SageMaker, AWS's machine learning management service, doesn't come up with the materials I want in a pinpoint manner, so I'll summarize them.
When operating an ML system in a production environment, the managed function shows its true value. A component for realizing MLops and MDLC. PoC model development alone does not provide real benefits.
Machine Learning Lens, AWS Well-Architected Framework https://d1.awsstatic.com/whitepapers/architecture/wellarchitected-Machine-Learning-Lens.pdf Guidelines for ML systems built on AWS
Developer guide, SDK reference, etc. https://docs.aws.amazon.com/ja_jp/sagemaker/index.html
GitHub https://github.com/awslabs/amazon-sagemaker-examples/blob/master/README.md
GitHub (Japanese) https://github.com/aws-samples/amazon-sagemaker-examples-jp
Amazon SageMaker Discussion Forums https://forums.aws.amazon.com/forum.jspa?forumID=285
Documents by AWS service https://aws.amazon.com/jp/aws-jp-introduction/aws-jp-webinar-service-cut/#ai-wn
MDLC : Model Development Life Cycle "Scaleable model deployment and automation on Amazon.com" https://aws.amazon.com/jp/blogs/news/aws-aiml-tokyo2/
In this session, we introduced the concept of model development life cycle (MDLC) of machine learning that customers should consider when introducing machine learning, and the workflow for executing MDLC. This MDLC is actually the method adopted by Amazon Consumer Payments and was also introduced in the session at re: Invent 2019.
Amazon SageMaker Basics https://pages.awscloud.com/rs/112-TZM-766/images/20191128_Amazon%20SageMaker_Basic.pdf
[AWS Black Belt Online Seminar] Amazon SageMaker Advanced Session materials and QA released https://aws.amazon.com/jp/blogs/news/webinar-bb-amazon-sagemaker-advanced-session-2019/?nc1=f_ls
Get the big picture of machine learning with SageMaker and xGBoost https://qiita.com/suzukihi724/items/3792f395fb22cf7fb311
How to throw when using the AWS SageMaker built-in method https://qiita.com/kazuhisa-nagashima/items/21d80271da733d15f4b0
How to pass environment variables from the host OS side to the guest OS (container) with Docker (various types) https://qiita.com/KEINOS/items/518610bc2fdf5999acf2 I was wondering why environment variables are used to pass parameters, but it seems that the theory is to use environment variables to convey information to the container.
I made a container image that can use LightGBM with Amazon SageMaker https://dev.classmethod.jp/articles/sagemaker-container-image-lightgbm/
Create your own learning / inference container image with Amazon SageMaker https://dev.classmethod.jp/articles/sagemaker-container-image-custom/
GitHub (Japanese) https://github.com/aws-samples/amazon-sagemaker-examples-jp/tree/master/workshop/lab_bring-your-own-containers
[Developer's Guide] Using your own inference code in your hosting service https://docs.aws.amazon.com/ja_jp/sagemaker/latest/dg/your-algorithms-inference-code.html
Application development with Amazon SageMaker inference endpoints https://www.slideshare.net/AmazonWebServicesJapan/amazon-sagemaker-122749918
Build a serverless front end for Amazon SageMaker endpoints https://aws.amazon.com/jp/blogs/news/build-a-serverless-frontend-for-an-amazon-sagemaker-endpoint/
Build, test, and deploy Amazon Sagemaker inference models to AWS Lambda https://aws.amazon.com/jp/blogs/news/build-test-and-deploy-your-amazon-sagemaker-inference-models-to-aws-lambda/
20190206 AWS Black Belt Online Seminar Amazon SageMaker Basic Session https://www.slideshare.net/AmazonWebServicesJapan/20190206-aws-black-belt-online-seminar-amazon-sagemaker-basic-session-130777850 Described in the reference at the end
ML Security on AWS https://pages.awscloud.com/rs/112-TZM-766/images/A2-07.pdf
How to solve the "Don't go out to the Internet" requirement on AWS-Private Link support- https://bcblog.sios.jp/aws-privatelink/
Explain how to use AWS Private Link https://blog.mmmcorp.co.jp/blog/2017/11/15/aws_privatelink/
How to use AWS Private Link and precautions-Use properly with VPC peering- https://devlog.arksystems.co.jp/2018/05/11/4896/
https://aws.amazon.com/jp/blogs/news/amazon-sagemaker-handson-20190517/
Machine Learning Model Development Flow on AWS SageMaker (PyTorch) https://qiita.com/noko_qii/items/41130f66afbb8e451f23
I tried using my own container with Amazon SageMaker Processing https://dev.classmethod.jp/articles/amazon-sagemaker-processing-original-container/ Almost the same as SKLearnProcessor except that a container with the necessary libraries installed is prepared.
