Upgrade to Pro
— share decks privately, control downloads, hide ads and more …
Speaker Deck
Features
Speaker Deck
PRO
Sign in
Sign up for free
Search
Search
Making Deployments Easy with TF Serving | TF Ev...
Search
Rishit Dagli
May 11, 2021
Programming
1
140
Making Deployments Easy with TF Serving | TF Everywhere India
My talk at TensorFlow Everywhere India
Rishit Dagli
May 11, 2021
Tweet
Share
More Decks by Rishit Dagli
See All by Rishit Dagli
Fantastic Models and Where to Find Them
rishitdagli
0
60
Plant AI: Project Showcase
rishitdagli
0
100
Deploying an ML Model as an API | Postman Student Summit
rishitdagli
0
78
APIs 101 with Postman
rishitdagli
0
65
Deploying Models to production with Azure ML | Scottish Summit
rishitdagli
1
73
Computer Vision with TensorFlow, Getting Started
rishitdagli
0
270
Teaching Your Models to Play Fair | Global AI Student Conf
rishitdagli
1
140
Deploying Models to Production with TF Serving
rishitdagli
1
180
Superpower Your Android apps with ML: Android 11 | Devfest 2020
rishitdagli
1
77
Other Decks in Programming
See All in Programming
のびしろを広げる巻き込まれ力:偶然を活かすキャリアの作り方/oso2024
takahashiikki
1
410
Jakarta Concurrencyによる並行処理プログラミングの始め方 (JJUG CCC 2024 Fall)
tnagao7
1
230
PHP でアセンブリ言語のように書く技術
memory1994
PRO
1
150
シールドクラスをはじめよう / Getting Started with Sealed Classes
mackey0225
3
400
Sidekiqで実現する 長時間非同期処理の中断と再開 / Pausing and Resuming Long-Running Asynchronous Jobs with Sidekiq
hypermkt
6
2.7k
Modern Angular: Renovation for Your Applications
manfredsteyer
PRO
0
210
Kubernetes for Data Engineers: Building Scalable, Reliable Data Pipelines
sucitw
1
200
Tuning GraphQL on Rails
pyama86
2
1k
Vue3の一歩踏み込んだパフォーマンスチューニング2024
hal_spidernight
3
3.1k
VR HMDとしてのVision Pro+ゲーム開発について
yasei_no_otoko
0
100
Identifying User Idenity
moro
6
7.9k
Universal Linksの実装方法と陥りがちな罠
kaitokudou
1
220
Featured
See All Featured
Speed Design
sergeychernyshev
24
570
RailsConf & Balkan Ruby 2019: The Past, Present, and Future of Rails at GitHub
eileencodes
131
33k
5 minutes of I Can Smell Your CMS
philhawksworth
202
19k
What's new in Ruby 2.0
geeforr
342
31k
個人開発の失敗を避けるイケてる考え方 / tips for indie hackers
panda_program
92
16k
Build The Right Thing And Hit Your Dates
maggiecrowley
32
2.4k
GitHub's CSS Performance
jonrohan
1030
460k
How GitHub (no longer) Works
holman
311
140k
The Pragmatic Product Professional
lauravandoore
31
6.3k
Producing Creativity
orderedlist
PRO
341
39k
Large-scale JavaScript Application Architecture
addyosmani
510
110k
Learning to Love Humans: Emotional Interface Design
aarron
272
40k
Transcript
Making Deployments Easy with TF Serving Rishit Dagli High School
TEDx, TED-Ed Speaker rishit_dagli Rishit-dagli
“Most models don’t get deployed.”
of models don’t get deployed. 90%
Source: Laurence Moroney
Source: Laurence Moroney
• High School Student • TEDx and Ted-Ed Speaker •
♡ Hackathons and competitions • ♡ Research • My coordinates - www.rishit.tech $whoami rishit_dagli Rishit-dagli
• Devs who have worked on Deep Learning Models (Keras)
• Devs looking for ways to put their model into production ready manner Ideal Audience
Why care about ML deployments? Source: memegenerator.net
None
• Package the model What things to take care of?
• Package the model • Post the model on Server
What things to take care of?
• Package the model • Post the model on Server
• Maintain the server What things to take care of?
• Package the model • Post the model on Server
• Maintain the server Auto-scale What things to take care of?
• Package the model • Post the model on Server
• Maintain the server Auto-scale What things to take care of?
• Package the model • Post the model on Server
• Maintain the server Auto-scale Global availability What things to take care of?
• Package the model • Post the model on Server
• Maintain the server Auto-scale Global availability Latency What things to take care of?
• Package the model • Post the model on Server
• Maintain the server • API What things to take care of?
• Package the model • Post the model on Server
• Maintain the server • API • Model Versioning What things to take care of?
Simple Deployments Why are they inefficient?
None
Simple Deployments Why are they inefficient? • No consistent API
• No model versioning • No mini-batching • Inefficient for large models Source: Hannes Hapke
TensorFlow Serving
TensorFlow Serving TensorFlow Data validation TensorFlow Transform TensorFlow Model Analysis
TensorFlow Serving TensorFlow Extended
• Part of TensorFlow Extended TensorFlow Serving
• Part of TensorFlow Extended • Used Internally at Google
TensorFlow Serving
• Part of TensorFlow Extended • Used Internally at Google
• Makes deployment a lot easier TensorFlow Serving
The Process
• The SavedModel format • Graph definitions as protocol buffer
Export Model
SavedModel Directory
auxiliary files e.g. vocabularies SavedModel Directory
auxiliary files e.g. vocabularies SavedModel Directory Variables
auxiliary files e.g. vocabularies SavedModel Directory Variables Graph definitions
TensorFlow Serving
TensorFlow Serving
TensorFlow Serving Also supports gRPC
TensorFlow Serving
TensorFlow Serving
TensorFlow Serving
TensorFlow Serving
Inference
• Consistent APIs • Supports simultaneously gRPC: 8500 REST: 8501
• No lists but lists of lists Inference
• No lists but lists of lists Inference
• JSON response • Can specify a particular version Inference
with REST Default URL http://{HOST}:8501/v1/ models/test Model Version http://{HOST}:8501/v1/ models/test/versions/ {MODEL_VERSION}: predict
• JSON response • Can specify a particular version Inference
with REST Default URL http://{HOST}:8501/v1/ models/test Model Version http://{HOST}:8501/v1/ models/test/versions/ {MODEL_VERSION}: predict Port Model name
Inference with REST
• Better connections • Data converted to protocol buffer •
Request types have designated type • Payload converted to base64 • Use gRPC stubs Inference with gRPC
Model Meta Information
• You have an API to get meta info •
Useful for model tracking in telementry systems • Provides model input/ outputs, signatures Model Meta Information
Model Meta Information http://{HOST}:8501/ v1/models/{MODEL_NAME} /versions/{MODEL_VERSION} /metadata
Batch Inferences
• Use hardware efficiently • Save costs and compute resources
• Take multiple requests process them together • Super cool😎 for large models Batch inferences
• max_batch_size • batch_timeout_micros • num_batch_threads • max_enqueued_batches • file_system_poll_wait
_seconds • tensorflow_session _paralellism • tensorflow_intra_op _parallelism Batch Inference Highly customizable
• Load configuration file on startup • Change parameters according
to use cases Batch Inference
Also take a look at...
• Kubeflow deployments • Data pre-processing on server🚅 • AI
Platform Predictions • Deployment on edge devices • Federated learning Also take a look at...
bit.ly/tf-everywhere-ind Demos!
bit.ly/serving-deck Slides
Thank You rishit_dagli Rishit-dagli