Moving data into and between ML pipeline steps (Python)

APPLIES TO: Python SDK azureml v1

This article provides code for importing, transforming, and moving data between steps in an Azure Machine Learning pipeline. For an overview of how data works in Azure Machine Learning, see Access data in Azure storage services. For the benefits and structure of Azure Machine Learning pipelines, see What are Azure Machine Learning pipelines?

This article shows you how to:

  • Use Dataset objects for pre-existing data
  • Access data within your steps
  • Split Dataset data into subsets, such as training and validation subsets
  • Create OutputFileDatasetConfig objects to transfer data to the next pipeline step
  • Use OutputFileDatasetConfig objects as input to pipeline steps
  • Create new Dataset objects from OutputFileDatasetConfig you wish to persist

Prerequisites

You need:

Use Dataset objects for pre-existing data

The preferred way to ingest data into a pipeline is to use a Dataset object. Dataset objects represent persistent data available throughout a workspace.

There are many ways to create and register Dataset objects. Tabular datasets are for delimited data available in one or more files. File datasets are for binary data (such as images) or for data that you parse. 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'))

datastore_path = [
    DataPath(datastore, 'animals/dog/1.jpg'),
    DataPath(datastore, 'animals/dog/2.jpg'),
    DataPath(datastore, 'animals/cat/*.jpg')
]
cats_dogs_dataset = Dataset.File.from_files(path=datastore_path)

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.

Pass datasets to your script

To pass the dataset's path to your script, use the Dataset object's as_named_input() method. 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[].

Once you've created a named input, you can choose its access mode(for FileDataset only): as_mount() or as_download(). 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. The download access mode avoids the overhead of streaming the data at runtime. If your script accesses a subset of the dataset or it's too large for your compute, use the mount access mode. For more information, read Mount vs. Download

To pass a dataset to your pipeline step:

  1. Use TabularDataset.as_named_input() or FileDataset.as_named_input() (no 's' at end) to create a DatasetConsumptionConfig object
  2. For FileDataset only:. Use as_mount() or as_download() to set the access mode. TabularDataset does not suppport set access mode.
  3. Pass the datasets to your pipeline steps using either the arguments or the inputs argument

The following snippet shows the common pattern of combining these steps within the PythonScriptStep constructor, using iris_dataset (TabularDataset):


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

Note

You would need to replace the values for all these arguments (that is, "train_data", "train.py", cluster, and iris_dataset) with your own data. The above snippet just shows the form of the call and is not part of a Microsoft sample.

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'), test.as_named_input('test')]
)

Access datasets within your script

Named inputs to your pipeline step script are available as a dictionary within the Run object. Retrieve the active Run object using Run.get_context() and then retrieve the dictionary of named inputs using input_datasets. If you passed the DatasetConsumptionConfig object using the arguments argument rather than the inputs argument, access the data using ArgParser code. Both techniques are demonstrated in the following snippets:

The 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,
    # datasets passed as arguments
    arguments=['--training-folder', train.as_named_input('train').as_download()],
    # datasets passed as inputs
    inputs=[test.as_named_input('test').as_download()]
)

The train.py script referenced from the PythonScriptStep

# In pipeline script
parser = argparse.ArgumentParser()
# Retreive the dataset passed as an argument
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
# Retrieve the dataset passed as an input
testing_data_folder = Run.get_context().input_datasets['test']

The passed value is the path to the dataset file(s).

It's also possible to access a registered Dataset directly. 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')

Note

The preceding snippets show the form of the calls and are not part of a Microsoft sample. You must replace the various arguments with values from your own project.

Use OutputFileDatasetConfig for intermediate data

While Dataset objects represent only persistent data, OutputFileDatasetConfig object(s) can be used for temporary data output from pipeline steps and persistent output data. OutputFileDatasetConfig supports writing data to blob storage, fileshare, adlsgen1, or adlsgen2. It supports both mount mode and upload mode. In mount mode, files written to the mounted directory are permanently stored when the file is closed. In upload mode, files written to the output directory are uploaded at the end of the job. If the job fails or is canceled, the output directory won't be uploaded.

OutputFileDatasetConfig object's default behavior is to write to the default datastore of the workspace. Pass your OutputFileDatasetConfig objects to your PythonScriptStep with the arguments parameter.

from azureml.data import OutputFileDatasetConfig
dataprep_output = OutputFileDatasetConfig()
input_dataset = Dataset.get_by_name(workspace, 'raw_data')

dataprep_step = PythonScriptStep(
    name="prep_data",
    script_name="dataprep.py",
    compute_target=cluster,
    arguments=[input_dataset.as_named_input('raw_data').as_mount(), dataprep_output]
    )

Note

Concurrent writes to a OutputFileDatasetConfig will fail. Do not attempt to use a single OutputFileDatasetConfig concurrently. Do not share a single OutputFileDatasetConfig in a multiprocessing situation, such as when using distributed training.

Use OutputFileDatasetConfig as outputs of a training step

Within your pipeline's PythonScriptStep, you can retrieve the available output paths using the program's arguments. If this step is the first and will initialize the output data, you must create the directory at the specified path. You can then write whatever files you wish to be contained in the OutputFileDatasetConfig.

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")

Read OutputFileDatasetConfig as inputs to non-initial steps

After the initial pipeline step writes some data to the OutputFileDatasetConfig path and it becomes an output of that initial step, it can be used as an input to a later step.

In the following code:

  • step1_output_data indicates that the output of the PythonScriptStep, step1 is written to the ADLS Gen 2 datastore, my_adlsgen2 in upload access mode. Learn more about how to set up role permissions in order to write data back to ADLS Gen 2 datastores.

  • After step1 completes and the output is written to the destination indicated by step1_output_data, then step2 is ready to use step1_output_data as an input.

# get adls gen 2 datastore already registered with the workspace
datastore = workspace.datastores['my_adlsgen2']
step1_output_data = OutputFileDatasetConfig(name="processed_data", destination=(datastore, "mypath/{run-id}/{output-name}")).as_upload()

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

)

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

Tip

Reading the data in the python script step2.py is the same as documented earlier in Access datasets within your script; use ArgumentParser to add an argument of --pd in your script to access the data.

Register OutputFileDatasetConfig objects for reuse

If you'd like to make your OutputFileDatasetConfig available for longer than the duration of your experiment, register it to your workspace to share and reuse across experiments.

step1_output_ds = step1_output_data.register_on_complete(
    name='processed_data', 
    description = 'files from step1'
)

Delete OutputFileDatasetConfig contents when no longer needed

Azure doesn't automatically delete intermediate data written with OutputFileDatasetConfig. To avoid storage charges for large amounts of unneeded data, you should either:

Caution

Only delete intermediate data after 30 days from the last change date of the data. Deleting the data earlier could cause the pipeline run to fail because the pipeline will assume the intermediate data exists within 30 day period for reuse.

  • Programmatically delete intermediate data at the end of a pipeline job, when it's no longer needed.
  • Use blob storage with a short-term storage policy for intermediate data (see Optimize costs by automating Azure Blob Storage access tiers). This policy can only be set to a workspace's non-default datastore. Use OutputFileDatasetConfig to export intermediate data to another datastore that isn't the default.
    # Get adls gen 2 datastore already registered with the workspace
    datastore = workspace.datastores['my_adlsgen2']
    step1_output_data = OutputFileDatasetConfig(name="processed_data", destination=(datastore, "mypath/{run-id}/{output-name}")).as_upload()
    
  • Regularly review and delete no-longer-needed data.

For more information, see Plan and manage costs for Azure Machine Learning.

Next steps