|
| 1 | +use ndarray::{ArrayBase, Data, Dimension, Zip}; |
| 2 | +use num_traits::{Float, Signed, ToPrimitive}; |
| 3 | +use std::convert::Into; |
| 4 | +use std::ops::AddAssign; |
| 5 | + |
| 6 | +/// Extension trait for `ArrayBase` providing functions |
| 7 | +/// to compute different deviation measures. |
| 8 | +pub trait DeviationExt<A, S, D> |
| 9 | +where |
| 10 | + S: Data<Elem = A>, |
| 11 | + D: Dimension, |
| 12 | +{ |
| 13 | + fn count_eq(&self, other: &ArrayBase<S, D>) -> usize |
| 14 | + where |
| 15 | + A: PartialEq; |
| 16 | + |
| 17 | + fn count_neq(&self, other: &ArrayBase<S, D>) -> usize |
| 18 | + where |
| 19 | + A: PartialEq; |
| 20 | + |
| 21 | + fn sq_l2_dist(&self, other: &ArrayBase<S, D>) -> A |
| 22 | + where |
| 23 | + A: AddAssign + Clone + Signed; |
| 24 | + |
| 25 | + fn l2_dist(&self, other: &ArrayBase<S, D>) -> f64 |
| 26 | + where |
| 27 | + A: AddAssign + Clone + Signed + ToPrimitive; |
| 28 | + |
| 29 | + fn l1_dist(&self, other: &ArrayBase<S, D>) -> A |
| 30 | + where |
| 31 | + A: AddAssign + Clone + Signed; |
| 32 | + |
| 33 | + fn linf_dist(&self, other: &ArrayBase<S, D>) -> A |
| 34 | + where |
| 35 | + A: Clone + PartialOrd + Signed; |
| 36 | + |
| 37 | + fn gkl_div(&self, other: &ArrayBase<S, D>) -> A |
| 38 | + where |
| 39 | + A: AddAssign + Clone + Float; |
| 40 | + |
| 41 | + fn mean_abs_dev(&self, other: &ArrayBase<S, D>) -> f64 |
| 42 | + where |
| 43 | + A: AddAssign + Clone + Signed + Into<f64>; |
| 44 | + |
| 45 | + fn max_abs_dev(&self, other: &ArrayBase<S, D>) -> A |
| 46 | + where |
| 47 | + A: Clone + PartialOrd + Signed; |
| 48 | + |
| 49 | + fn mean_sq_dev(&self, other: &ArrayBase<S, D>) -> f64 |
| 50 | + where |
| 51 | + A: AddAssign + Clone + Signed + Into<f64>; |
| 52 | + |
| 53 | + fn root_mean_sq_dev(&self, other: &ArrayBase<S, D>) -> f64 |
| 54 | + where |
| 55 | + A: AddAssign + Clone + Signed + Into<f64>; |
| 56 | + |
| 57 | + fn peak_signal_to_noise_ratio(&self, other: &ArrayBase<S, D>, maxv: A) -> f64 |
| 58 | + where |
| 59 | + A: AddAssign + Clone + Signed + Into<f64>; |
| 60 | +} |
| 61 | + |
| 62 | +impl<A, S, D> DeviationExt<A, S, D> for ArrayBase<S, D> |
| 63 | +where |
| 64 | + S: Data<Elem = A>, |
| 65 | + D: Dimension, |
| 66 | +{ |
| 67 | + fn count_eq(&self, other: &ArrayBase<S, D>) -> usize |
| 68 | + where |
| 69 | + A: PartialEq, |
| 70 | + { |
| 71 | + let mut c = 0; |
| 72 | + |
| 73 | + Zip::from(self).and(other).apply(|a, b| { |
| 74 | + if a == b { |
| 75 | + c += 1; |
| 76 | + } |
| 77 | + }); |
| 78 | + |
| 79 | + c |
| 80 | + } |
| 81 | + |
| 82 | + fn count_neq(&self, other: &ArrayBase<S, D>) -> usize |
| 83 | + where |
| 84 | + A: PartialEq, |
| 85 | + { |
| 86 | + self.len() - self.count_eq(other) |
| 87 | + } |
| 88 | + |
| 89 | + fn sq_l2_dist(&self, other: &ArrayBase<S, D>) -> A |
| 90 | + where |
| 91 | + A: AddAssign + Clone + Signed, |
| 92 | + { |
| 93 | + let mut r = A::zero(); |
| 94 | + |
| 95 | + Zip::from(self).and(other).apply(|self_i, other_i| { |
