ML_Development_with_Sagemaker.pptx
- 1. DEVELOPING ML SOLUTIONS WITH SAGEMAKER Knowledge Share by Temiloluwa Adeoti | October 12
- 2. 2 Strategies for Developing ML Solutions on AWS Polly – Text to Speech Rekognition – Computer Vision Comprehend – Natural Language Processing Lex – Chatbots AWS AI/ML Services 01 Pretrained Models on AWS Market place 02 A collection of: - Pre-trained models - Solution Templates 03 - EMR Jupyter or Zepplin Notebooks - Notebooks with Glue Dev endpoint - Cloud 9 development - Sagemaker Jupyter Notebooks - Sagemaker Studio 04 Subscribe to pre-trained models and deploy them to Sagemaker from web console Accelerated ML development with Sagemaker JumpStart Dev Enviroments for .Py and .Ipynb files
- 3. 3 What is Sagemaker? Sagemaker is a fully-managed AWS Service that facilitates almost every part of the ML development lifecycle. Data • Sagemaker Groundtruth - automated data labelling • Sagemaker Data Wrangler – data preprocessing pipelines • Sagemaker Clarify – bias detection Training • Sagemaker Experiments – experiments tracking • Sagemaker Autopilot – auto ml • Sagemaker Feature Store – feature store Inference • Sagemaker Monitor – detect drift in models • Sagemaker Elastic Inference – lower cost inference acceleration • Sagemaker Neo – compile model for multiple hardware platforms Full list of Features: https://docs.aws.amazon.com/sagemaker/latest/dg/whatis.html
- 4. 4 My Personal Challenges with Sagemaker 1. Poor Documentation 2. APIs with a deluge of Parameters 3. Docker containers are king and debugging them are difficult
- 5. 5 Ways to interact with Sagemaker resources 1 AWS Console – Web Interface 2 AWS CLI – command line 3 AWS SDK for python (Boto3) – Low level API 4 Sagemaker Python SDK – High level API
- 6. 6 Sagemaker Development Environments 01 Sagemaker Jupyter Notebooks Similar to Google Colab but you have to manually create your notebook instance 02 Sagemaker Studio Jupyter Lab type of experience that provides full access to Sagemaker’s resources
- 7. 7 Jupyter Notebooks Disadvantages • They are slow to startup: 5 – 10 times slower than Sagemaker Studio Notebooks • Lacks the integrated Notebook sharing features present in Sagemaker Notebooks • Development environment has a fixed instance type o You can switch the instance type on which your notebook should run in Sagemaker Studio o Better cost savings on with Sagemaker Studio
- 8. 8 Sagemaker Studio Architecture Source: https://docs.aws.amazon.com/sagemaker/latest/dg/notebooks.html AWS Service Account - JupyterServer App - KernelGateway Apps Customer Account - Amazon EFS
- 9. 9 Basic workflow with a Sagemaker Studio Notebook JupyterServerApp created from Jupyter notebooks Infrastructure: (EFS, Jupyter Server) KernelGatewayApp created from Docker Images Infrastructure: (ml.t3.medium) CreateTrainingJob Infrastructure: (ml.c5.xlarge) Source: https://docs.aws.amazon.com/sagemaker/latest/dg/notebooks.html
- 10. 10 Sagemaker Instance Types • ml.t3.medium (Free-Tier Eligible) >> Fast launch • ml.m5.large >> Fast launch General purpose (no GPUs) 1 Compute Optimized (No GPUs) 2 • ml.c5.large >> Fast launch Accelerated computing (1+ GPUs) 3 • ml.g4dn.xlarge >> Fast launch Memory optimized (no GPUs) 4 • ml.r5.large Full list can be found here: https://docs.aws.amazon.com/sagemaker/latest/dg/notebooks-available- instance-types.html Fast Launch: Optimized to launch in under 2 minutes
- 11. 11 Sagemaker Input Source: https://aws.amazon.com/blogs/machine-learning/using-pipe-input-mode-for-amazon-sagemaker-algorithms/ Data Sources • S3 is the first-class data source • File-mode: Dataset is downloaded to the training instance • Pipe mode: Dataset is streamed to your training instance • Gives higher throughput • Preferrable for large datasets • Only supported for Probobuf record-IO encoded data • FileSystems: EFS or Amazon FSx are other data sources
