linfa-nn 0.6.1

use linfa::Float;
use ndarray::{Array2, ArrayBase, ArrayView, Axis, Data, Dimension, Ix2, Zip};
use ndarray_stats::DeviationExt;

/// A distance function that can be used in spatial algorithms such as nearest neighbour.
pub trait Distance<F: Float>: Clone + Send + Sync + Unpin {
    /// Computes the distance between two points. For most spatial algorithms to work correctly,
    /// **this metric must satisfy the Triangle Inequality.**
    ///
    /// Panics if the points have different dimensions.
    fn distance<D: Dimension>(&self, a: ArrayView<F, D>, b: ArrayView<F, D>) -> F;

    /// A faster version of the distance metric that keeps the order of the distance function. That
    /// is, `dist(a, b) > dist(c, d)` implies `rdist(a, b) > rdist(c, d)`. For most algorithms this
    /// is the same as `distance`. Unlike `distance`, this function does **not** need to satisfy
    /// the Triangle Inequality.
    #[inline]
    fn rdistance<D: Dimension>(&self, a: ArrayView<F, D>, b: ArrayView<F, D>) -> F {
        self.distance(a, b)
    }

    /// Converts the result of `rdistance` to `distance`
    #[inline]
    fn rdist_to_dist(&self, rdist: F) -> F {
        rdist
    }

    /// Converts the result of `distance` to `rdistance`
    #[inline]
    fn dist_to_rdist(&self, dist: F) -> F {
        dist
    }
}

/// L1 or [Manhattan](https://en.wikipedia.org/wiki/Taxicab_geometry) distance
#[derive(Debug, Clone, PartialEq, Eq)]
pub struct L1Dist;
impl<F: Float> Distance<F> for L1Dist {
    #[inline]
    fn distance<D: Dimension>(&self, a: ArrayView<F, D>, b: ArrayView<F, D>) -> F {
        a.l1_dist(&b).unwrap()
    }
}

/// L2 or [Euclidean](https://en.wikipedia.org/wiki/Euclidean_distance) distance
#[derive(Debug, Clone, PartialEq, Eq)]
pub struct L2Dist;
impl<F: Float> Distance<F> for L2Dist {
    #[inline]
    fn distance<D: Dimension>(&self, a: ArrayView<F, D>, b: ArrayView<F, D>) -> F {
        F::from(a.l2_dist(&b).unwrap()).unwrap()
    }

    #[inline]
    fn rdistance<D: Dimension>(&self, a: ArrayView<F, D>, b: ArrayView<F, D>) -> F {
        F::from(a.sq_l2_dist(&b).unwrap()).unwrap()
    }

    #[inline]
    fn rdist_to_dist(&self, rdist: F) -> F {
        rdist.sqrt()
    }

    #[inline]
    fn dist_to_rdist(&self, dist: F) -> F {
        dist.powi(2)
    }
}

/// L-infinte or [Chebyshev](https://en.wikipedia.org/wiki/Chebyshev_distance) distance
#[derive(Debug, Clone, PartialEq, Eq)]
pub struct LInfDist;
impl<F: Float> Distance<F> for LInfDist {
    #[inline]
    fn distance<D: Dimension>(&self, a: ArrayView<F, D>, b: ArrayView<F, D>) -> F {
        a.linf_dist(&b).unwrap()
    }
}

/// L-p or [Minkowsky](https://en.wikipedia.org/wiki/Minkowski_distance) distance
#[derive(Debug, Clone, PartialEq)]
pub struct LpDist<F: Float>(pub F);
impl<F: Float> LpDist<F> {
    pub fn new(p: F) -> Self {
        LpDist(p)
    }
}
impl<F: Float> Distance<F> for LpDist<F> {
    #[inline]
    fn distance<D: Dimension>(&self, a: ArrayView<F, D>, b: ArrayView<F, D>) -> F {
        Zip::from(&a)
            .and(&b)
            .fold(F::zero(), |acc, &a, &b| acc + (a - b).abs().powf(self.0))
            .powf(F::one() / self.0)
    }
}

/// Computes a similarity matrix with gaussian kernel and scaling parameter `eps`
///
/// The generated matrix is a upper triangular matrix with dimension NxN (number of observations) and contains the similarity between all permutations of observations
/// similarity
pub fn to_gaussian_similarity<F: Float>(
    observations: &ArrayBase<impl Data<Elem = F>, Ix2>,
    eps: F,
    dist_fn: &impl Distance<F>,
) -> Array2<F> {
    let n_observations = observations.len_of(Axis(0));
    let mut similarity = Array2::eye(n_observations);

    for i in 0..n_observations {
        for j in 0..n_observations {
            let a = observations.row(i);
            let b = observations.row(j);

            let distance = dist_fn.distance(a, b);
            similarity[(i, j)] = (-distance / eps).exp();
        }
    }

    similarity
}

#[cfg(test)]
mod test {
    use approx::assert_abs_diff_eq;
    use ndarray::arr1;

    use super::*;

    #[test]
    fn autotraits() {
        fn has_autotraits<T: Send + Sync + Sized + Unpin>() {}
        has_autotraits::<L1Dist>();
        has_autotraits::<L2Dist>();
        has_autotraits::<LInfDist>();
        has_autotraits::<LpDist<f64>>();
    }

    fn dist_test<D: Distance<f64>>(dist: D, result: f64) {
        let a = arr1(&[0.5, 6.6]);
        let b = arr1(&[4.4, 3.0]);
        let ab = dist.distance(a.view(), b.view());
        assert_abs_diff_eq!(ab, result, epsilon = 1e-3);
        assert_abs_diff_eq!(dist.rdist_to_dist(dist.dist_to_rdist(ab)), ab);

        let a = arr1(&[f64::INFINITY, 6.6]);
        let b = arr1(&[4.4, f64::NEG_INFINITY]);
        assert!(dist.distance(a.view(), b.view()).is_infinite());

        // Triangle equality
        let a = arr1(&[0.5, 6.6]);
        let b = arr1(&[4.4, 3.0]);
        let c = arr1(&[-4.5, 3.3]);
        let ab = dist.distance(a.view(), b.view());
        let bc = dist.distance(b.view(), c.view());
        let ac = dist.distance(a.view(), c.view());
        assert!(ab + bc > ac)
    }

    #[test]
    fn l1_dist() {
        dist_test(L1Dist, 7.5);
    }

    #[test]
    fn l2_dist() {
        dist_test(L2Dist, 5.3075);

        // Check squared distance
        let a = arr1(&[0.5, 6.6]);
        let b = arr1(&[4.4, 3.0]);
        assert_abs_diff_eq!(L2Dist.rdistance(a.view(), b.view()), 28.17, epsilon = 1e-3);
    }

    #[test]
    fn linf_dist() {
        dist_test(LInfDist, 3.9);
    }

    #[test]
    fn lp_dist() {
        dist_test(LpDist(3.3), 4.635);
    }
}