rapl

注意： rapl 处于早期开发阶段，未针对性能优化，不建议用于生产应用。

rapl 是一个 Rust 库，提供了与 N 维数组简单操作的方式，并附带一系列用于操作它们的数学函数。它受到了 NumPy 和 APL 的启发，主要目标是实现最大的易用性和用户友好性，同时保持通用性。

我们的目标是使 Rust 脚本尽可能高效，使 Rust 成为数值计算和数据科学的一个真实选择。查看 示例。

开箱即用的 rapl 提供了诸如 协同广播、秩类型检查、原生复数支持 等功能

use rapl::*;
fn main() {
    let a = Ndarr::from([1 + 1.i(), 2 + 1.i()]);
    let b = Ndarr::from([[1, 2], [3, 4]]);
    let r = a + b - 2;
    assert_eq!(r, Ndarr::from([[1.i(), 2 + 1.i()],[2 + 1.i(), 4 + 1.i()]]));
}

数组初始化

有多种方便的方法可以初始化 N 维数组（或 Ndarr）。

• 从原生 Rust 数组到 Ndarr。

let a = Ndarr::from(["a","b","c"]); 
let b = Ndarr::from([[1,2],[3,4]]);

• 从范围。

let a = Ndarr::from(1..7).reshape(&[2,3])

• 从 &str

let chars = Ndarr::from("Hello rapl!"); //Ndarr<char,U1>

• 其他

let ones: Ndarr<f32, 2> = Ndarr::ones(&[4,4]);
let zeros : Ndarr<i32, 3>= Ndarr::zeros(&[2,3,4]);
let letter_a = Ndarr::fill("a", &[5]);
let fold = Ndarr::new(data: &[0, 1, 2, 3], shape: [2, 2]).expect("Error initializing");

• linspace, logspace, geomspace

    let linear = Ndarr::linspace(0, 9, 10);
    assert_eq!(linear,Ndarr::from(0..10));

    let logarithmic = Ndarr::logspace(0.,9., 10., 10);
    assert!(logarithmic.approx(&Ndarr::from([1.,1e1, 1e2, 1e3, 1e4, 1e5, 1e6, 1e7, 1e8, 1e9])));

    let geom = Ndarr::geomspace(1.,256., 9);
    assert!(geom.approx(&Ndarr::from([1., 2., 4., 8., 16., 32., 64., 128., 256.])));

随机数组创建

您可以轻松地创建任何形状的随机数组

//Normal distribution
let arr_norm = NdarrRand::normal(low: 0f32, high: 1f32, shape: [2, 2], Seed: Some(1234));
//Normal distribution
let arr_uniform = NdarrRand::uniform(low: 0f32, high: 1f32, shape: [10], Seed: None);
//Choose between values
let arr_choose = NdarrRand::choose(&[1, 2, 3, 4, 5], [3, 3], Some(1234));

逐元素操作

• 与标量进行算术运算

let ones: Ndarr<i32, 2> = Ndarr::ones(&[4,4]);
let twos = ones + 1;
let sixes = twos * 3;

• 在 Ndarr 之间进行算术运算

let a = Ndarr::from([[1,2],[3,4]]);
let b = Ndarr::from([[1,2],[-3,-4]]);

assert_eq!(a + b, Ndarr::from([[2,4],[0,0]]))

注意：如果形状不相等，rapl 将自动将数组广播到兼容的形状（如果存在）并执行操作。

• 包括三角函数和激活函数在内的数学运算。

let x = Ndarr::from([-1.0 , -0.8, -0.6, -0.4, -0.2, 0.0, 0.2, 0.4, 0.6, 0.8, 1.0]);
let sin_x = x.sin();
let cos_x = x.cos();
let tanh_x = x.tanh();

let abs_x = x.abs();
let relu_x = x.relu();

• 映射函数

let a = Ndarr::from([[1,2],[3,4]]);
let mapped = a.map(|x| x*2-1);

单调张量操作

• 转置

let arr = Ndarr::from([[1,2,3],[4,5,6]]);	
assert_eq!(arr.shape(), [2,3]);
assert_eq!(arr.clone().t().shape, [3,2]); //transpose

• 重塑

let a = Ndarr::from(1..7).reshape(&[2,3]).unwrap();

• 切片

let arr = Ndarr::from([[1,2],[3,4]]);

assert_eq!(arr.slice_at(1)[0], Ndarr::from([1,3]))

• 减少

let sum_axis = arr.clone().reduce(1, |x,y| x + y).unwrap();
assert_eq!(sum_axis, Ndarr::from([6, 15])); //sum reduction

• 向右和向左扫描

 let s = Ndarr::from([1,2,3]);
 let cumsum = s.scanr( 0, |x,y| x + y);
 assert_eq!(cumsum, Ndarr::from([1,3,6]));

• 滚动

let a = Ndarr::from([[1, 2], [3, 4]]);
assert_eq!(a.roll(1, 1), Ndarr::from([[2, 1], [4, 3]]))

动态张量操作

• 兼容数组之间的一般矩阵乘法

use rapl::*
use rapl::ops::{mat_mul};
let a = Ndarr::from(1..7).reshape(&[2,3]).unwrap();
let b = Ndarr::from(1..7).reshape(&[3,2]).unwrap();
    
let matmul = mat_mul(a, b))

• APL 启发的内积。

    let a = Ndarr::from(1..7).reshape(&[2,3]).unwrap();
    let b = Ndarr::from(1..7).reshape(&[3,2]).unwrap();
    
    let inner = rapl::ops::inner_product(|x,y| x*y, |x,y| x+y, a.clone(), b.clone());
    assert_eq!(inner, rapl::ops::mat_mul(a, b))

• 外积。

    let suits = Ndarr::from(["♣","♠","♥","♦"]);
    let ranks = Ndarr::from(["2","3","4","5","6","7","8","9","10","J","Q","K","A"]);

    let add_str = |x: &str, y: &str| (x.to_owned() + y);

    let deck = ops::outer_product( add_str, ranks, suits).flatten(); //All cards in a deck

复数

您可以通过简单且干净的用户界面，在原生数值类型和复数类型 C<T> 之间进行操作。

use rapl::*;
// Complex sclars
    let z = 1 + 2.i();
    assert_eq!(z, C(1,2));
    assert_eq!(z - 3, -2 + 2.i());

use rapl::*;
// Complex tensors
let arr = Ndarr::from([1, 2, 3]);
let arr_z = arr + -1 + 2.i();
assert_eq!(arr_z, Ndarr::from([C(0,2), C(1,2), C(2,2)]));
assert_eq!(arr_z.im(), Ndarr::from([2,2,2]));

    let signal = Ndarr::linspace(-10., 10., 100).sin();
    let signal_fft = signal.to_complex().fft();

use rapl::*;
use rapl::utils::rapl_img;

fn main() {
    //open RGB image as  Ndarr<u8,3>
    let img: Ndarr<u8,U3> = rapl_img::open_rgbu8(&"image_name.jpg").unwrap();
    //Split RGB channels by Slicing along 3'th axis.
    let channels: Vec<Ndarr<u8,U2>> = img.slice_at(2);
    //select blue channel and save it as black and white image.
    channels[2].save_as_luma(&"blue_channel.png", rapl_img::ImageFormat::Png);
}