Show HN: Microcrad – Micrograd Reimplemented in C
github.com
github.com
The basic idea is the same as micrograd: each number is a `Value` node in a computation graph, ops connect nodes, and the `backward` function topologically sorts the graph before applying the chain rule in reverse. The C-specific parts are memory management and two simple data structures I needed to implement backprop: sets and vectors.
The source code is about 1,350 lines, MIT licensed, and well documented. Dependencies are just the standard library and libm. In addition, the repo contains two examples to showcase how the engine works: a toy regression and an MNIST task.
What this is not: a framework to build and train neural networks in production. Being scalar-valued makes it slow, and it wasn't built for numerical robustness or large datasets. There's no commercial aim here; it's a learning project.
If you read through it, I'd like to hear thoughts, both on the ML engineering aspect and on anything that reads as un-idiomatic C.
The rest looks fairly nice but there are a couple of things I would do differently: I would not have the tests for NULL, but use signed integers for indices and dimensions, use a flexible array member to integrate the data into the vector type directly, and omit the capacity field (as long as benchmarking does not show it is really needed). I would also use variably modified types for bounds checking, and with C23 the include guards become largely unnecessary.
(edit: minor edit for clarity)
I can answer about the include guards, though. I consciously added them for portability, following the same general approach that led me to handle the big-endian to little-endian conversion explicitly in examples/mnist/idx.c: even if that safeguard is not strictly necessary on most modern systems, I love the idea that this project is potentially buildable and runnable in most environments.
All identifiers that begin with an underscore are reserved for use as identifiers with file scope in both the ordinary and tag name spaces.
About the second paragraph, first of all, thank you for the suggestions. Can I ask you to elaborate a little on the reasons for your proposals? For instance, even if redundant in some cases, I thought to myself it couldn’t be a bad thing to check for null pointers (though I could improve the error handling itself).
For a similar reason, it makes sense to use signed integers. A signed overflow sanitizer will find the overflow bugs or safely terminate the program. Finding unsigned wraparound bugs is much harder.
I would suggest you keep checking for NULLs. It's a good habit to have to watch over details and to remain cognizant of edge cases. There are tools of course but they are neither standard nor as ubiquitous as C.
Sloppiness becomes a habit and creeps into other aspects.
The first version of the code had relatively few null checks. Later I went through and added them consistently at function boundaries, and that process ended up revealing a surprising number of bugs and bad assumptions.
Sanitizers and static analyzers could have surely caught many of these issues later, but adding the checks was a useful way to reason about the code while writing it. It felt less defensive and more proactively preventive.
btw, I went down the micrograd path with numpy-primitives all the way to building a PyTorch clone that can pre-train and post-train LLMs (https://github.com/workofart/ml-by-hand). My learning focus was on the math/calculus <-> high-level APIs, instead of efficiency. I'm glad to see more people tackling this problem from different angles.
> Reference counting buys correctness and composability, but at a cost.
> Disadvantage #1: you must balance every reference. Each value_create, value_retain, and each operation's implicit retain of its operands has to be matched by a value_release. Forget one and you leak; do one too many and you free memory that is still in use. The training examples in examples/ are verbose precisely because they are scrupulous about this in their error paths; that verbosity is the price of leak-free C.
I’d put it this way: though the idea of ref counting felt very natural at the beginning for my use case, at some point I realized there were probably better techniques to achieve the same goal. I found myself multiple times writing nonsense like:
sum = value_add(v1, v2) mul = value_mul(sum, v3) [...] value_release(sum) value_release(mul)
so that later I could release the sum. When you only have one intermediate value it’s still acceptable, but at 3-4 it starts getting cumbersome.
After asking for feedback, someone rightfully pointed out that the better and faster approach for an autograd engine is using an arena allocator. My reason for saying “rightfully” is that arenas are ideal when you have many “objects” that have the same lifespan, such as the values involved in the forward/backward pass. RC is better when you have a lot of “objects” with independent lifespans.
My guess is that a scalar engine might have an advantage on tiny optimization problems (among others) with only a handful of variables, where the overhead of tensor frameworks dominates the runtime.