ferris_grad

A PyTorch-like autograd engine in under 1000 lines of Rust code.🦀

How ferris_grad works?

ferris_grad is just build in 3 files:

file	description
scalar.rs	implement the scalar data structure, compute graph, backward rule, and basic operations.
tensor.rs	implement the tensor data structure and basic operations.
nn.rs	implement the neural network layers and modules.

How to use ferris_grad?

Add this to your Cargo.toml:

[dependencies]
ferris_grad = "*"

and then you can use ferris_grad in your project:

use anyhow::Result;
use ferris_grad::{Tensor, nn::Module};

fn main() -> Result<()> {
    let a = Tensor::from_vec(vec![1.0.into(), 2.0.into(), 3.0.into()], [3, 1].into())?;
    let b = Tensor::rand([3, 1].into())?;
    let c = &a * &b;
    println!("{}", c);
    Ok(())
}

Examples

All the examples are in the examples folder.

file	description
examples/tensor_example.rs	basic usage of tensor api.
examples/backward_scalar.rs	basic usage of backward api.
examples/sgd.rs	training a MLP example using sgd.
examples/sgd_with_plotter.rs	training a MLP example using sgd with plotter.
examples/rnn.rs	training a recurrent neural network in ferris_grad.
examples/gpt.rs	training a mini GPT in ferris_grad!

License

This project is licensed under the MIT License - see the LICENSE file for details.