[go: up one dir, main page]
Scribd
Upload
133 views38 pages

Deep Learning With Databricks: Srijith Rajamohan, Ph.D. John O'Dwyer

This document discusses how Databricks can be used for deep learning and machine learning workflows. It covers topics like model building, distributed training, model tracking with MLflow, automating machine learning with AutoML, and deploying models with MLflow Projects and model serving. The integrated Databricks platform provides notebooks, compute clusters, job scheduling, experiments tracking, and other tools to support the full machine learning lifecycle from data preparation to model deployment.

Uploaded by

NiharikaNic
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
133 views38 pages

Deep Learning With Databricks: Srijith Rajamohan, Ph.D. John O'Dwyer

This document discusses how Databricks can be used for deep learning and machine learning workflows. It covers topics like model building, distributed training, model tracking with MLflow, automating machine learning with AutoML, and deploying models with MLflow Projects and model serving. The integrated Databricks platform provides notebooks, compute clusters, job scheduling, experiments tracking, and other tools to support the full machine learning lifecycle from data preparation to model deployment.

Uploaded by

NiharikaNic
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 38

Deep Learning with

Databricks
Srijith Rajamohan, Ph.D.
John O’Dwyer
Open

30+ Million
Unify your data
ecosystem with open
source, standards and
formats
monthly downloads

Built on the innovation of


some of the most
successful open source
data projects in the world
Data Analysts Data Scientists Collaborative
Unify your data teams
to collaborate across the
Models entire data and
AI workflow
Dashboards

Notebooks

Datasets

Data Engineers
Questions for Scalable ML
▪ Track the provenance and reason for model creation
▪ What training data was used, if any?
▪ Proprietary data, sensitive data, storage, data retention period?
▪ Real-time or batch?
▪ How are the models being used and who is using it?
▪ Exploratory analysis and production environment?
▪ Is model performance being measured regularly and is the model being updated?
▪ Is the model well documented to ensure reuse?
▪ Is the model deployment process being automated?
▪ Institutional adoption and support
Best Practices for ML
▪ Software engineering practices
▪ Code quality best practices
▪ Validate your data
▪ Ensure proper data types and format are fed to your model (Schema validation)
▪ Ensure no data drift, can render a supervised model ineffective
▪ Version and track your experiments like code!
▪ Changing hyperparameters, inputs, code etc.
▪ Monitor predictive performance over time
▪ Ensure model performance does not degrade over time
▪ Ensure model fairness across different classes of data (bias)
What is MLOps?

MLOps = ML + DataOps + DevOps

Build -> Test -> Deploy -> Monitor -> Feedback -> Build

Model management
Databricks Ecosystem for ML/DL
▪ Integrated Environment
▪ Use compute instances from AWS, Azure or GCP
▪ Centered around a notebook environment
▪ Version control them with GitHub
▪ Integrated ‘DBFS’ filesystem that can mount cloud filesystems like S3
▪ Mix SQL, Python, R and Bash in the same notebook
▪ Schedule jobs to run anytime
▪ Databricks Runtimes (DBRs)
▪ Preinstalled with packages for ML/DL
▪ Additional packages can be installed per cluster or per notebook
▪ MLflow integrated into the Databricks platform
▪ Model tracking for experiment management/reproducibility
▪ MLflow projects for packaging an experiment
▪ Model serving with MLflow
Workspace
Workspace
Notebooks
Job scheduling
Job page
Experiments
Registered models
The Data Preparation
The Delta Lake Architecture
Data Store and Versioning
Delta Lake Feature Store
▪ Scalable metadata ▪ Data stored needs to be transformed
▪ Time travel into features to be useful
▪ Feature tables are Delta tables
▪ Open format ▪ Feature Stores can save these features
▪ Unified Batch and Streaming ▪ Discoverable and reusable across an
▪ Schema enforcement organization
▪ Ensures consistency for Data Engineers,
Data Scientists and ML Engineers
▪ Track feature lineage in a model
ETL and EDA
▪ Delta lake
▪ Save data in scalable file formats like Parquet
▪ Delta file formats can let you version control your data
▪ ETL
▪ Read data
▪ PySpark - Ideal for large data
▪ Tensorflow (tf.data) and Pytorch (DataLoader)
▪ Clean and process data
▪ PySpark/Pandas API on Spark can work with large datasets across clusters
▪ Clean and prepare the data
▪ Extract features and save them using Feature Stores
▪ EDA
▪ Preliminary data analysis such as inspecting records, summary statistics
▪ Visualize the data and its distribution
The Model Build
Model training
▪ DBRs provide your favorite DL frameworks such as Tensorflow, Pytorch,
Keras etc.
▪ Integration with MLflow for model tracking
▪ Hyperparameter tuning with Hyperopt/Optuna
▪ Seamlessly run single node but multi-CPU/multi-GPU jobs
▪ Distributed training on multiple nodes with Horovod
▪ NVlink/NCCL enabled instances available for accelerating DL workloads
▪ Tightly coupled - Train directly on Spark Dataframes with Horovod Estimator
▪ Train on distributed Spark clusters with Horovod Runner
Distributed Training with Spark/Horovod
Distributed Training with Spark/Horovod contd...
Invoke training across multiple nodes

