HN Explorer

tysam_and

↗ HN profile [ 102 ms ] full profile
Karma
1,158
Created
December 12, 2022 (3y ago)
Submissions
0
get a completed local CIFAR10 run @94% in <1 minute with one single command (w/ a couple of reasonable conditions): https://github.com/tysam-code/hlb-CIFAR10

great for students. better for researchers. best for unhinged scientists who work at the microscopic level.

it's built to be experimented with. many small hacks take 20 seconds or less from ideation to next completed run. very few take over 5 minutes to implement in the right conditions. using the power of science, version 0.7.0 will be coming soon with one less layer (what?!?!)

------------------------------------------------------------

Rest of the bio:

okay so what i really need tbh is someone to take https://github.com/tysam-code/hlb-gpt and make a plot of a trained network using the top n% of the token mass with banded colors representing the lyapunov exponents of the full following sequences (up to about 128 characters or so) mapped onto a poincare disk

bonus if it is clickable so ...

  1. 1 points 0
  2. 1 points 0
  3. 1 points 0
  4. Show HN: Part 4: Train CIFAR10 to 94% in under 8 seconds on a single A100 (github.com)

    Hello everyone, It's been two weeks since we moved under the 10 second mark, and as we've made some more progress with (some very hard) work on the issue in the past two weeks, and we've past our internal 8 second…

    2 points 0
  5. Train CIFAR10 to 94% in under 10 seconds on a single A100 (github.com)

    Hi, My career is currently in this field, and I created this project as (effectively, among other things) a living resume, and to also be a really great workbench for hacking/experimenting on different methods. Testing…

    151 points 54
  6. Show HN: Hlb-CIFAR10 0.2.0: New world record (~<12.38s) on single-GPU CIFAR10 (github.com)

    Hello everyone, After recreating the accuracy/rough speed from David Page's implementation in hlb-CIFAR10 0.1.0 (18.1s on an A100, SXM4, Colab), it was down to some basic NVIDIA kernel profiling to figure out which…

    7 points 5
  7. Show HN: Hlb-CIFAR10 – 94% accuracy in 18.1 seconds, A100, ground-up hackable (github.com)

    Quick Start To get started (if you've done your code scan due diligence and you're setup on CUDA), you can just run `git clone https://github.com/tysam-code/hlb-CIFAR10 && cd hlb-CIFAR10 && python -m pip install -r…

    1 points 0