A friendly introduction to distributed training (ML Tech Talks)

preview_player
Показать описание
Google Cloud Developer Advocate Nikita Namjoshi introduces how distributed training models can dramatically reduce machine learning training times, explains how to make use of multiple GPUs with Data Parallelism vs Model Parallelism, and explores Synchronous vs Asynchronous Data Parallelism.

Chapters:
0:00 - Introduction
00:17 - Agenda
00:37 - Why distributed training?
1:49 - Data Parallelism vs Model Parallelism
6:05 - Synchronous Data Parallelism
18:20 - Asynchronous Data Parallelism
23:41 Thank you for watching

#TensorFlow #MachineLearning #ML

product: TensorFlow - General;
Рекомендации по теме
Комментарии
Автор

Wow Ring All-Reduce is just... beautiful. 😍

giovannimurru
Автор

this is really well explained. more on this series please. thanks

wryltxw
Автор

awesome video, crystal clear with the content design and easy to understand

kkmm
Автор

Thank you very much for the insightful presentation.

lukasvandewiel
Автор

Good an easy to understand explanations!

extrememike
Автор

GPU must be same, what happens if i use different GPU???

ahmadnoroozi
Автор

What if i dont have gpus as you said in the video, i have 32 systems with i5 CPU..can i run this mirrored strategy on multiple CPUs?

Ajaytshaju
Автор

At time code 15:23 in the Ring Reduce algorithm the subscripts for the c vector are incorrect.

VanWarren
Автор

Hello is there any docs about federated learning with differentiel privacy, thank you

hannibalbra
Автор

That tf.distribute.experimental worries me.. not sure when the api will be deprecated.

liuauto