Asynchronous Hyperparameter Tuning and Ablation Studies with Apache - - PowerPoint PPT Presentation

▶
asynchronous hyperparameter tuning and ablation studies
SMART_READER_LITE
LIVE PREVIEW

Asynchronous Hyperparameter Tuning and Ablation Studies with Apache - - PowerPoint PPT Presentation

Nov 11, 2022 445 likes •892 views

Asynchronous Hyperparameter Tuning and Ablation Studies with Apache Spark Sina Sheikholeslami Distributed Computing Group, KTH Royal Institute of Technology @cutlash CASTOR Software Days 2019 October 16 2019 sinash@kth.se The Machine

slide-1
SLIDE 1

Asynchronous Hyperparameter Tuning and Ablation Studies with Apache Spark

sinash@kth.se

Sina Sheikholeslami

Distributed Computing Group, KTH Royal Institute of Technology

October 16 2019 CASTOR Software Days 2019

@cutlash

slide-2
SLIDE 2

The Machine Learning System

October 16 2019 Sina Sheikholeslami - KTH Royal Institute of Technology 2

Dataset Machine Learning Model Optimizer Evaluate Problem Definition Data Preparation Model Selection

Repeat if needed Model Training

slide-3
SLIDE 3

Artificial Neural Networks

October 16 2019 Sina Sheikholeslami - KTH Royal Institute of Technology 3

Input Layer Hidden Layer Output Layer

slide-4
SLIDE 4

How We Study the Brain

  • Early 19th Century,

ablative brain surgeries by Jean Pierre Flourens

(1794 - 1867)

October 16 2019 Sina Sheikholeslami - KTH Royal Institute of Technology 4

slide-5
SLIDE 5

Ablation for Machine Learning?

October 16 2019 Sina Sheikholeslami - KTH Royal Institute of Technology 5

Dataset Machine Learning Model Optimizer Evaluate Problem Definition Data Preparation Model Selection

Repeat if needed Model Training

area rooms floors price

slide-6
SLIDE 6

Talk of the Town

October 16 2019 Sina Sheikholeslami - KTH Royal Institute of Technology 6

“Too frequently, authors propose many tweaks absent proper ablation studies … Sometimes just one of the changes is actually responsible for the improved results … this practice misleads readers to believe that all of the proposed changes are necessary.”

(Lipton & Steinhardt, “Troubling Trends in Machine Learning Scholarship”)

slide-7
SLIDE 7

October 16 2019 Sina Sheikholeslami - KTH Royal Institute of Technology 7

Accuracy: 78%

Example: Layer Ablation (1/6)

The Base Model

slide-8
SLIDE 8

October 16 2019 Sina Sheikholeslami - KTH Royal Institute of Technology 8

Accuracy: 73%

Example: Layer Ablation (2/6)

slide-9
SLIDE 9

October 16 2019 Sina Sheikholeslami - KTH Royal Institute of Technology 9

Example: Layer Ablation (3/6)

The Base Model

slide-10
SLIDE 10

October 16 2019 Sina Sheikholeslami - KTH Royal Institute of Technology 10

Accuracy: 67%

Example: Layer Ablation (4/6)

slide-11
SLIDE 11

October 16 2019 Sina Sheikholeslami - KTH Royal Institute of Technology 11

Example: Layer Ablation (5/6)

The Base Model

slide-12
SLIDE 12

October 16 2019 Sina Sheikholeslami - KTH Royal Institute of Technology 12

Accuracy: 63%

Example: Layer Ablation (6/6)

slide-13
SLIDE 13

Ablation Study

October 16 2019 Sina Sheikholeslami - KTH Royal Institute of Technology 13

Machine Learning System

Ablation

New Dataset / Model Configuration Evaluate

slide-14
SLIDE 14

Hyperparameter Tuning

October 16 2019 Sina Sheikholeslami - KTH Royal Institute of Technology 14

Machine Learning System

Hyperparameter Tuner

New Hyperparameter Values Evaluate

slide-15
SLIDE 15

System Experimentation (Search)

October 16 2019 Sina Sheikholeslami - KTH Royal Institute of Technology 15

Machine Learning System

Global Experiment Controller

New Trial Evaluate

slide-16
SLIDE 16

Better Parallel

  • Ability to train better models, faster
  • Ability to modify and inspect, easier

October 16 2019 Sina Sheikholeslami - KTH Royal Institute of Technology 18

(“Parallel Training” - by Maxim Melnikov)

slide-17
SLIDE 17

Parallelization in Practice

October 16 2019 Sina Sheikholeslami - KTH Royal Institute of Technology 19

(TensorFlow, the TensorFlow logo and any related marks are trademarks of Google Inc.)

Machine Learning Deep Learning Parallel Processing

slide-18
SLIDE 18

Hopsworks

October 16 2019 Sina Sheikholeslami - KTH Royal Institute of Technology 20

Open-source Platform for Data-intensive AI

slide-19
SLIDE 19

Hopsworks

October 16 2019 Sina Sheikholeslami - KTH Royal Institute of Technology 21

Open-source Platform for Data-intensive AI

What is Hopsworks? https://tinyurl.com/y4ze79d4

slide-20
SLIDE 20

ML/DL in Hopsworks

October 16 2019 Sina Sheikholeslami - KTH Royal Institute of Technology 22

Data Pipelines Ingest & Prep Feature Store Machine Learning Experiments Data Parallel Training Model Serving

