MyTorch – Minimalist autograd in 450 lines of Python | Hacker News Reader

MyTorch – Minimalist autograd in 450 lines of Python

github.com

100 pointsiguana200019 comments

Having written a slightly more involved version of this recently myself I think you did a great job of keeping this compact while still readable. This style of library requires some design for sure.

Supporting higher order derivatives was also something I considered, but it’s basically never needed in production models from what I’ve seen.

Thanks! I agree about the style

Karpathy’s micrograd did it first (and better); start here: https://karpathy.ai/zero-to-hero.html

Harsh.

Why is it better

Because it's an acclaimed, often cited course by a preeminent AI Researcher (and founding member of OAI) rather than four undocumented python files.

Objective measures like branch depth, execution speed, memory use and correctness of the results be damned.

Karpathy's implementation is explicitly for teaching purposes. It's meant to be taken in alongside his videos, which are pretty awesome.

it being acclaimed is a poor measure of success, theres always room for improvement, how about some objective comparisons?

Ironically the reason Karpathy's is better is because he livecoded it and I can be sure it's not some LLM vomit. Unfortunately, we are now indundated with newbies posting their projects/tutorials/guides in the hopes that doing so will catch the eye of a recuiter and land them a high paying AI job. That's not so bad in itself except for the fact that most of these people are completely clueless and posting AI slop.

Haha, couldn't agree with you more. This, however, isn't AI slop. You can see in the commit history that this is from 3 years ago

Cleaner, more straightforward, more compact code, and considered complete in its scope (i.e. implement backpropagation with a PyTorch-y API and train a neural network with it). MyTorch appears to be an author's self-experiment without concrete vision/plan. This is better for author but worse for outsiders/readers.

P.S. Course goes far beyond micrograd, to makemore (transfomers), minbpe (tokenization), and nanoGPT (LLM training/loading).

Imho, we should let people experiment as much as they want. Having more examples is better than less. Still, thanks for the link for the course, this is a top-notch one

Karpathy's material is excellent! This was a project I made for fun, and hopefully provides a different perspective on how this can look

I'm very sorry, I should have phrased my original post in a kinder, less dismissive way, and kudos to you for not reacting badly to my rudeness. It is a cool repo and a great accomplishment. Implementing autograd is great as a learning exercise, but my opinion is that you're not going to get the performance or functionality of one of the large, mainstream autograd libraries. Karpathy, for example, throws away micrograd after implementing it and uses pytorch in his later exercises. So it's great that you did this, but for others to learn how autograd works, Karpathy is usually a better route, because the concepts are built up one by one and explained thoroughly.

No worries, you're good, yes Karpathy is for sure the better route

Better readme would be way to go

In iguana2000's defense, the code is highly self-documenting.

It arguably reads cleaner than Karpathy's in some respects, as he occasionally gets a little ahead of his students with his '1337 Python skillz.

HmcKk

Reply on news.ycombinator.com