Use MLflow to manage and deploy Machine Learning model on Spark
4 likes3,283 views
MLflow is a platform for managing the machine learning lifecycle, including tracking experiments, packaging models into reproducible projects, and deploying models into production. It provides APIs and tools to help with experiment tracking, model deployment, and model monitoring. MLflow supports many frameworks like PyTorch, TensorFlow, Keras, and Spark ML. It integrates with cloud platforms like Azure ML, AWS SageMaker, and Databricks to enable deployment and management of models on those platforms.
![mlflow.set_tracking_uri(). [remote tracking URIs]
Local file path (specified as file:/my/local/dir)
Database encoded as
<dialect>+<driver>://<username>:<password>@<host>:<port>/<d
atabase
MLFlow tracking server (specified as https://my-server:5000
Databricks workspace (specified as databricks or as
databricks://<profileName>e.](https://image.slidesharecdn.com/sparkmlflow-190906150820/85/Use-MLflow-to-manage-and-deploy-Machine-Learning-model-on-Spark-15-320.jpg)
![MLFLOW_TRACKING_URI=http://0.0.0.0:5000 mlflow sklearn serve
--port 5001
--run_id XXXXXXXXXXXXXXXXXXXXXX
--model-path model
curl -X POST
http://127.0.0.1:5001/invocations
-H 'Content-Type: application/json'
-d '[
{
“XXX": 1.111,
“YYYY": 1.22,
“ZZZZ": 1.888
}
]'](https://image.slidesharecdn.com/sparkmlflow-190906150820/85/Use-MLflow-to-manage-and-deploy-Machine-Learning-model-on-Spark-31-320.jpg)
Use MLflow to manage and deploy Machine Learning model on Spark
- 3. Building a model Building a model Data ingestion Data analysis Data transformation Data validation Data splitting Trainer Model validation Training at scale LoggingRoll-out Serving Monitoring
- 4. Train Model Validate Model Deploy ModelPackage Model Monitor Model Retrain Model
- 8. 模型追蹤：記錄和查詢模型訓練的資料，如Accuracy 和各種參數 專案管理：將模型封裝在 pipeline 中，以便與可重複執行 模型管理：管理模型部署並提供 呼叫 API
- 13. Amazon S3 Azure Blob Storage Google Cloud Storage FTP server SFTP Server NFS HDFS
- 14. mlflow server --backend-store-uri /home/hermanwu/mlflowdata --default-artifact-root wasbs://artifacts@hmlflow.blob.core.windows.net --host 13.75.XXX.XXX export AZURE_STORAGE_ACCESS_KEY=ukmcWZA1l9ZK1M17V/SfHXzQN7jRL5+/I8KAIk2Mjwe emCFSmBJ85V18kz7Qvt7Aj5JihKxxxxxxxxxxxxxx==
- 15. mlflow.set_tracking_uri(). [remote tracking URIs] Local file path (specified as file:/my/local/dir) Database encoded as <dialect>+<driver>://<username>:<password>@<host>:<port>/<d atabase MLFlow tracking server (specified as https://my-server:5000 Databricks workspace (specified as databricks or as databricks://<profileName>e.