| 96 | + let (a, b) = (self_i.clone(), other_i.clone()); |
| 97 | + let abs_diff = (a - b).abs(); |
| 98 | + r += abs_diff.clone() * abs_diff; |
| 99 | + }); |
| 100 | + |
| 101 | + r |
| 102 | + } |
| 103 | + |
| 104 | + fn l2_dist(&self, other: &ArrayBase<S, D>) -> f64 |
| 105 | + where |
| 106 | + A: AddAssign + Clone + Signed + ToPrimitive, |
| 107 | + { |
| 108 | + self.sq_l2_dist(other).to_f64().unwrap().sqrt() |
| 109 | + } |
| 110 | + |
| 111 | + fn l1_dist(&self, other: &ArrayBase<S, D>) -> A |
| 112 | + where |
| 113 | + A: AddAssign + Clone + Signed, |
| 114 | + { |
| 115 | + let mut r = A::zero(); |
| 116 | + |
| 117 | + Zip::from(self).and(other).apply(|self_i, other_i| { |
| 118 | + let (a, b) = (self_i.clone(), other_i.clone()); |
| 119 | + r += (a - b).abs(); |
| 120 | + }); |
| 121 | + |
| 122 | + r |
| 123 | + } |
| 124 | + |
| 125 | + fn linf_dist(&self, other: &ArrayBase<S, D>) -> A |
| 126 | + where |
| 127 | + A: Clone + PartialOrd + Signed, |
| 128 | + { |
| 129 | + let mut max = A::zero(); |
| 130 | + |
| 131 | + Zip::from(self).and(other).apply(|self_i, other_i| { |
| 132 | + let (a, b) = (self_i.clone(), other_i.clone()); |
| 133 | + let diff = (a - b).abs(); |
| 134 | + if diff > max { |
| 135 | + max = diff; |
| 136 | + } |
| 137 | + }); |
| 138 | + |
| 139 | + max |
| 140 | + } |
| 141 | + |
| 142 | + fn gkl_div(&self, other: &ArrayBase<S, D>) -> A |
| 143 | + where |
| 144 | + A: AddAssign + Clone + Float, |
| 145 | + { |
| 146 | + let mut r = A::zero(); |
| 147 | + |
| 148 | + Zip::from(self).and(other).apply(|self_i, other_i| { |
| 149 | + let (a, b) = (self_i.clone(), other_i.clone()); |
| 150 | + r += a * (a / b).ln() - a + b; |
| 151 | + }); |
| 152 | + |
| 153 | + r |
| 154 | + } |
| 155 | + |
| 156 | + fn mean_abs_dev(&self, other: &ArrayBase<S, D>) -> f64 |
| 157 | + where |
| 158 | + A: AddAssign + Clone + Signed + Into<f64>, |
| 159 | + { |
| 160 | + let a: f64 = self.l1_dist(other).into(); |
| 161 | + let b = self.len().to_f64().unwrap(); |
| 162 | + a / b |
| 163 | + } |
| 164 | + |
| 165 | + #[inline] |
| 166 | + fn max_abs_dev(&self, other: &ArrayBase<S, D>) -> A |
| 167 | + where |
| 168 | + A: Clone + PartialOrd + Signed, |
| 169 | + { |
| 170 | + self.linf_dist(other) |
| 171 | + } |
| 172 | + |
| 173 | + fn mean_sq_dev(&self, other: &ArrayBase<S, D>) -> f64 |
| 174 | + where |
| 175 | + A: AddAssign + Clone + Signed + Into<f64>, |
| 176 | + { |
| 177 | + let a: f64 = self.sq_l2_dist(other).into(); |
| 178 | + let b = self.len().to_f64().unwrap(); |
| 179 | + a / b |
| 180 | + } |
| 181 | + |
| 182 | + fn root_mean_sq_dev(&self, other: &ArrayBase<S, D>) -> f64 |
| 183 | + where |
| 184 | + A: AddAssign + Clone + Signed + Into<f64>, |
| 185 | + { |
| 186 | + self.mean_sq_dev(other).sqrt() |
| 187 | + } |
| 188 | + |
| 189 | + fn peak_signal_to_noise_ratio(&self, other: &ArrayBase<S, D>, maxv: A) -> f64 |
| 190 | + where |
| 191 | + A: AddAssign + Clone + Signed + Into<f64>, |