Inference using Horovod


Distributed Training
Data parallelism Model parallelism
▪ Data is divided among the different ▪ Model is divided among all the nodes
nodes ▪ Only works if you can take advantage of
▪ Entire model is copied to all the nodes task parallelism in the model
▪ Gradients are communicated back to ▪ Model size is less of a concern
all other nodes to update the model
▪ Synchronous or asynchronous updates
▪ Model size is a concern
Deep Learning Synchronization
Model parameter server All-reduce
▪ Central servers hold all shared ▪ All the machines store the shared
parameters parameters
▪ Workers receive updates from the ▪ No central server
central server ▪ Several architectures for this
▪ Harder to scale ▪ Ring All-reduce
▪ Speedup now depends on the overhead ▪ Tree All-reduce
of communication with the central
server
Other Topics in Training

▪ Quantization-aware training
▪ Lower-precision training to minimize memory/compute requirements
▪ Federated learning
▪ Decentralized learning with the Federated Averaging algorithm (Google)
▪ Keep data on device
▪ Model is updated with data on device and updates sent back to central server
▪ Updates from all devices are averaged
▪ Privacy-preserving learning
▪ Learn from data that is encrypted or with minimal exposure to the data
Model tracking with MLflow

▪ The MLflow Tracking API


▪ Integrations with common ML/DL tools such as Scikit-learn, Pytorch,
Tensorflow, Spark etc.
▪ Logs metrics and artifacts (output files)
▪ Can log this locally or a remote tracking server
▪ Tracking UI to query runs and visualize the results of a run
▪ Save and load models from a run
Model tracking with MLflow - Keras
Model tracking with MLflow - Autolog

With many of the popular libraries,


you can use the autologging feature
AutoML

▪ Only ML algorithms for now


▪ Works with 9.1 LTS ML DBRs and above
▪ Classification and Regression
▪ Decision trees, Random Forests, Logistic Regression, XGBoost, LightGBM
▪ Forecasting with Prophet
▪ Run from the UI or use the command line API
AutoML
AutoML contd...
AutoML - Load the best model
AutoML - Experiments
The Model Inference and Deployment
Model Inference - Pandas UDF
▪ Use a compiled DL model with Pandas UDF for distributed inference
▪ Scalar pandas UDF (batch of data) vs. Iterator pandas UDF (iterator of
batches ) here so model is no initialized for every batch
Model Packaging with MLflow Projects

MLProject file for


reproducible executions

File under folder


sklearn_elasticnet_wine

Execute this project using


the command below

mlflow run sklearn_elasticnet_wine -P alpha=0.42


Model Serve with MLflow
Serve the model
mlflow models serve -m
/Users/mlflow/mlflow-prototype/mlruns/0/7c1a0d5c42844dcdb8f5191146925
174/artifacts/model -p 1234

Send a request
curl -X POST -H "Content-Type:application/json; format=pandas-split"
--data '{"columns":["alcohol", "chlorides", "citric acid",
],"data":[[12.8, 0.029, 0.48]]}' http://127.0.0.1:1234/invocations
Thank you!

You might also like