from_config(path="your/file/path") 指定一个路径参数，使之指向该文件。This function looks for the JSON file in the current directory by default, but you can also specify a path parameter to point to the file using from_config(path="your/file/path").

  import azureml.core
  from azureml.core import Workspace, Datastore
  
  ws = Workspace.from_config()
  

• 某些预先存在的数据。Some pre-existing data. 本文简要介绍了如何使用 Azure blob 容器。This article briefly shows the use of an Azure blob container.

• 可选：现有机器学习管道，例如使用 Azure 机器学习 SDK 创建和运行机器学习管道中所述的管道。Optional: An existing machine learning pipeline, such as the one described in Create and run machine learning pipelines with Azure Machine Learning SDK.

将 Dataset 对象用于预先存在的数据Use Dataset objects for pre-existing data

将数据引入到管道的首选方法是使用 Dataset 对象。The preferred way to ingest data into a pipeline is to use a Dataset object. Dataset 对象表示在整个工作区中可用的持久性数据。Dataset objects represent persistent data available throughout a workspace.

可以通过许多方法来创建和注册 Dataset 对象。There are many ways to create and register Dataset objects. 表格数据集用于一个或多个文件中可用的分隔数据。Tabular datasets are for delimited data available in one or more files. 文件数据集用于二进制数据（例如图像）或你要分析的数据。File datasets are for binary data (such as images) or for data that you'll parse. 创建 Dataset 对象的最简单编程方式是使用工作区存储或公共 URL 中的现有 blob：The simplest programmatic ways to create Dataset objects are to use existing blobs in workspace storage or public URLs:

datastore = Datastore.get(workspace, 'training_data')
iris_dataset = Dataset.Tabular.from_delimited_files(DataPath(datastore, 'iris.csv'))

cats_dogs_dataset = Dataset.File.from_files(
    paths='https://download.microsoft.com/download/3/E/1/3E1C3F21-ECDB-4869-8368-6DEBA77B919F/kagglecatsanddogs_3367a.zip',
    archive_options=ArchiveOptions(archive_type=ArchiveType.ZIP, entry_glob='**/*.jpg')
)

若要详细了解如何使用各种选项以及基于各种源创建数据集、在 Azure 机器学习 UI 中注册数据集并查看它们、了解数据大小如何与计算容量交互，以及如何对数据集进行版本控制，请参阅创建 Azure 机器学习数据集。For more options on creating datasets with different options and from different sources, registering them and reviewing them in the Azure Machine Learning UI, understanding how data size interacts with compute capacity, and versioning them, see Create Azure Machine Learning datasets.

将数据集传递给脚本Pass datasets to your script

若要将数据集的路径传递给你的脚本，请使用 Dataset 对象的 as_named_input() 方法。To pass the dataset's path to your script, use the Dataset object's as_named_input() method. 你可以将生成的 DatasetConsumptionConfig 对象作为参数传递给你的脚本，也可以通过使用管道脚本的 inputs 参数来使用 Run.get_context().input_datasets[] 检索数据集。You can either pass the resulting DatasetConsumptionConfig object to your script as an argument or, by using the inputs argument to your pipeline script, you can retrieve the dataset using Run.get_context().input_datasets[].

创建命名输入后，可以选择其访问模式：as_mount() 或 as_download()。Once you've created a named input, you can choose its access mode: as_mount() or as_download(). 如果你的脚本处理数据集中的所有文件，而计算资源上的磁盘足以容纳数据集，则下载访问模式是更好的选择。If your script processes all the files in your dataset and the disk on your compute resource is large enough for the dataset, the download access mode is the better choice. 下载访问模式可避免在运行时流式传输数据所产生的开销。The download access mode will avoid the overhead of streaming the data at runtime. 如果你的脚本访问数据集的子集或它对于你的计算而言太大，请使用装载访问模式。If your script accesses a subset of the dataset or it's too large for your compute, use the mount access mode. 有关详细信息，请阅读装载与下载For more information, read Mount vs. Download

若要将数据集传递给管道步骤，请执行以下操作：To pass a dataset to your pipeline step:

1. 使用 TabularDataset.as_named_input() 或 FileDataset.as_named_input()（末尾没有“s”）创建一个 DatasetConsumptionConfig 对象Use TabularDataset.as_named_input() or FileDataset.as_named_input() (no 's' at end) to create a DatasetConsumptionConfig object
2. 使用 as_mount() 或 as_download() 设置访问模式Use as_mount() or as_download() to set the access mode
3. 使用 arguments 或 inputs 参数将数据集传递给管道步骤Pass the datasets to your pipeline steps using either the arguments or the inputs argument