| 192 | + { |
| 193 | + let maxv_f: f64 = maxv.into(); |
| 194 | + 10. * f64::log10(maxv_f * maxv_f / self.mean_sq_dev(&other)) |
| 195 | + } |
| 196 | +} |
| 197 | + |
| 198 | +#[cfg(test)] |
| 199 | +mod tests { |
| 200 | + use super::*; |
| 201 | + use approx::assert_abs_diff_eq; |
| 202 | + use ndarray::*; |
| 203 | + use ndarray_rand::RandomExt; |
| 204 | + use rand::distributions::Uniform; |
| 205 | + use std::f64; |
| 206 | + |
| 207 | + #[test] |
| 208 | + fn test_count_eq() { |
| 209 | + let a = array![0., 0.]; |
| 210 | + let b = array![1., 0.]; |
| 211 | + let c = array![0., 1.]; |
| 212 | + let d = array![1., 1.]; |
| 213 | + |
| 214 | + assert_eq!(a.count_eq(&a), 2); |
| 215 | + assert_eq!(a.count_eq(&b), 1); |
| 216 | + assert_eq!(a.count_eq(&c), 1); |
| 217 | + assert_eq!(a.count_eq(&d), 0); |
| 218 | + } |
| 219 | + |
| 220 | + #[test] |
| 221 | + fn test_count_neq() { |
| 222 | + let a = array![0., 0.]; |
| 223 | + let b = array![1., 0.]; |
| 224 | + let c = array![0., 1.]; |
| 225 | + let d = array![1., 1.]; |
| 226 | + |
| 227 | + assert_eq!(a.count_neq(&a), 0); |
| 228 | + assert_eq!(a.count_neq(&b), 1); |
| 229 | + assert_eq!(a.count_neq(&c), 1); |
| 230 | + assert_eq!(a.count_neq(&d), 2); |
| 231 | + } |
| 232 | + |
| 233 | + #[test] |
| 234 | + fn test_sq_l2_dist() { |
| 235 | + let a = array![0., 1., 4., 2.]; |
| 236 | + let b = array![1., 1., 2., 4.]; |
| 237 | + |
| 238 | + assert_eq!(a.sq_l2_dist(&b), 9.); |
| 239 | + } |
| 240 | + |
| 241 | + #[test] |
| 242 | + fn test_l2_dist() { |
| 243 | + let a = array![0., 1., 4., 2.]; |
| 244 | + let b = array![1., 1., 2., 4.]; |
| 245 | + |
| 246 | + assert_eq!(a.l2_dist(&b), 3.); |
| 247 | + } |
| 248 | + |
| 249 | + #[test] |
| 250 | + fn test_l1_dist() { |
| 251 | + let a = array![0., 1., 4., 2.]; |
| 252 | + let b = array![1., 1., 2., 4.]; |
| 253 | + |
| 254 | + assert_eq!(a.l1_dist(&b), 5.); |
| 255 | + } |
| 256 | + |
| 257 | + #[test] |
| 258 | + fn test_linf_dist() { |
| 259 | + let a = array![0., 0.]; |
| 260 | + let b = array![1., 0.]; |
| 261 | + let c = array![1., 2.]; |
| 262 | + |
| 263 | + assert_eq!(a.linf_dist(&a), 0.); |
| 264 | + |
| 265 | + assert_eq!(a.linf_dist(&b), 1.); |
| 266 | + assert_eq!(b.linf_dist(&a), 1.); |
| 267 | + |
| 268 | + assert_eq!(a.linf_dist(&c), 2.); |
| 269 | + assert_eq!(c.linf_dist(&a), 2.); |
| 270 | + } |
| 271 | + |
| 272 | + #[test] |
| 273 | + fn test_gkl_div() { |
| 274 | + let a = Array::random((5,), Uniform::new(0., 1.)); |
| 275 | + let b = Array::random((5,), Uniform::new(1., 2.)); |
| 276 | + let c = Array::random((5,), Uniform::new(-1., 0.)); |
| 277 | + |
| 278 | + assert_eq!(a.gkl_div(&a), 0.); |
| 279 | + assert!(a.gkl_div(&b) > 0.); |
| 280 | + assert!(b.gkl_div(&a) > 0.); |
| 281 | + assert_ne!(a.gkl_div(&b), b.gkl_div(&a)); |
| 282 | + |
| 283 | + // TODO: what is the sign check logic doing in StatsBase.jl impl? |
| 284 | + assert!(f64::is_nan(a.gkl_div(&c))); |
| 285 | + } |
| 286 | + |