- 16. Framework Metrics Parameters Tags Artifacts Keras Training loss; validation loss; user-specified metrics Number of layers; optimizer name; learning rate; epsilon Model summary MLflow Model (Keras model), TensorBoard logs; on training end tf.keras Training loss; validation loss; user-specified metrics Number of layers; optimizer name; learning rate; epsilon Model summary MLflow Model (Keras model), TensorBoard logs; on training end tf.estimator TensorBoard metrics – – MLflow Model (TF saved model); on call to tf.estimator.export_saved_ model TensorFlow Core All tf.summary.scalar cal ls – – –
- 17. “ ”
- 18. “ ”
- 21. mlflow run sklearn_elasticnet_wine -P alpha=0.5 mlflow run https://github.com/mlflow/mlflow-example.git -P alpha=5
- 24. ❖ Custom Flavors
- 26. mlflow models --help mlflow models serve --help mlflow models predict --help mlflow models build-docker --help
- 27. azure_image, azure_model = mlflow.azureml.build_image(model_uri="<path-to-model>", workspace=azure_workspace, description="Wine regression model 1", synchronous=True) webservice_deployment_config = AciWebservice.deploy_configuration() webservice = Webservice.deploy_from_image( image=azure_image, workspace=azure_workspace, name="<deployment-name>") webservice.wait_for_deployment()
- 28. “ ”
- 29. pyfunc_udf = mlflow.pyfunc.spark_udf(<path-to-model>) df = spark_df.withColumn("prediction", pyfunc_udf(<features>))
- 30. %%PySpark import mlflow from mlflow import pyfunc pyfunc_udf = mlflow.pyfunc.spark_udf(<path-to-model>) spark.udf.register("pyfunc ", pyfunc_udf ) %%SQL SELECT id, pyfunc( feature01, feature02, feature03, ….. ) AS prediction FROM tempPredict LIMIT 20 df.createOrReplaceTempView(“tempPredict")
- 31. MLFLOW_TRACKING_URI=http://0.0.0.0:5000 mlflow sklearn serve --port 5001 --run_id XXXXXXXXXXXXXXXXXXXXXX --model-path model curl -X POST http://127.0.0.1:5001/invocations -H 'Content-Type: application/json' -d '[ { “XXX": 1.111, “YYYY": 1.22, “ZZZZ": 1.888 } ]'
- 32. mlflow models serve -m runs:/<RUN_ID>/model --port 5050
- 33. mlflow mlflow.azureml mlflow.entities mlflow.h2o mlflow.keras mlflow.mleap mlflow.models mlflow.onnx mlflow.projects mlflow.pyfunc Filesystem format Inference API Creating custom Pyfunc models mlflow.pytorch mlflow.sagemaker mlflow.sklearn mlflow.spark mlflow.tensorflow mlflow.tracking
- 35. Create Experiment List Experiments Get Experiment Delete Experiment Restore Experiment Update Experiment Create Run Delete Run Restore Run Get Run Log Metric Log Batch Set Experiment Tag Set Tag Delete Tag Log Param Get Metric History Search Runs List Artifacts Update Run Data Structures
- 44. Use tempfile.TemporaryDirectory + mlflow.log_artifacts To upload artifices with TemporaryDirectory(prefix='temp_arti_', dir='temp_artifacts') as dirname: …… (create artifcats ) …….. mlflow.log_artifacts(dirname)
- 45. Train model Validate model Deploy model Monitor model Retrain model Model reproducibility Model retrainingModel deploymentModel validation Build appCollaborate Test app Release app Monitor app ML DevOps integration App developer using DevOps Services Data scientist using Machine Learning
- 48. A M L & M L F L O W M O D E L S The mlflow.azureml module can export python_function models as Azure ML compatible models. It can also be used to directly deploy and serve models on Azure ML, provided the environment has been correctly set up. ▪ export the model in Azure ML-compatible format. MLflow will output a directory with the dependencies necessary to deploy the model. ▪ deploy deploys the model directly to Azure ML. You first need to set up your environment to work with the Azure ML CLI. You also have to set up all accounts required to run and deploy on Azure ML. Where the model is deployed is dependent on your active Azure ML environment. If the active environment is set up for local deployment, the model will be deployed locally in a Docker container (Docker is required). mlflow.azureml.build_image(model_path, workspace, run_id=None,image_name=None, model_name=None,mlflow_home=None, description=None, tags=None, synchronous=True)
- 49. ▪ Experiment Tracking ▪ MLflow lets you run experiments with any ML library, framework, or language, and automatically keeps track of parameters, results, code, and data from each experiment so that you can compare results and find the best performing runs. ▪ With Managed MLflow on Databricks, you can now track, share, visualize, and manage experiments securely from within the Databricks Workspace and notebooks. ▪ Reproducible Projects ▪ MLflow lets you package projects with a standard format that integrates with Git and Anaconda and capture dependencies like libraries, parameters, and data. ▪ With Managed MLflow on Databricks, now you can quickly launch reproducible runs remotely from your laptop as a Databricks job. ▪ Productionize models faster ▪ MLflow lets you quickly deploy production models for batch inference on Apache SparkTM, or as REST APIs using built-in integration with Docker containers, Azure ML, or Amazon SageMaker. ▪ With Managed MLflow on Databricks, now you can operationalize and monitor production models using Databricks Jobs Scheduler and auto-managed Clusters to scale as needed based on business needs.
- 50. A M L & D A T A B R I C K S Choose only one option Easily install the AML Python SDK in the Azure Databricks clusters and use it for: ✓ logging training run metrics ✓ containerize Spark ML models ✓ deploy them into ACI or AKS