Alibaba PAI Python SDK

AliPAI Python SDK is provided by PAI team of Alibaba computing platform. It provides convenience for users to access PAI service in Alibaba Cloud.

In current, PAI SDK supports PAIFlow(ML Pipeline Service of PAI) service, other PAI services, such as EAS(Elastic Algorithm Service) and Blade will be included soon.

Installation

To install the PAI sdk, use the below command in terminal.

python -m pip install alipai

Usage

Setup default PAI session

Before use PAI service via SDK, developer should initialize the default PAI session by providing credential and region_id of service.

Pipeline service of PAI is currently provided in cn-shanghai region only.

from pai.core.session import setup_default_session

session = setup_default_session(access_key_id="your_access_key", access_key_secret="your_access_secret", region_id="your_region_id")

Access Pipeline Service

Use PipelineTemplate

PipelineTemplate instance includes the definition of "Workflow" use in PAI pipeline service. It could be fetched from remote PAI service or constructed from local Pipeline/Component.

Saved pipeline template has unique pipeline_id which is generated by pipeline service. Remote pipeline template could be fetched using identifier-provider-version or pipeline_id.

PAI provides a list of public pipeline templates which could be used as workflow template to run or to build pipeline. These templates are accessible by the specific provider pai.common.ProviderAlibabaPAI in PipelineTemplate.list.

from pai.pipeline import PipelineTemplate
from pai.common import ProviderAlibabaPAI

# search PipelineTemplate which provide by `PAI` and include `xflow` in identifier.
template = next(PipelineTemplate.list(identifie="xflow", provider=ProviderAlibabaPAI))

# view template inputs/outputs.
template
template.inputs
template.outputs

After submitting run job, users are able to inspect the detailed workflow DAG, execution log and outputs of the pipeline by visiting the job detail URL printed in console.

from pai.common import ProviderAlibabaPAI
from pai.pipeline import PipelineTemplate

# Get specific template by Identifier-Provider-Version
template = PipelineTemplate.get_by_identifier(identifier="split-xflow-maxCompute",
    provider=ProviderAlibabaPAI, version="v1")

xflow_execution = {
    "odpsInfoFile": "/share/base/odpsInfo.ini",
    "endpoint": "http://service.cn-shanghai.maxcompute.aliyun.com/api",
    "logViewHost": "http://logview.odps.aliyun.com",
    "odpsProject": "your_odps_project",
}

# run pipeline use provide arguments.
job = template.run(job_name="demo-split-job", arguments={
    "inputArtifact": "odps://pai_online_project/tables/mnist_data",
    "execution": xflow_execution, "fraction": 0.7}, wait=True)
job.get_outputs()

Build runnable and reusable pipeline

PAI Pipeline Service supports nested user-defined workflow. Composite pipeline is runnable by providing required arguments. Saved pipeline template could be used as a step to build a new pipeline.

def create_composite_pipeline():
    # Definite the inputs parameters in pipeline
    execution_input = PipelineParameter(name="execution", typ=dict)
    cols_to_double_input = PipelineParameter(name="cols_to_double")

    table_input = PipelineArtifact(name="data_source", metadata=ArtifactMetadata(
            data_type=ArtifactDataType.DataSet,
            location_type=ArtifactLocationType.MaxComputeTable))

    # Pipeline step from remote PAI service.
    type_transform_step = PipelineStep(
        identifier="type-transform-xflow-maxCompute", provider=ProviderAlibabaPAI,
        version="v1", name="typeTransform", inputs={
            "inputArtifact": table_input, "execution": execution_input, 
            "outputTable": gen_temp_table(), "cols_to_double": cols_to_double_input,
        }
    )

    split_template = PipelineTemplate.get_by_identifier(identifier="split-xflow-maxCompute",
     provider=ProviderAlibabaPAI, version="v1")
    split_step = split_template.as_step(inputs={"inputArtifact": type_transform_step.outputs[0],
            "execution": execution_input, "output1TableName": gen_temp_table(),
            "fraction": 0.5, "output2TableName": gen_temp_table(),
        })

    # Initialize the pipeline instance by specific the steps and outputs.   
    p = Pipeline(
        steps=[split_step],
        outputs=split_step.outputs[:2],
    )
    return p

p = create_composite_pipeline()
# Run pipeline with required arguments.
pipeline_run = p.run(job_name="demo-composite-pipeline-run", arguments={
            "execution": xflow_execution,
            "cols_to_double": "time,hour,pm2,pm10,so2,co,no2",
            "data_source": "odps://pai_online_project/tables/wumai_data",
        }, wait=True)

# Save Pipeline
p.save(identifier="demo-composite-pipeline", version="v1")