Corgi

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

BLAS

• The BLAS feature can be enabled, and requires CBLAS if used.

Important Design Notes

• Array values should never be modified from operations; instead, new arrays should be created.
• Arrays are untracked by default, so if gradients are required, tracked(), or start_tracking() must be used (see the documentation for details).

Examples

• Fully-connected MNIST (convolutional neural networks are in-progress).
• Fully-connected neural network (full version): ```rust let initializer = initializer::makehe(); let sigmoid = activation::makesigmoid(); let mse = cost::makemse(); let gd = GradientDescent::new(learningrate); let l1 = Dense::new(inputsize, hiddensize, initializer.clone(), Some(sigmoid)); let l2 = Dense::new(hiddensize, outputsize, initializer.clone(), None); let mut model = Model::new(vec![Box::new(l1), Box::new(l2)], Box::new(gd), mse);

for _ in 0..8 { let mut input = vec![0.0; inputsize * batchsize]; let mut target = vec![0.0; outputsize * batchsize]; for j in 0..batchsize { let x: Float = rng.genrange(-1.0..1.0); let y: Float = rng.genrange(-1.0..1.0); input[inputsize * j] = x; input[inputsize * j + 1] = y; target[outputsize * j] = x.exp(); target[output_size * j + 1] = x.exp() + y.sin(); }

let input = Arrays::new((vec![batch_size, input_size], input));
let target = Arrays::new((vec![batch_size, output_size], target));

let _result = model.forward(input.clone());
let loss = model.backward(target.clone());

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

} * Dynamic computational graph: rust 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); asserteq!(c.gradient(), arr![1.0]); asserteq!(b.gradient(), arr![97650.0]); assert_eq!(a.gradient(), arr![232420.0]); * Custom operation (still needs some work): rust // note proper implementations should handle tracked, and untracked cases let op: array::ForwardOp = Arc::new(|x: &[&Array]| { Arrays::new((x[0].dimensions(), x[0].values().iter().zip(x[1].values()).map(|(x, y)| x * y).collect::>())) });

let opclone = Arc::clone(&op); let backwardop: array::BackwardOp = Arc::new(move |c: &mut Vec, x: &Array| { vec![Some(Array::op(&vec![&c[1], x], Arc::clone(&opclone), None)), Some(Array::op(&vec![&c[0], x], Arc::clone(&opclone), None))] });

let a = arr![1.0, 2.0, 3.0].tracked(); let b = arr![3.0, 2.0, 1.0].tracked(); let mut product = Array::op(&vec![&a, &b], op, Some(backwardop)); asserteq!(product, arr![3.0, 4.0, 3.0]); product.backward(None); asserteq!(b.gradient(), arr![1.0, 2.0, 3.0]); asserteq!(a.gradient(), arr![3.0, 2.0, 1.0]); ```

Design

• Originally worked around the ergonomics of the arr! macro (which however, currently still needs more work).
• Dynamic-as-possible computational graph.
• Did not want to have to manage any 'graph' structures when using Corgi (the Arrays should represent the graph alone).
• Graph became more, and more dependent on threading for the backward pass, and the use of Arc, and Mutex.
• Graphs do note store consumers (at the moment). They store consumer counts instead.

Tracked Arrays

• 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.

Backward Pass

• An informal UML sequence diagram (it's not entirely up to specs, but should give an overview of the process):

Name

• Original name was going to be 'cog-(something)', since Rust's logo is a cog, and since cognition (get it?). But as it turns out, many AI libraries are named 'cog-(something)'. Attempts at permutations of 'cog' sounded awkward, such as 'cogi', for 'cog-intelligence', so the name Corgi was chosen.

Acknowledgements

• Shields are from shields.io.

Licence

• MIT