Cloud Practices for
Machine Learning

@LynnLangit

August 2020

Topics

  • 0%       Data & Questions 
  • 20%    Dev & Cloud
  • 40%    Start with Hello (ML)
  • 60%    Data Samples & Security
  • 80%    Model Quality
  • 100% Scale & Reproduce

Solve for What?

# A hybrid quantum-classical model.
model = tf.keras.Sequential([
    # Quantum circuit data comes in inside of tensors.
    tf.keras.Input(shape=(), dtype=tf.dtypes.string),

    # Parametrized Quantum Circuit (PQC) provides output
    # data from the input circuits run on a quantum computer.
    tfq.layers.PQC(my_circuit, [cirq.Z(q1), cirq.X(q0)]),

    # Output data from quantum computer passed through model.
    tf.keras.layers.Dense(50)
])

Data

  • What - type(s)
  • Where - source locations
  • When - frequency
  • How much -> CLOUD-scale

Get Sample Data

  • File Type
  • Format
  • Quality
  • Quantity

Get Cloud Open Data

Dev Environment

The cloud is your laptop, really.

Dev Environment

  • Set up a Virtual Machine Image
  • Use a Jupyter Notebook
  • Use a ML Docker Container Image

Which Cloud?

Vendor/Location

Get Started

  • Vendor
  • Account Type
  • Data Center

 

Do the Basics

  • Account
  • Billing
  • Security

Hello
ML Cloud
World

Types of Cloud Machine Learning 

  • Virtual Machines w/your model
  • Docker containers w/your (vendor) model
  • Vendor APIs

Build vs Buy - AWS

Build vs Buy - GCP

Learn the Cloud ML Services

  • SaaS
  • PaaS
  • IaaS

Hello ML Algorithm w/mnist

  • Works?
  • Verified?
  • Time/Cost?
  • Reuse?
  • Bias?

Your
Data 
Matters

Bias in the Training Data

Model

Quality
Matters

MyCloudML - Example

  • Why the cloud? - data volume
  • How long will it take? cost?
  • What to do first? - ????

Hello MyML

  • Works?
  • Verified?
  • Time/Cost?
  • Reuse?
  • Bias?

Tools for Fairness - TensorFlow

Cost-effective Scaling

Scaling / Reproducibility

  • Works?
  • Verified?
  • Time/Cost?
  • Reuse?
  • Bias?

One-page Diagram

MLDevOps - AWS

 

  • Config as Code
    • Automated Tests
    • Deployment Templates
    • CI/CD

Full MLDevOps - AWS

MLDevOps by Example

Domain-specific cloud ML - Bioinformatics

Cloud Practices for
Machine Learning

@LynnLangit

August 2020

Cloud for ​Machine Learning

By Lynn Langit

Cloud for ​Machine Learning

Resources and patterns for implementing effective machine learning workloads on the public cloud.

  • 68
Loading comments...

More from Lynn Langit