PyTorch vs TensorFlow | Ishan Misra and Lex Fridman

Lex Fridman Podcast full episode: • Ishan Misra: Self-Supe...
Please support this podcast by checking out our sponsors:
- Onnit: lexfridman.com/onnit to get up to 10% off
- The Information: theinformation.com/lex to get 75% off first month
- Grammarly: grammarly.com/lex to get 20% off premium
- Athletic Greens: athleticgreens.com/lex and use code LEX to get 1 month of fish oil
GUEST BIO:
Ishan Misra is a research scientist at FAIR working on self-supervised visual learning.
PODCAST INFO:
Podcast website: lexfridman.com/podcast
Apple Podcasts: apple.co/2lwqZIr
Spotify: spoti.fi/2nEwCF8
RSS: lexfridman.com/feed/podcast/
Full episodes playlist: • Lex Fridman Podcast
Clips playlist: • Lex Fridman Podcast Clips
SOCIAL:
- Twitter: / lexfridman
- LinkedIn: / lexfridman
- Facebook: / lexfridman
- Instagram: / lexfridman
- Medium: / lexfridman
- Reddit: / lexfridman
- Support on Patreon: / lexfridman

[@ogrendang7810]

Asking a Facebook employee about pytorch vs tensorflow. 🤔

[@ghostman3329]

Yup ..

[@kidzeus2433]

Lol

[@niranjan_7891]

They make the largest no of open-source AI models so why not.

[@niranjan_7891]

They have currently 1907models.Refer Hugging Face

[@BapiKAR]

PyTorch is what I made a transition to from Tf2.x(x>=7). So far I see it quite cool & friendly to research community. And yes, I converted my entire TF code to PyTorch in less than 2weeks.

[@Fordance100]

TF usually has some features 1-2 year ahead pytorch. Pytorch is more flexible for tweaking the model. TF2.4-2.6 was very buggy with strange errors that took long time to fix. That was the time I switched to mainly pytorch. I think TF is better now.

[@adityanjsg99]

Tensorflow...! Great support, model integration and model deployment.

[@qedqubit]

wow 😀 ! you are diverse in your topics ! luv 'it 😘!

[@GChief117]

PyTorch has more applications over tensorflow

[@os2002n77]

I like calling pytorch imperative! Because it you would understand machine learning in better way! Unlike declarative approaches similar to Unix like command and SQL you give it a command or query and it would run. Even through that Tensorflow uses procedural language, coding with it similar to declaring command. And static graph structuring making it difficult especially for beginners debugging code!

[@perceptron5983]

Both are great. TF with Keras has better performance, strong community support and robustness 👍PyTorch better for research and experimentation. Easier to use and debug since is more pythonic. Better Dynamic computation graph. Easier to deploy on web and mobile. Your choice will be determine on: use case and developer's preferences.

[@GChief117]

I mean there are more applications with PyTorch for computer vision purposes over tensorflow

[@carlossegura403]

PyTorch’s source code is elegant and well thought out.

[@noli-timere-crede-tantum]

In other words, TF's source code is complete garbage with no forethought taken? Jeff Dean just sat down one day and started typing until a framework emerged. Got it...

[@cristian-bull]

@@noli-timere-crede-tantum Pytorch's source code is more elegant and better thought out.

[@noli-timere-crede-tantum]

@@cristian-bull I'm sure you can give examples of that, as well as examples of when TF's source code lesser elegance makes it less effective at run time or solving a bug or adding features?

[@generichuman_]

@@noli-timere-crede-tantum When someone tells you they like pepperoni pizza, do you say "oh... so what you're saying is that ham and pineapple pizza is garbage?!". No one is saying what you are accusing them of saying. Here's a better, more productive pathway. Why don't you tell us why you think tensorflow is better, instead of asking everyone else to do that for you.

[@manny7662]

@@generichuman_ would PyTorch better for startups?

[@stephennfernandes]

Pytorch Forever 🔥🔥🔥 ... I know the XLA TPU support isn't almost there yet but yeah Pytorch 🔥

[@anmolmonga1933]

Using TPU very niche thought

[@Ikaikaalika95]

pytorch is more intuitive for SWEs, tensorflow is killer now with the keras integration

[@ravipratapmishra7013]

For me pytorch can give you more granular control with small learning curve, but when it comes to deplotment and documentation tf is way ahead.

[@MrEmbrance]

jax

[@abrahamsong6913]

RIP tensorflow

[@ritapravadutta7939]

After keras integration tf is winner

[@user-gp3ub5td2t]

#throwbacksundays

[@BobF510]

This content is incredibly moving. I read a book with akin material that reshaped my worldview. "AWS Unleashed: Mastering Amazon Web Services for Software Engineers" by Harrison Quill

[@MrGeorgetudo]

WHY THE FUCK IS LEX FRIDMAN TALKING ABOUT MACHINE LEARNING

[@BobbyDaBunnyRabbit]

Tensorflow is bread and butter

[@generichuman_]

so it's boring?

[@gorgolyt]

Pretty useless clip tbh, all he says is "I prefer Pytorch because I've been using it for longer". He mentions that the imperative style is easier to debug, but TensorFlow 2 also uses an imperative style.

[@tashinahmed5745]

doesn't answer the question at all :|

[@noli-timere-crede-tantum]

I think his initial disclaimer completely disqualifies him from answering the question. If he hasn't built something meaningful in TF2, how can he make any comparisons whatsoever?

[@chillydoog]

read between the lines

[@generichuman_]

@@noli-timere-crede-tantum Have you used Pytorch?

[@francismumbi49]

Interesting

[@jrich5682]

Indeed

[@chillydoog]

Rather

[@davisonyeoguzoro9232]

Bias Try asking a Google employee the same question

[@totheknee]

More like collaboration than competition. The open source _COMMUN_-ity shows how a _commune_ environment that _shares_ resources can be so vastly superior to the capitalistic wealth redistribution scheme, it isn't even funny.

[@paveldanilov4869]

trrrrrrrrrrrrr