Azure Machine Learning SDK v2 for Python。用于ML工作区、作业、模型、数据集、计算和管道。下载地址agentic-awesome-skills/skills/azure-ai-ml-py at main · sickn33/agentic-awesome-skills · GitHub技能概述azure-ai-ml-py 技能是管理Azure ML资源的Python客户端库。该SDK提供了现代化的API来管理工作区、作业、模型、数据和计算资源。支持构建端到端的机器学习工作流程从数据准备到模型训练、注册和部署。安装pip install azure-ai-ml环境变量AZURE_SUBSCRIPTION_ID AZURE_RESOURCE_GROUP AZURE_ML_WORKSPACE_NAME认证from azure.ai.ml import MLClientfrom azure.identity import DefaultAzureCredentialml_client MLClient(credentialDefaultAzureCredential(),subscription_idos.environ[AZURE_SUBSCRIPTION_ID],resource_group_nameos.environ[AZURE_RESOURCE_GROUP],workspace_nameos.environ[AZURE_ML_WORKSPACE_NAME])从配置文件认证from azure.ai.ml import MLClientfrom azure.identity import DefaultAzureCredential# Uses config.json in current directory or parentml_client MLClient.from_config(credentialDefaultAzureCredential())工作区管理创建工作区from azure.ai.ml.entities import Workspacews Workspace(namemy-workspace,locationeastus,display_nameMy Workspace,descriptionML workspace for experiments,tags{purpose: demo})ml_client.workspaces.begin_create(ws).result()列出工作区for ws in ml_client.workspaces.list():print(f{ws.name}: {ws.location})数据资产注册数据from azure.ai.ml.entities import Datafrom azure.ai.ml.constants import AssetTypes# Register a filemy_data Data(namemy-dataset,version1,pathazureml://datastores/workspaceblobstore/paths/data/train.csv,typeAssetTypes.URI_FILE,descriptionTraining data)ml_client.data.create_or_update(my_data)注册文件夹my_data Data(namemy-folder-dataset,version1,pathazureml://datastores/workspaceblobstore/paths/data/,typeAssetTypes.URI_FOLDER)ml_client.data.create_or_update(my_data)模型注册表注册模型from azure.ai.ml.entities import Modelfrom azure.ai.ml.constants import AssetTypesmodel Model(namemy-model,version1,path./model/,typeAssetTypes.CUSTOM_MODEL,descriptionMy trained model)ml_client.models.create_or_update(model)列出模型for model in ml_client.models.list(namemy-model):print(f{model.name} v{model.version})计算资源创建计算集群from azure.ai.ml.entities import AmlComputecluster AmlCompute(namecpu-cluster,typeamlcompute,sizeStandard_DS3_v2,min_instances0,max_instances4,idle_time_before_scale_down120)ml_client.compute.begin_create_or_update(cluster).result()列出计算资源for compute in ml_client.compute.list():print(f{compute.name}: {compute.type})作业命令作业from azure.ai.ml import command, Inputjob command(code./src,commandpython train.py --data ${{inputs.data}} --lr ${{inputs.learning_rate}},inputs{data: Input(typeuri_folder, pathazureml:my-dataset:1),learning_rate: 0.01},environmentAzureML-sklearn-1.0-ubuntu20.04-py38-cpulatest,computecpu-cluster,display_nametraining-job)returned_job ml_client.jobs.create_or_update(job)print(fJob URL: {returned_job.studio_url})监控作业ml_client.jobs.stream(returned_job.name)管道from azure.ai.ml import dsl, Input, Outputfrom azure.ai.ml.entities import Pipelinedsl.pipeline(computecpu-cluster,descriptionTraining pipeline)def training_pipeline(data_input):prep_step prep_component(datadata_input)train_step train_component(dataprep_step.outputs.output_data,learning_rate0.01)return {model: train_step.outputs.model}pipeline training_pipeline(data_inputInput(typeuri_folder, pathazureml:my-dataset:1))pipeline_job ml_client.jobs.create_or_update(pipeline)环境创建自定义环境from azure.ai.ml.entities import Environmentenv Environment(namemy-env,version1,imagemcr.microsoft.com/azureml/openmpi4.1.0-ubuntu20.04,conda_file./environment.yml)ml_client.environments.create_or_update(env)数据存储列出数据存储for ds in ml_client.datastores.list():print(f{ds.name}: {ds.type})获取默认数据存储default_ds ml_client.datastores.get_default()print(fDefault: {default_ds.name})MLClient操作属性操作workspacescreate, get, list, deletejobscreate_or_update, get, list, stream, cancelmodelscreate_or_update, get, list, archivedatacreate_or_update, get, listcomputebegin_create_or_update, get, list, deleteenvironmentscreate_or_update, get, listdatastorescreate_or_update, get, list, get_defaultcomponentscreate_or_update, get, list最佳实践使用版本控制- 为数据、模型和环境使用版本控制配置空闲缩减- 设置空闲时间自动缩减以降低计算成本使用环境- 为可重复训练创建可重用环境流式传输作业日志- 监控训练进度注册模型- 成功训练后注册模型使用管道- 为多步骤工作流使用管道标记资源- 为组织和成本跟踪标记资源触发条件需要在Python应用中管理Azure ML工作区需要提交和管理训练作业需要注册和管理ML模型需要创建和管理计算集群需要构建ML管道
