Corgi

A neural network, and tensor dynamic automatic differentiation implementation for Rust.

---

let l1 = Dense::new(input_size, hidden_size, &initializer, Some(&relu));
let l2 = Dense::new(hidden_size, output_size, &initializer, Some(&softmax));
let mut model = Model::new(vec![&mut l1, &mut l2], &gradient_descent, &cross_entropy);

for _ in 0..iterations {
    // array operations are never in-place for corgi, so values never change
    let input = Array::from((vec![batch_size, input_size], vec![...]));
    let target = Array::from((vec![batch_size, output_size], vec![...]));

    let _result = model.forward(input);
    let loss = model.backward(target);
    // update the parameters, and clear gradients (backward pass only sets gradients)
    model.update();

    println!("loss: {}", loss);
}

Design

• Array operations are never in-place, meaning array values are never modified.
• Eager execution.
• Dynamic-as-possible computational graph.

for _ in 0..10 {
    c = &c + &(&a * &b);
    if c[0] > 50.0 {
        c = &c * &a;
    }
}

c.backward(None);

• The Array is responsible differentiates operations done on it for the backward pass.
• No graph structure for ergonomics - an Array contains only its children.
• Arrays do note store consumers (at the moment). They store consumer counts instead.

BLAS

• The openblas, or netlib features can be enabled.
• Versions prior to 0.9.7 of Corgi did not prioritise optimisation, and will be slow.

Tracked Arrays

• Arrays are untracked by default, so if gradients are required, tracked(), or start_tracking() must be used (see the documentation for details).
• Tracked arrays are arrays which require gradients to be computed, and stored.
• For more information, see the documentation for tracked(), and untracked() in array.rs.

Examples

• Fully-connected neural network (full version):

let initializer = initializer::he();
let relu = activation::relu();
let softmax = activation::softmax();
let ce = cost::cross_entropy();
let gd = GradientDescent::new(learning_rate);
let l1 = Dense::new(input_size, hidden_size, &initializer, Some(&relu));
let l2 = Dense::new(hidden_size, output_size, &initializer, Some(&softmax));
let mut model = Model::new(vec![&mut l1, &mut l2], &gd, &ce);

for _ in 0..iterations {
    let mut input = vec![0.0; input_size * batch_size];
    let mut target = vec![0.0; output_size * batch_size];

    // set inputs, and targets

    // arrays in corgi should not be mutated after creation, so we initialise the values first
    let input = Array::from((vec![batch_size, input_size], input));
    let target = Array::from((vec![batch_size, output_size], target));

    let _result = model.forward(input);
    let loss = model.backward(target);
    // update the parameters, and clear gradients (backward pass only sets gradients)
    model.update();

    println!("loss: {}", loss);
}

• Dynamic computational graph:

let a = arr![5.0].tracked();
let b = arr![2.0].tracked();
let mut c = arr![0.0].tracked();

for _ in 0..10 {
    c = &c + &(&a * &b);
    if c[0] > 50.0 {
        c = &c * &a;
    }
}

assert_eq!(c, arr![195300.0]);

c.backward(None);
assert_eq!(c.gradient(), arr![1.0]);
assert_eq!(b.gradient(), arr![97650.0]);
assert_eq!(a.gradient(), arr![232420.0]);

• Custom operation (still needs some work).

Resources

• Shields are from shields.io.
• MIT 6.034 on OpenCourseWare for a primer on Backward Propagation.
• CS231n YouTube recordings for a primer on Convolutional Neural Networks.

A lot of the library was built around being as dynamic as possible, meaning if chosen well, some design choices might be similar to other dynamic computational graph libraries.

Third-party libraries were used, and can be found in Cargo.toml.

Licence

• MIT