以下代码片段显示了在 PythonScriptStep 构造函数中组合这些步骤的常见模式：The following snippet shows the common pattern of combining these steps within the PythonScriptStep constructor:

train_step = PythonScriptStep(
    name="train_data",
    script_name="train.py",
    compute_target=cluster,
    inputs=[iris_dataset.as_named_input('iris').as_mount()]
)

你还可以使用 random_split() 和 take_sample() 等方法来创建多个输入或减少传递给管道步骤的数据量：You can also use methods such as random_split() and take_sample() to create multiple inputs or reduce the amount of data passed to your pipeline step:

seed = 42 # PRNG seed
smaller_dataset = iris_dataset.take_sample(0.1, seed=seed) # 10%
train, test = smaller_dataset.random_split(percentage=0.8, seed=seed)

train_step = PythonScriptStep(
    name="train_data",
    script_name="train.py",
    compute_target=cluster,
    inputs=[train.as_named_input('train').as_download(), test.as_named_input('test').as_download()]
)

在脚本中访问数据集Access datasets within your script

管道步骤脚本的命名输入可以在 Run 对象中用作字典。Named inputs to your pipeline step script are available as a dictionary within the Run object. 使用 Run.get_context() 检索活动 Run 对象，然后使用 input_datasets 检索命名输入的字典。Retrieve the active Run object using Run.get_context() and then retrieve the dictionary of named inputs using input_datasets. 如果你使用 arguments 参数而非 inputs 参数传递 DatasetConsumptionConfig 对象，请使用 ArgParser 代码访问数据。If you passed the DatasetConsumptionConfig object using the arguments argument rather than the inputs argument, access the data using ArgParser code. 以下代码片段演示了这两种技术。Both techniques are demonstrated in the following snippet.

# In pipeline definition script:
# Code for demonstration only: It would be very confusing to split datasets between `arguments` and `inputs`
train_step = PythonScriptStep(
    name="train_data",
    script_name="train.py",
    compute_target=cluster,
    arguments=['--training-folder', train.as_named_input('train').as_download()]
    inputs=[test.as_named_input('test').as_download()]
)

# In pipeline script
parser = argparse.ArgumentParser()
parser.add_argument('--training-folder', type=str, dest='train_folder', help='training data folder mounting point')
args = parser.parse_args()
training_data_folder = args.train_folder

testing_data_folder = Run.get_context().input_datasets['test']

传递的值将是数据集文件的路径。The passed value will be the path to the dataset file(s).

还可以直接访问已注册的 Dataset。It's also possible to access a registered Dataset directly. 由于已注册的数据集是持久性的，并且在整个工作区中共享，因此你可以直接检索它们：Since registered datasets are persistent and shared across a workspace, you can retrieve them directly:

run = Run.get_context()
ws = run.experiment.workspace
ds = Dataset.get_by_name(workspace=ws, name='mnist_opendataset')

将 PipelineData 用于中间数据Use PipelineData for intermediate data

尽管 Dataset 对象表示持久性数据，但 PipelineData 对象用于从管道步骤输出的临时数据。While Dataset objects represent persistent data, PipelineData objects are used for temporary data that is output from pipeline steps. 由于 PipelineData 对象的寿命比单个管道步骤长，因此你在管道定义脚本中定义它们。Because the lifespan of a PipelineData object is longer than a single pipeline step, you define them in the pipeline definition script. 创建 PipelineData 对象时，必须提供一个名称和用于存储数据的数据存储。When you create a PipelineData object, you must provide a name and a datastore at which the data will reside. 同时使用 arguments 和 outputs 参数将 PipelineData 对象传递给 PythonScriptStep：Pass your PipelineData object(s) to your PythonScriptStep using both the arguments and the outputs arguments:

default_datastore = workspace.get_default_datastore()
dataprep_output = PipelineData("clean_data", datastore=default_datastore)

dataprep_step = PythonScriptStep(
    name="prep_data",
    script_name="dataprep.py",
    compute_target=cluster,
    arguments=["--output-path", dataprep_output]
    inputs=[Dataset.get_by_name(workspace, 'raw_data')],
    outputs=[dataprep_output]
)

