Similar User
@punk1685
@jaxayshah
@Miisssliz
@Prophet87_eth
@1903SavasBJK
@maesocio
@drah_cula
@tdl8_
@LimaJmr8
@FesslerThedfess
A comparison of the exact same model training for all 3 approaches 1) Pure PyTorch: github.com/rasbt/pytorch-… (25 min) 2) PyTorch + Fabric: github.com/rasbt/pytorch-… (1.7 min) 3) PyTorch + Trainer: github.com/rasbt/faster-p… (2.7 min, the 1 extra min is for logging+checkpointing)
Easy to miss in the @pytorch 2.0 release notes, they've added a small, but useful feature: torch.device, which previously just returned a device object, can now be used as a context manager. 0/8
What an awesome week for open source and the PyTorch ecosystem with three big launches! - PyTorch 2.0 - Lightning Trainer 2.0 for PyTorch - Fabric for PyTorch! Just updated my "faster PyTorch" article to include the latest tools! 🔗 sebastianraschka.com/blog/2023/pyto…
Btw if you have an hour to spare on the weekend, I am covering how to use the Trainer to fit PyTorch models in the newly released Unit 5: lightning.ai/pages/courses/…
What an awesome week for open source and the PyTorch ecosystem with three big launches! - PyTorch 2.0 - Lightning Trainer 2.0 for PyTorch - Fabric for PyTorch! Just updated my "faster PyTorch" article to include the latest tools! 🔗 sebastianraschka.com/blog/2023/pyto…
But note that vision transformers (ViTs) are not free from any inductive biases! ViTs focus more on global relationships due to the patchification & self-attention mechanism, which often leads to the perception that they act as low-pass filters, emphasizing (or recognizing)…
Similar to fully-connected networks, the ViT architecture (and transformer architecture in general) lacks the inductive bias for spatial invariance/equivariance that convolutional networks have. Consequently, ViTs require more data for pretraining to acquire useful "priors" from…
Plot twist: they didn’t disclose any details because there was actually no innovation to report, just a bit more finetuning, which would have looked underwhelming given all the hype. By not sharing any details they made GPT-4 seem like a bigger innovation/deal than it really is.
GPT-4 was interesting for a hot second, I'll admit. But today is a new day: time to move on and get back to discussing research and open source.
Exactly! This is a great example of how CVPR publicity restrictions very effectively prevent unfair public visibility on social media for research from prestigious institutes like Ivy League. oh wait…
For a hot second, I was wondering how relevant weight decay still is. Instead of asking ChatGPT, I ran a simple experiment (*when a picture says more than a thousand words*)
Some unsolicited writing advice when you are trying to fit things into an 8-page paper limit / 1-page rebuttal limit: ... without any use of bias units ... ... without using bias units ... Does LayerNorm have an effect on ...? Does LayerNorm affect ...? The theorem does not…
Classical Theory: garbage in garbage out Minor domain shift: gold in garbage out Diffusion models: garbage in gold out
New AI research & news everywhere! A short post on my personal approach to keeping up with things. sebastianraschka.com/blog/2023/keep…
🤔 Want to improve your PyTorch model's training performance without sacrificing accuracy? Learn how you can cut training time on a single GPU from 22.53 mins to 2.75 mins and maintain prediction accuracy🤯🚀 Check out this blog by @rasbt: lightning.ai/pages/communit… #PyTorch…
This blog on attention and the intuition by @rasbt is so well written! Bonus: It also has the code to run with. sebastianraschka.com/blog/2023/self…
Reviewers: If you ask the authors to do something and they have followed through successfully, or you made a claim authors successfully refuted, then you need to be prepared to change your recommendation to positive. #ICML2023 1/3
We're excited to release Lit-LLaMA🦙, a minimal, optimized rewrite of LLaMA for training and inference licensed under Apache 2.0 🎉 Check out the repo👉👉 github.com/Lightning-AI/l…
I am old enough to remember people cheering AI when it was defeating the human Go champion
Want to get into AI? My book is a 775-page journey from the fundamentals of machine learning to finetuning large language models. Today is the last day to catch "Machine Learning with PyTorch & Scikit-Learn book" during the Spring Sale – 25% off! 🌱📚 amazon.com/Machine-Learni…
I didn't sign "the letter". Current AI poses lots of risks, but describing these systems as "ever more powerful digital minds" that no one can control is likely to make the problem even worse. What's needed: more transparency and better public discourse.
Speaking of better public discourse, @MelMitchell1 has an excellent newsletter/blog on the state & problems of large language models. (It's free from attention-seeking headlines, and thus probably not as popular as it should be.) Highly recommended: aiguide.substack.com
I didn't sign "the letter". Current AI poses lots of risks, but describing these systems as "ever more powerful digital minds" that no one can control is likely to make the problem even worse. What's needed: more transparency and better public discourse.
Open Source Sunday! Just released a new version of MLxtend: rasbt.github.io/mlxtend/ Featuring - a snappier ExhaustiveFeatureSelector - the H-Mine frequent pattern mining algo - multiprocessing for plot_decision_regions Thx to contributors, Fatih Sen, Nima Sarajpoor & others
United States Trends
- 1. $CUTO 7.298 posts
- 2. #WednesdayMotivation 6.395 posts
- 3. Taina 5.211 posts
- 4. WNBA 39,3 B posts
- 5. Good Wednesday 30,4 B posts
- 6. #wednesdayfeelings 2.129 posts
- 7. Hump Day 15,6 B posts
- 8. $ASTROS 2.076 posts
- 9. Herbo 1.666 posts
- 10. Dreamville 4.101 posts
- 11. #WednesdayWisdom N/A
- 12. #TheHeartKillersEP4 177 B posts
- 13. Sports Illustrated Stadium N/A
- 14. McConnell 70,9 B posts
- 15. Core CPI 3.590 posts
- 16. Cutoshi Farming N/A
- 17. #Wednesdayvibe 2.046 posts
- 18. Caitlin 60,7 B posts
- 19. Happy Hump 9.685 posts
- 20. USPS 20,3 B posts
Something went wrong.
Something went wrong.