NumRst

NumRst is a Rust implementation of multi-dimensional arrays (NdArray) and numerical computation, inspired by NumPy.

It provides a rich set of numerical operations, including broadcasting, indexing, matrix operations, reductions, logical and comparison ops, aiming to build a high-performance numerical computing tool in the Rust ecosystem.

🚀Features

• Data types supported: Bool, U8, I8, U16, I16, U32, I32, USize, F32, F64

• Operators & Arithmetic:

  • Basic ops: + - / minimum maximum
  • Comparison: eq ne lt le gt ge
  • Logic: and or xor
  • Float functions: exp sin cos sqrt tanh floor ceil round abs neg ln

• NdArray operations:

  • Shape manipulation: reshape, transpose, squeeze, unsqueeze, narrow, narrow_range, repeat, flatten, split
  • Concatenation/stacking: cat, stack

• Broadcasting: NumPy-like broadcasting (broadcast_add, broadcast_sub, etc.)

• Reductions:

  • Global: sum, product, min, max, mean, var, std
  • Along axis: sum_axis, max_axis, argmax_axis, etc.

• Array creation:

  • zeros, ones, arange, rand, randn, trues, falses
  • from_vec, full, new

• Indexing & slicing:

  • Single index: arr.index(2)?
  • Range: arr.index(rng!(1:3))?
  • Mixed multi-dim: arr.index((rng!(1:3), .., 1..2))?

• Matrix operations:

  • matmul

• File format:

  • Support operate with .npy and .npz
  • Custom file format .nrst and .nrsz

• View:

  • Matrix view: MatrixView and MatrixViewMut
  • Matrix view: MatrixView and MatrixViewMut
  • Unsafe view: MatrixViewUsf and VectorViewUsf are unsafe view, use ptr as storage

• Linear algebra:

  • basic: dot, matmul, trace, det, rank
  • norm: frobenius, infinity, spectral, nuclear, pl
  • decompose: lu, qr, cholesky, bidiagonal
  • solve: lu, qr, cholesky, eig(only support symmetric), svd
  • inv: plu_inv, qr_inv, cholesky_inv and inv(using solve)

Quick Start

use numrst::prelude::*;

fn main() -> Result<(), Box<dyn std::error::Error>> {
    // Create a 5x5x5 zero array
    let arr = NdArray::<u32>::zeros((5, 5, 5))?;

    // Basic arithmetic
    let b = NdArray::<u32>::ones((5, 5, 5))?;
    let c = arr.add(&b)?;
    
    // Indexing and slicing
    let sub_arr = c.index((rng!(1:3), .., 2))?;
    
    // Matrix multiplication
    let m1 = NdArray::<f32>::rand(0.0, 1.0, (3, 4))?;
    let m2 = NdArray::<f32>::rand(0.0, 1.0, (4, 5))?;
    let m3 = m1.matmul(&m2)?;
    
    println!("Result shape: {:?}", m3.shape());
    Ok(())
}

License

MIT

Reference

• https://github.com/numpy/numpy
• https://github.com/huggingface/candle