Ablation Studies Hyperparameter Tuning

Bottleneck, due to

  • iterative nature
  • human interaction
slide-21
SLIDE 21

Spark and Bulk Synchronous Parallel Model

October 16 2019 Sina Sheikholeslami - KTH Royal Institute of Technology 23

Task11

Driver

Task12 Task13 Task1N

…

HDFS

Task21 Task22 Task23 Task2N

…

Barrier Barrier

Task31 Task32 Task33 Task3N

…

Barrier

Metrics1 Metrics2 Metrics3

slide-22
SLIDE 22

Example: Synchronous Hyperparameter Search

October 16 2019 Sina Sheikholeslami - KTH Royal Institute of Technology 24

Task11

Driver

Task12 Task13 Task1N

…

HDFS

Task21 Task22 Task23 Task2N

…

Barrier Barrier

Task31 Task32 Task33 Task3N

…

Barrier

Metrics1 Metrics2 Metrics3 Wasted Compute Wasted Compute Wasted Compute

slide-23
SLIDE 23

Critical Requirements

  • Parallel execution of trials
  • Support for early stopping of trials
  • Support for global control of the experiment
  • Resilience to stragglers
  • Simple, “Unified” User & Developer API

October 16 2019 Sina Sheikholeslami - KTH Royal Institute of Technology 25

slide-24
SLIDE 24

Maggy

An Open-source Framework for Asynchronous Computation on top of Apache Spark

October 16 2019 Sina Sheikholeslami - KTH Royal Institute of Technology 26

slide-25
SLIDE 25

Key Idea: Long Running Tasks

October 16 2019 Sina Sheikholeslami - KTH Royal Institute of Technology 27 Task11

Driver

Task12 Task13 Task1N

…

Barrier

Metrics New Trial

slide-26
SLIDE 26

Maggy Core Architecture

October 16 2019 Sina Sheikholeslami - KTH Royal Institute of Technology 28

slide-27
SLIDE 27

Back to Ablation

October 16 2019 Sina Sheikholeslami - KTH Royal Institute of Technology 29

slide-28
SLIDE 28

LOCO: Leave One Component Out

  • A simple, “natural” ablation policy: an implementation of an ablator
  • Currently supports Feature Ablation + Layer Ablation

October 16 2019 Sina Sheikholeslami - KTH Royal Institute of Technology 30

slide-29
SLIDE 29

Feature Ablation

  • Uses the Feature Store to access the dataset metadata
  • Generates Python callables that once called, will return modified datasets
  • Removes one-feature-at-a-time

October 16 2019 Sina Sheikholeslami - KTH Royal Institute of Technology 31

area rooms floors price rooms floors price

slide-30
SLIDE 30

Layer Ablation

  • Uses a base model function
  • Generates Python callables that once called, will return modified models
  • Uses the model configuration to find and remove layer(s)
  • Removes one-layer-at-a-time (or one-layer-group-at-a-time)

October 16 2019 Sina Sheikholeslami - KTH Royal Institute of Technology 32

slide-31
SLIDE 31

Ablation User & Developer API

October 16 2019 Sina Sheikholeslami - KTH Royal Institute of Technology 33

(Example Notebook Available!)

slide-32
SLIDE 32

User API: Initialize the Study and Add Features

October 16 2019 Sina Sheikholeslami - KTH Royal Institute of Technology 34

slide-33
SLIDE 33

User API: Define Base Model

October 16 2019 Sina Sheikholeslami - KTH Royal Institute of Technology 35

slide-34
SLIDE 34

User API: Setup Model Ablation

October 16 2019 Sina Sheikholeslami - KTH Royal Institute of Technology 36

slide-35
SLIDE 35

User API: Wrap the Training Function

October 16 2019 Sina Sheikholeslami - KTH Royal Institute of Technology 37

slide-36
SLIDE 36

User API: Lagom!

October 16 2019 Sina Sheikholeslami - KTH Royal Institute of Technology 38

slide-37
SLIDE 37

Developer API: Policy Implementation (1/2)

October 16 2019 Sina Sheikholeslami - KTH Royal Institute of Technology 39

slide-38
SLIDE 38

Developer API: Policy Implementation (2/2)

October 16 2019 Sina Sheikholeslami - KTH Royal Institute of Technology 40

slide-39
SLIDE 39

Hyperparameter Tuning: User API

October 16 2019 Sina Sheikholeslami - KTH Royal Institute of Technology 41

slide-40
SLIDE 40

Hyperparameter Tuning: Developer API

October 16 2019 Sina Sheikholeslami - KTH Royal Institute of Technology 42

slide-41
SLIDE 41

Maggy is Open-source

  • Code Repository: https://github.com/logicalclocks/maggy
  • API Documentation: https://maggy.readthedocs.io/en/latest/

October 16 2019 Sina Sheikholeslami - KTH Royal Institute of Technology 43

slide-42
SLIDE 42

Next Steps

  • More Ablators
  • More Tuners
  • Support for More Frameworks

October 16 2019 Sina Sheikholeslami - KTH Royal Institute of Technology 44

slide-43
SLIDE 43

Thank you! J

@logicalclocks @hopsworks GitHub

https://github.com/hopshadoop/maggy https://maggy.readthedocs.io/en/latest/ https://logicalclocks.com/whitepapers/ Thanks to the entire Logical Clocks Team J Specially: Moritz Meister @morimeister Jim Dowling @jim_dowling Robin Andersson @robzor92 Kim Hammar @KimHammar1 Alex Ormenisan @alex_ormenisan

(Example Notebook Available!)

sinash@kth.se October 16 2019 CASTOR Software Days 2019

@cutlash