Use built-in container: SKLearnProcessor https://docs.aws.amazon.com/ja_jp/sagemaker/latest/dg/use-scikit-learn-processing-container.html
Use your own container: ScriptProcessor
Amazon SageMaker Processing – Fully Managed Data Processing and Model Evaluation https://aws.amazon.com/jp/blogs/news/amazon-sagemaker-processin-fully-managed-data-processing-and-model-evaluation/
CI / CD pipeline for ML with Amazon SageMaker https://pages.awscloud.com/rs/112-TZM-766/images/E-3.pdf
How to combine your own containers in Amazon SageMaker's inference pipeline https://qiita.com/yaiwase/items/79f99d2c38ed66729a47
AWS Step Functions AWS Step Functions https://docs.aws.amazon.com/ja_jp/step-functions/latest/dg/connect-sagemaker.html
Pattern 1: StepFunctions-> SageMaker API (Available with the SageMaker API installed in Setp Functions)
Create a Step Functions API using API Gateway https://docs.aws.amazon.com/ja_jp/step-functions/latest/dg/tutorial-api-gateway.html If you want to accept from API Gateway.
AWS Step Functions Data Science SDK for Python https://docs.amazonaws.cn/en_us/step-functions/latest/dg/concepts-python-sdk.html
Introducing "AWS Step Functions Data Science SDK" to easily create a state machine https://dev.classmethod.jp/articles/yoshim_step_functions_datascience_sdk/
The API of SageMaker that can be handled by Step Functions has increased https://dev.classmethod.jp/articles/stepfunctions-sagemaker-api-update/
Pattern 2: StepFunctions-> Lambda-> SageMaker API Script execution using your own processing container https://docs.aws.amazon.com/ja_jp/sagemaker/latest/dg/processing-container-run-scripts.html
Manage ML workflows serverless with AWS Step Functions: Amazon SageMaker Advent Calendar 2018 https://dev.classmethod.jp/articles/2018advent-calendar-sagemaker-20181210/
Manage machine learning workflows with AWS StepFunctions https://qiita.com/kurakura0916/items/5e89cb86e86d22fdc5d8
Create a batch process that combines Lambda with AWS Step Functions https://dev.classmethod.jp/articles/aws-step-functions-batch-service/
[GitHub] AWS Step Functions Data Science SDK - Hello World https://github.com/awslabs/amazon-sagemaker-examples/blob/master/step-functions-data-science-sdk/hello_world_workflow/hello_world_workflow.ipynb
Automating Amazon SageMaker workflows with AWS Step Functions https://www.youtube.com/watch?v=0kMdOi69tjQ
Orchestrate Machine Learning Workflows with Amazon SageMaker and AWS Step Functions https://www.youtube.com/watch?v=dNb5jVffzPs
20190522 AWS Black Belt Online Seminar AWS Step Functions https://www.slideshare.net/AmazonWebServicesJapan/20190522-aws-black-belt-online-seminar-aws-step-functions
Building an AWS Serverless ML Pipeline with Step Functions https://tech.olx.com/building-an-aws-serverless-ml-pipeline-with-step-functions-b39feed12bab
I tried looping and branching for the first time using Step Functions! https://dev.classmethod.jp/articles/first-aws-step-functions/ Step Functions + Lambda hands-on
AWS Lambda Build a CI / CD pipeline at once using AWS Lambda application creation https://qiita.com/shonansurvivors/items/b223fbb362aed3c1c536
AWS GLUE AWS GLUE concept https://docs.aws.amazon.com/ja_jp/glue/latest/dg/components-key-concepts.html
How to use Glue https://qiita.com/pioho07/items/32f76a16cbf49f9f712f
What I did to handle 5TB / day data with AWS Glue (Overview) https://future-architect.github.io/articles/20180828/ The comparison between EMR and Glue is very helpful.
