Implement numpy's `triu` and `tril` methods

[geetmankar]

Hi, I want to propose an implementation of numpy.triu and numpy.tril methods (link here) for arrays that are atleast 2 dimensions.

import numpy as np

np.triu([[1,2,3],[4,5,6],[7,8,9],[10,11,12]], k=0)

returns:

array([[ 1,  2,  3],
       [ 0,  5,  6],
       [ 0,  0,  9],
       [ 0,  0,  0]])

and

np.triu([[1,2,3],[4,5,6],[7,8,9],[10,11,12]], k=1)

returns:

array([[ 0,  2,  3],
       [ 0,  0,  6],
       [ 0,  0,  0],
       [ 0,  0,  0]])

and

np.triu([[1,2,3],[4,5,6],[7,8,9],[10,11,12]], k=-1)

returns:

array([[ 1,  2,  3],
       [ 4,  5,  6],
       [ 0,  8,  9],
       [ 0,  0, 12]])

I implemented a basic version of np.triu and np.tril for 2D owned float square ndarrays for personal use, however I did not write it to support negative values of k.

pub trait Tri {
    fn triu(&self, k: usize) -> Self;
    fn tril(&self, k: usize) -> Self;
}

impl Tri for ArrayBase<OwnedRepr<f64>, Ix2> {
    fn triu(&self, k: usize) -> Self {
        let cols = self.len_of(Axis(1));
        let copied_arr = self.clone();

        let mut result_arr = Array2::<f64>::zeros((cols, cols).f());

        for (i, row) in copied_arr.axis_iter(Axis(0)).enumerate() {
            let slice_main = (i as usize + k)..self.len_of(Axis(0));
            row.to_owned()
                .slice(s![slice_main.clone()])
                .assign_to(result_arr.slice_mut(s![i, slice_main.clone()]));
        }

        return result_arr;
    }

   fn tril(&self, k: usize) -> Self {
        return self.clone().t().to_owned().triu(k).t().to_owned();
    }
}

[geetmankar]

A test of this code:

use ndarray::Array

// ... paste the `Tri` implementation and trait here or import it

fn main() {
    let a = Array::<f64, _>::ones((3, 3).f());
    println!("a.triu(0) = {:?}\n", a.clone().triu(0));
    println!("a.triu(1) = {:?}\n\n", a.clone().triu(1));
    println!("a.tril(0) = {:?}\n", a.clone().tril(0));
    println!("a.tril(1) = {:?}\n", a.clone().tril(1));
}

The result is:

Compiling nbody_code v0.1.0 (C:\Users\Rebel1\projects\nbody_code)
    Finished dev [unoptimized + debuginfo] target(s) in 1.40s
     Running `target\debug\nbody_code.exe`
a.triu(0) = [[1.0, 1.0, 1.0],
             [0.0, 1.0, 1.0],
             [0.0, 0.0, 1.0]], shape=[3, 3], strides=[1, 3], layout=Ff (0xa), const ndim=2

a.triu(1) = [[0.0, 1.0, 1.0],
             [0.0, 0.0, 1.0],
             [0.0, 0.0, 0.0]], shape=[3, 3], strides=[1, 3], layout=Ff (0xa), const ndim=2


a.tril(0) = [[1.0, 0.0, 0.0],
             [1.0, 1.0, 0.0],
             [1.0, 1.0, 1.0]], shape=[3, 3], strides=[3, 1], layout=Cc (0x5), const ndim=2  

a.tril(1) = [[0.0, 0.0, 0.0],
             [1.0, 0.0, 0.0],
             [1.0, 1.0, 0.0]], shape=[3, 3], strides=[3, 1], layout=Cc (0x5), const ndim=2  

[nilgoyette]

ndarray should probably have those functions.

Making it generic, and with less to_owned and clone will require some work. Thank you for writing the proposition.

[geetmankar]

Sweet, I'll be waiting for the upgrade.

[nilgoyette]

I'll be waiting for the upgrade.

Just to be clear, we're mostly offering bug fixes and support at the moment because all maintainers are busy. If you are actually waiting for it, I advise you to be patient :)

[geetmankar]

No worries, the code snippet I wrote is currently good enough for my pet project, so I won't need to hold my breath. Thanks for the heads-up though!

[akern40]

I've been wanting to get involved in ndarray development, and this seemed like a good place to get my feet wet! Hope you don't mind, but I've put in a PR that creates these methods (along with some tests and documentation). They're designed to precisely mimic NumPy's approach. Would love feedback; in particular, I added them to an impl block of ArrayBase, but I wasn't sure whether they should instead be free functions or have their own traits.