I think that the approach taken with Julia will be interesting to watch as well. It's not as mainstream as Python (yet?), but source-to-source AD in Flux / Zygote.jl is particularly cool because it means that any package/code can become gradient-enabled!
Have you ever wondered what will be the ML frameworks of the '20s? In this essay, I examine the directions AI research might take and the requirements they impose, concluding with an overview of what I believe to be two strong candidates: JAX and S4TF.http://inoryy.com/post/next-gen-ml-tools/ …
-
-
-
Agreed, Julia based approaches are quite exciting to follow as well. However, I think it's a mistake to disregard over a decade of scientific community development in Python, which is why I'm leaning more towards S4TF.
- Još 2 druga odgovora
Novi razgovor -
-
-
I love JAX, but it might be a little early to declare a winner for the 2020s, I think there is plenty of innovation still in store :). TF and PyTorch didn't come out until the second half of the 2010s.
-
Of course, and I'm much more reserved with conclusions in the essay. Just a little click-bait tweet to pique interest :)
Kraj razgovora
Novi razgovor -
-
-
I'm not familiar with the project, but the impression I get from the article is that Weld is on the lower level, closer to XLA. In theory, it could even act as a backend to JAX. Though I see no mention of TPUs in the blog, which are on track to become the ML hardware of the '20s.
Kraj razgovora
Novi razgovor -
-
Which of these tools will solve "seamless scalability from a single machine to gigantic clusters"?
-
I neglected to illustrate it, but JAX's `pmap` makes it really easy to distribute a computation across multiple devices and machines (still working out the kinks on the last part). It's demonstrated in the cloud TPU notebooks:https://github.com/google/jax/blob/master/cloud_tpu_colabs/README.md …
- Još 1 odgovor
Novi razgovor -
-
-
Indeed. Agree about Jax. Here's another use case in competitive optimization for training GANs and constrained RLhttps://twitter.com/animaanandkumar/status/1219133562651176965?s=21 …
-
Thank you for sharing! Looks quite interesting, I can't believe I missed it while surveying use cases. I've referenced it in the blog post as well now.
- Još 1 odgovor
Novi razgovor -
Čini se da učitavanje traje već neko vrijeme.
Twitter je možda preopterećen ili ima kratkotrajnih poteškoća u radu. Pokušajte ponovno ili potražite dodatne informacije u odjeljku Status Twittera.