你可以选择使用一个提供即时上传功能的访问模式来创建 PipelineData 对象。You may choose to create your PipelineData object using an access mode that provides an immediate upload. 在这种情况下，当你创建 PipelineData 时，请将 upload_mode 设置为 "upload"，并使用 output_path_on_compute 参数指定要将数据写入到其中的路径：In that case, when you create your PipelineData, set the upload_mode to "upload" and use the output_path_on_compute argument to specify the path to which you'll be writing the data:

PipelineData("clean_data", datastore=def_blob_store, output_mode="upload", output_path_on_compute="clean_data_output/")

将 PipelineData 用作训练步骤的输出Use PipelineData as outputs of a training step

在管道的 PythonScriptStep 中，可以使用程序的参数检索可用输出路径。Within your pipeline's PythonScriptStep, you can retrieve the available output paths using the program's arguments. 如果此步骤是第一个步骤并将初始化输出数据，则必须在指定的路径创建目录。If this step is the first and will initialize the output data, you must create the directory at the specified path. 然后，你可以写入要包含在 PipelineData 中的任何文件。You can then write whatever files you wish to be contained in the PipelineData.

parser = argparse.ArgumentParser()
parser.add_argument('--output_path', dest='output_path', required=True)
args = parser.parse_args()

# Make directory for file
os.makedirs(os.path.dirname(args.output_path), exist_ok=True)
with open(args.output_path, 'w') as f:
    f.write("Step 1's output")

如果你已在创建 PipelineData 时将 is_directory 参数设置为 True，则只需执行 os.makedirs() 调用即可，然后就可以随意将所需的任何文件写入该路径。If you created your PipelineData with the is_directory argument set to True, it would be enough to just perform the os.makedirs() call and then you would be free to write whatever files you wished to the path. 有关更多详细信息，请查看 PipelineData 参考文档。For more details, see the PipelineData reference documentation.

读取 PipelineData 作为非初始步骤的输入Read PipelineData as inputs to non-initial steps

当初始管道步骤将一些数据写入到 PipelineData 路径并且该数据成为初始步骤的输出后，可将其用作后面步骤的输入：After the initial pipeline step writes some data to the PipelineData path and it becomes an output of that initial step, it can be used as an input to a later step:

step1_output_data = PipelineData("processed_data", datastore=def_blob_store, output_mode="upload")
# get adls gen 2 datastore already registered with the workspace
datastore = workspace.datastores['my_adlsgen2']

step1 = PythonScriptStep(
    name="generate_data",
    script_name="step1.py",
    runconfig = aml_run_config,
    arguments = ["--output_path", step1_output_data],
    inputs=[],
    outputs=[step1_output_data]
)

step2 = PythonScriptStep(
    name="read_pipeline_data",
    script_name="step2.py",
    compute_target=compute,
    runconfig = aml_run_config,
    arguments = ["--pd", step1_output_data],
    inputs=[step1_output_data]
)

pipeline = Pipeline(workspace=ws, steps=[step1, step2])

PipelineData 输入的值是上一输出的路径。The value of a PipelineData input is the path to the previous output.

如前所示，如果第一个步骤写入了单个文件，则可以如下所示使用它：If, as shown previously, the first step wrote a single file, consuming it might look like:

parser = argparse.ArgumentParser()
parser.add_argument('--pd', dest='pd', required=True)
args = parser.parse_args()

with open(args.pd) as f:
    print(f.read())

将 PipelineData 对象转换为 DatasetConvert PipelineData objects to Datasets

如果要使 PipelineData 的可用时间超过某个运行的持续时间，请使用其 as_dataset() 函数将其转换为 Dataset。If you'd like to make your PipelineData available for longer than the duration of a run, use its as_dataset() function to convert it to a Dataset. 然后，你可以注册 Dataset，使其在你的工作区中处于第一等级。You may then register the Dataset, making it a first-class citizen in your workspace. 由于每次运行管道时，PipelineData 对象的路径都不同，因此强烈建议你在注册基于 PipelineData 对象创建的 Dataset 时将 create_new_version 设置为 True。Since your PipelineData object will have a different path every time the pipeline runs, it's highly recommended that you set create_new_version to True when registering a Dataset created from a PipelineData object.

step1_output_ds = step1_output_data.as_dataset()
step1_output_ds.register(name="processed_data", create_new_version=True)

后续步骤Next steps

• 创建 Azure 机器学习数据集Create an Azure machine learning dataset
• 使用 Azure 机器学习 SDK 创建和运行机器学习管道Create and run machine learning pipelines with Azure Machine Learning SDK