RAI

ML framework with Ergonomic APIs in Rust. Lazy computation and composable transformations.

Installation

cargo add rai

Code snippets

Function transformations (jvp, vjp, grad, value_and_grad)

use rai::{backend::Cpu, grad, DType, Func, Tensor};

fn f(x: &Tensor) -> Tensor {
    x.sin()
}

fn main() {
    let grad_fn = grad(grad(f));

    let backend = &Cpu;
    let x = Tensor::ones([1], DType::F32, backend);
    let grad = grad_fn.apply((&x,));

    println!("{}", grad.dot_graph());
    println!("{}", grad);
}

NN Modules, Optimizer and loss functions

fn loss_fn<M: DifferentiableModule + 'static>(
    model: &M,
    input: &Tensor,
    labels: &Tensor,
) -> (Tensor, Aux<Tensor>) {
    let logits = model.forward(input);
    let loss = softmax_cross_entropy(&logits, labels).mean(..);
    (loss, Aux(logits))
}

fn train_step<O: Optimizer, M: DifferentiableModule + 'static>(
    optimizer: &mut O,
    model: &M,
    input: &Tensor,
    labels: &Tensor,
) {
    let vg_fn = value_and_grad(loss_fn);
    let ((_loss, Aux(_logits)), (grads, ..)) = vg_fn.apply((model, input, labels));
    let mut params = optimizer.step(&grads);
    eval(&params);
    model.update(&mut params);
}

Examples

• linear_regression
  • cargo run --bin linear_regression --release
• mnist
  • cargo run --bin mnist --release

LICENSE

This project is licensed under either of

• Apache License, Version 2.0, (LICENSE-APACHE or http://www.apache.org/licenses/LICENSE-2.0)
• MIT license (LICENSE-MIT or http://opensource.org/licenses/MIT)

at your option.