Lucas Beyer clarified BiternionNet was developed at RWTH Aachen
if you're doing AI research at all; I recommend doing the "ETH zurich" route
Train models that use a single GPU. Make sure that it takes less than a minute to train models. Pufferlib is a great example.
The more models you train the more you learn
BiternionNet、１分で学習が終わってしまったんだが。
Many users endorse fast single-GPU AI model training because it enables rapid iterations and more effective learning without needing large compute clusters.
No Digg Deeper questions have been answered for this story yet.
💯AI research often isn't so much research, as it is "throwing stuff at the wall and seeing what sticks". If you can throw more stuff, you will learn more, and also more stuff will stick. @kaggle was a great place for me to internalise this, but it is also true in the real world!
if you're doing AI research at all; I recommend doing the "ETH zurich" route
Train models that use a single GPU. Make sure that it takes less than a minute to train models. Pufferlib is a great example.
The more models you train the more you learn
@yacineMTB These two posts (Yacine and PINTO) motivated me to write up more info about this project, if anyone is curious:
Alright, it's time for a paper thread about my own first ever vision paper, which is having a bit of a moment on twitter rn thanks to @PINTO03091 and @yacineMTB.
BiternionNets: continuous head orientation from discrete labels.
Demo video from ~11y ago:
@yacineMTB Btw this is my model/code i wrote while i was at RWTH Aachen, Long before my Zürich time :) (And I've never been student at ETH)
if you're doing AI research at all; I recommend doing the "ETH zurich" route
Train models that use a single GPU. Make sure that it takes less than a minute to train models. Pufferlib is a great example.
The more models you train the more you learn
@guojing0 i'm talking about their research. it's often really great results with a single gpu
**Sure!** The "ETH Zurich route" points to their fast-iteration RL style using massive parallelism on a **single workstation GPU** for quick training (minutes instead of hours/days).
Key source: "Learning to Walk in Minutes Using Massively Parallel Deep Reinforcement Learning" (ETH Zurich + NVIDIA, 2021) → https://arxiv.org/abs/2109.11978
PufferLib is an excellent modern example built for exactly this: tiny models that train in seconds to minutes on a single GPU (or even CPU), letting you run way more experiments.
GitHub: https://github.com/pufferai/pufferlib Site: https://puffer.ai/
Fast feedback loops > big clusters for learning.
@yacineMTB Pufferlib is the gold standard for this. It’s a masterclass in proving that you don't need a massive compute cluster if your underlying engine is actually written efficiently.
@yacineMTB what is eth Zurich route?
@thkostolansky yeah. i learned a lot about RNNs
@OGALANGLEY @yacineMTB Check out https://github.com/Entrpi/eemicrogpt you don't even need a GPU.
@yacineMTB unconditionally endorsed
@yacineMTB @JoelDeTeves @grok gove me some sources for this
@yacineMTB Cuda, metal or whatever?
@yacineMTB I see, thank you for the clarifications. I forgot if it’s you or someone else, also mentioned in the past that EPFL does really good ML research (without relying on too much compute).
Out of curiosity, which GPU(s) would you recommend, how about 5060 Ti 16 GB?
@yacineMTB Why ETH? They teach courses this way?
@yacineMTB Hands-on iteration beats reading papers every time.
@WageRoti @yacineMTB Public university in Switzerland that provides cloud compute to scientists and researchers
@giffmana Wow..
@yacineMTB did u do this
@yacineMTB Single GPU, fast convergence. I prototype the same way. Optimizing information density per parameter before you scale is how you actually learn what matters.
@yacineMTB What kind of tiny models are you training that take less than a minute? I have a 6000 pro at home and never reached that on a full run