glowstick

This crate makes working with tensors in Rust safe, easy, and fun by tracking their shapes in the type system!

Example usage with candle:

use candle::{DType, Device};
use glowstick::{assert_shape_type_eq, num::{U1, U2}, Shape2};
use glowstick_candle::Tensor;

let a: Tensor<Shape2<U2, U1>> = Tensor::zeros(DType::F32, &Device::Cpu).expect("tensor A");
let b: Tensor<Shape2<U1, U2>> = Tensor::zeros(DType::F32, &Device::Cpu).expect("tensor B");
let c = a.matmul(&b).expect("matmul");
assert_shape_type_eq!(c, Shape2<U2, U2>);

Several operations are available:

use candle::{DType, Device};
use glowstick::{assert_shape_type_eq, num::{U0, U1, U8, U64}, Shape1, Shape2};
use glowstick_candle::Tensor;

let my_tensor: Tensor<Shape2<U8, U8>> = Tensor::zeros(DType::F32, &Device::Cpu).expect("tensor");
let reshaped: Tensor<Shape1<U64>> = my_tensor.reshape().expect("reshape");
let unsqueezed = reshaped.unsqueeze::<U0>().expect("unsqueeze");
assert_shape_type_eq!(unsqueezed, Shape2<U1, U64>);
assert_eq!(unsqueezed.inner().dims(), &[1, 64]);

For examples of more extensive usage and integration with popular Rust ML frameworks like candle and Burn, check out the examples directory.

The project is currently pre-1.0: breaking changes will be made!

Features

• Express tensor shapes as types
• Support for dynamic dimensions (gradual typing)
• Human-readable error messages (sort of)