MLOps Automate building machine learning models with Lambda and lowering makers https://medium.com/@yuyasugano/lambda%E3%81%A8%E4%B8%8B%E3%81%92%E3%83%A1%E3%83%BC%E3%82%AB%E3%83%BC%E3%81%A7%E6%A9%9F%E6%A2%B0%E5%AD%A6%E7%BF%92%E3%83%A2%E3%83%87%E3%83%AB%E3%81%AE%E6%A7%8B%E7%AF%89%E3%82%92%E8%87%AA%E5%8B%95%E5%8C%96-73161d316c0e
An event-driven machine learning platform that orchestrate with Step functions and aws batch https://www.slideshare.net/yuyamada777/step-functionsaws-batch
Articles that I referred to when studying around Machine Learning Platform / MLOps (as of December 2018) https://qiita.com/noko_qii/items/f31901817dbed86f2b25
Mechanism of hyperparameter adjustment https://docs.aws.amazon.com/ja_jp/sagemaker/latest/dg/automatic-model-tuning-how-it-works.html
Understanding the Optuna (TPE) Algorithm-Part 1- https://qiita.com/nabenabe0928/items/708d221dbccebf31f01c
GitHub (Japanese) https://github.com/aws-samples/amazon-sagemaker-examples-jp/blob/master/hpo_pytorch_mnist/pytorch_mnist.ipynb
Optuna hyperparameter optimization framework https://www.slideshare.net/pfi/pydatatokyo-meetup-21-optuna Define-by-Run: Define the search space when the objective function is executed
Use Optuna with SageMaker https://aws.amazon.com/jp/blogs/machine-learning/implementing-hyperparameter-optimization-with-optuna-on-amazon-sagemaker/
SageMaker HPO: First, the search space is decided. Determine the parameters in Optuna: for loop. Dynamically determine the search space ・ Multiple trials run in one training job. ・ After one trial, write the parameters to the aurora. Look at the past parameters and decide the next parameter. pytorch-simple https://github.com/aws-samples/amazon-sagemaker-optuna-hpo-blog/blob/master/examples/pytorch_simple/src/pytorch_simple.py
・ Objective (trial) Trials, experiments, etc. -Tuning the number of layer units. Tuning the number of layers 46th line n_layers = trial.suggest_int("n_layers", 1, 3) -Optuna has an RDS back end, so RDS and Aulora are needed. -The number of units is tuned on the 51st line. out_features = trial.suggest_int("n_units_l{}".format(i), 4, 128)
AutoML Can't you automate that machine learning process? https://qiita.com/Hironsan/items/30fe09c85da8a28ebd63
What is Amazon QuickSight? https://docs.aws.amazon.com/ja_jp/quicksight/latest/user/welcome.html
AWS Manga Episode 8: Visualize All Data! (1/8) https://aws.amazon.com/jp/campaigns/manga/vol8-1/ There is hands-on material at the bottom
10 Visualizations to Try with Amazon QuickSight Using Sample Data https://aws.amazon.com/jp/blogs/news/10-visualizations-to-try-in-amazon-quicksight-with-sample-data/
[Official] Use of ML Insight https://docs.aws.amazon.com/ja_jp/quicksight/latest/user/making-data-driven-decisions-with-ml-in-quicksight.html Random cut forest is used for both anomaly detection and numerical prediction.
Visualizing Amazon SageMaker machine learning predictions with Amazon QuickSight https://aws.amazon.com/jp/blogs/machine-learning/making-machine-learning-predictions-in-amazon-quicksight-and-amazon-sagemaker/ Cooperation between SageMaker and QuickSight
Use a custom container Use your own algorithms and models with Amazon SageMaker https://docs.aws.amazon.com/ja_jp/sagemaker/latest/dg/your-algorithms.html
GitHub https://github.com/awslabs/amazon-sagemaker-examples/tree/master/advanced_functionality/r_bring_your_own
Introducing Amazon SageMaker Operators for Kubernetes https://aws.amazon.com/jp/blogs/news/introducing-amazon-sagemaker-operators-for-kubernetes/ Can be used from EKS. It can also be used with plain Kubernetes.
Amazon SageMaker Components for Kubeflow Pipelines https://aws.amazon.com/about-aws/whats-new/2020/06/amazon-sagemaker-components-kubeflow-pipelines/ SageMaker can be incorporated into Kubeflow
Distributed learning and workflow for large-scale data realized by Amazon SageMaker | AWS Summit Tokyo 2019 https://www.youtube.com/watch?v=NUnIiYD-PEU&t=714s AWS CodeCommit EFS: Data sharing between instances / users File access / pipe mode Parameters for large data of 100GB HPO Amazon Elastic Inference Batch conversion Data Preparation Best Practices (Glue / Athena) Pack the dataset into a RecordIO / TFRecord file Use batch conversion if you don't need online forecasting Use inference pipeline instead of multiple endpoints
"Common problems in actual machine learning operations and solutions / case studies using AWS" | AWS Summit Tokyo 2019 https://www.youtube.com/watch?v=a3smIzBC6BQ
[Beginner] Introduction to AWS Machine Learning Services | AWS Summit Tokyo 2019 https://www.youtube.com/watch?v=1gC46ODyudE
https://www.youtube.com/watch?v=L4bLDNRSYC8
【AWS Black Belt Online Seminar】Amazon SageMaker Advanced Session https://www.youtube.com/watch?v=G-s67PmTCjo&t=2496s
【AWS Black Belt Online Seminar】AWS AI Language Services https://www.youtube.com/watch?v=Q0Ety9Z7oWM
【AWS Black Belt Online Seminar】AWS AI Services https://www.youtube.com/watch?v=xvUyKjuv-Z4&t=1183s Rekognition inference architecture available
Amazon API Gateway 【AWS Black Belt Online Seminar】Amazon API Gateway https://www.youtube.com/watch?v=EpEETIox03s&list=RDCMUCnjKWUK2t5QJYfeqqilhJhQ&index=8
Recommended Posts