- 12. 12 Sagemaker Algorithms: JumpStart Source: https://docs.aws.amazon.com/sagemaker/latest/dg/notebooks.html Pre-trained Models and Solutions Templates • Pretrained models and frameworks are stored as docker images • Model API – Load pretrained model • Prepocessor API – Run preprocessing jobs • Transform API – Run Batch transform jobs • Predictor API – Supply class to model to make real-time predictions using sagemaker endpoints • Estimator API – Train or Finetune pre-trained Model Docker Images: https://sagemaker.readthedocs.io/en/stable/doc_utils/pretrainedmodels.html
- 13. 13 Sagemaker Algorithms: Amazon Estimators aka Amazon Algorithm Estimator Source: https://docs.aws.amazon.com/sagemaker/latest/dg/notebooks.html base class: sagemaker.amazon.amazon_estimator.AmazonAlgorithmEstimatorBase 1.Linear Learner 2.Latent Dirichlet Allocation (LDA) 3.Neural Topic Modelling (NTM) 4.Object2Vec 5.Principal Component Analysis (PCA) 6.RandomCutForest 7.Kmeans 8.Knearest Neighbors 9.Ip-Insights 10.Factorization machines
- 14. 14 Amazon Estimator Example: Linear Learner Source: https://sagemaker- examples.readthedocs.io/en/latest/scientific_details_of_algorithms/linear_learner_multiclass_classification/linear_learner_multiclass_classification.html Easy to use: Supply inputs then call Estimator.fit() Supported input formats: protobuf, csv, json Common information: https://docs.aws.amazon.com/sagemaker/latest/dg/common-info-all-im-models.html
- 15. 15 Sagemaker Algorithms: Frameworks Source: https://sagemaker.readthedocs.io/en/v2.38.0/frameworks/index.html base class: sagemaker.estimator.Framework 1.Scikit-learn 2.SparkMLServing 3.Tensorflow 4.Xgboost 5.Pytorch 6.HuggingFace 7.MXNet 8.Chainer 9.REinforcement Learning
- 16. 16 Frameworks Example: Xgboost Source: https://sagemaker.readthedocs.io/en/v2.38.0/frameworks/index.html Use as a Framework: custom logic is defined in script-mode
- 17. 17 Frameworks Example: Xgboost Source: https://sagemaker.readthedocs.io/en/v2.38.0/frameworks/index.html Use a builtin algorithm: retrieved from docker images Estimator used here is the base sagemaker estimator
- 18. 18 Best Resources Source: https://sagemaker.readthedocs.io/en/v2.38.0/frameworks/index.html 1. Sagemaker Examples Documentation: https://sagemaker- examples.readthedocs.io/en/latest/intro.html 2. Amazon SageMaker Python SDK: https://sagemaker.readthedocs.io/en/stable/
Editor's Notes
- #4: Some features are accessible directly for the AWS web console While others are accessible programmatically
- #5: Some features are accessible directly for the AWS web console While others are accessible programmatically
- #7: Switch to Deployed Jupyter Notebook
- #8: Advantages of Studio over Jupyter notebooks Easy to spin up and shut down training instances Cost savings because you can decide on the type of instance you want Collaboration between user profiles because it makes it easier to share notebooks than using Github
- #9: ****** Switch to Show Domain Creation ************ - You create a domain using AWS IAM identity or AWS IAM to authenticate - Afterwards, you create user profiles: this corresponds to a single user with a unique home directory in the EFS ****** Switch to Show Notebook Creation from docker images ***** Show notebook creation from docker images Show terminals: System terminal and Docker Image terminal Show JumpStart
- #12: Algorithms are already precoded Supply your data to these Algorithms
- #13: ***** Link to common properties of algorithms ******** Algorithms are already precoded Supply your data to these Algorithms
- #14: Algorithms are already precoded Supply your data to these Algorithms
- #20: Demo notes: sagemaker-python-sdk/scikit_learn_iris/