Top all-time · #25013
Another data science student's blog
Personal blog · sgugger.github.ioaiprogramming
Another data science student's blog is a personal blog at sgugger.github.io about ai and programming. Since 2018, 3 different Hacker News readers have linked it 4 times, most often “The 1cycle policy - Train Your Deep Learning Models 5x Faster”. The last post was in 2018, but the archive is still widely read.
Too few sampled comments (2) to say how often HN recommends it.?
Why HN links it
Linked for posts on deep learning training methods, including the 1cycle and superconvergence learning-rate policy and AdamW, alongside discussion of the fast.ai course approach. No comment judges the writing.
Summarised by Claude from a sample of 2 HN comments that link here. Each comment is marked recommending, critical or neutral (citing without judging); how this works.
Most-linked posts?
- The 1cycle policy - Train Your Deep Learning Models 5x Faster2 people · 2 mentions · first linked 2018-04-10
What Hacker News says?
Top-down vs bottom-up depends to some extent on what you enjoy, vs what you need to be patient about. With our top-down approach you do get to all of the 'why', but only after understanding 'how'. So it's good for people who want to get started doing stuff right away, and don't mind waiting a bit to understand the details. It means you can start experimenting and building a good intuition for training models, which…
Comment linking /deep-painterly-harmonization.html
on “Cutting Edge Deep Learning for Coders, Part 2”Submitted Integrate random objects in paintings while preserving their style as a story · 2 points
HN discussionIt's a desirable emergent property. AFAIK, no one knows of another way to achieve it other than by this kind of LR/momentum policy, so informally I think it's OK to use the term "superconvergence" (or "SuperC" for short) for the policy itself. "1 cycle" refers to having a "single cycle" over all epochs, as opposed to multiple cycles, for example as in SGDR. PS. Here's a good walkthrough of the training policy by a c…
Comment linking /the-1cycle-policy.html
on “AdamW and Super-convergence is now the fastest way to train neural nets”Submitted The 1cycle policy - Train Your Deep Learning Models 5x Faster as a story · 1 points
HN discussion