| 287 | + #[test] |
| 288 | + fn test_mean_abs_dev() { |
| 289 | + let a = array![1., 1.]; |
| 290 | + let b = array![3., 5.]; |
| 291 | + |
| 292 | + assert_eq!(a.mean_abs_dev(&a), 0.); |
| 293 | + assert_eq!(a.mean_abs_dev(&b), 3.); |
| 294 | + assert_eq!(b.mean_abs_dev(&a), 3.); |
| 295 | + } |
| 296 | + |
| 297 | + #[test] |
| 298 | + fn test_max_abs_dev() { |
| 299 | + // This is effectively an alias for linf_dist, so not retesting deeply |
| 300 | + let a = array![0., 0.]; |
| 301 | + let b = array![2., 4.]; |
| 302 | + |
| 303 | + assert_eq!(a.max_abs_dev(&a), 0.); |
| 304 | + assert_eq!(a.max_abs_dev(&b), 4.); |
| 305 | + assert_eq!(b.max_abs_dev(&a), 4.); |
| 306 | + } |
| 307 | + |
| 308 | + #[test] |
| 309 | + fn test_mean_sq_dev() { |
| 310 | + let a = array![1., 1.]; |
| 311 | + let b = array![3., 5.]; |
| 312 | + |
| 313 | + assert_eq!(a.mean_sq_dev(&a), 0.); |
| 314 | + assert_eq!(a.mean_sq_dev(&b), 10.); |
| 315 | + assert_eq!(b.mean_sq_dev(&a), 10.); |
| 316 | + } |
| 317 | + |
| 318 | + #[test] |
| 319 | + fn test_root_mean_sq_dev() { |
| 320 | + let a = array![1., 1.]; |
| 321 | + let b = array![3., 5.]; |
| 322 | + |
| 323 | + assert_eq!(a.root_mean_sq_dev(&a), 0.); |
| 324 | + assert_abs_diff_eq!(a.root_mean_sq_dev(&b), 10.0.sqrt()); |
| 325 | + assert_abs_diff_eq!(b.root_mean_sq_dev(&a), 10.0.sqrt()); |
| 326 | + } |
| 327 | + |
| 328 | + #[test] |
| 329 | + fn test_peak_signal_to_noise_ratio() { |
| 330 | + let a = array![1., 1.]; |
| 331 | + assert!(a.peak_signal_to_noise_ratio(&a, 1.).is_infinite()); |
| 332 | + |
| 333 | + let a = array![1., 2., 3., 4., 5., 6., 7.]; |
| 334 | + let b = array![1., 3., 3., 4., 6., 7., 8.]; |
| 335 | + let maxv = 8.; |
| 336 | + let expected = 20. * Float::log10(maxv) - 10. * Float::log10(a.mean_sq_dev(&b)); |
| 337 | + let actual = a.peak_signal_to_noise_ratio(&b, maxv); |
| 338 | + |
| 339 | + assert_abs_diff_eq!(actual, expected); |
| 340 | + } |
| 341 | + |
| 342 | + #[test] |
| 343 | + fn test_deviations_with_n_by_m_ints() { |
| 344 | + let a = array![[0, 1], [4, 2]]; |
| 345 | + let b = array![[1, 1], [2, 4]]; |
| 346 | + |
| 347 | + assert_eq!(a.count_eq(&a), 4); |
| 348 | + assert_eq!(a.count_neq(&a), 0); |
| 349 | + assert_eq!(a.sq_l2_dist(&b), 9); |
| 350 | + assert_eq!(a.l2_dist(&b), 3.); |
| 351 | + assert_eq!(a.l1_dist(&b), 5); |
| 352 | + assert_eq!(a.linf_dist(&b), 2); |
| 353 | + |
| 354 | + assert_abs_diff_eq!(a.mean_abs_dev(&b), 1.25); |
| 355 | + assert_eq!(a.max_abs_dev(&b), 2); |
| 356 | + assert_abs_diff_eq!(a.mean_sq_dev(&b), 2.25); |
| 357 | + assert_abs_diff_eq!(a.root_mean_sq_dev(&b), 1.5); |
| 358 | + assert_abs_diff_eq!( |
| 359 | + a.peak_signal_to_noise_ratio(&b, 4), |
| 360 | + 8.519374645445623, |
| 361 | + epsilon = f64::EPSILON |
| 362 | + ); |
| 363 | + |
| 364 | + // TODO: gkl_div |
| 365 | + } |
| 366 | + |
| 367 | + #[test] |
| 368 | + fn test_deviations_with_empty_inputs() { |
| 369 | + let a: Array1<f64> = array![]; |
| 370 | + |
| 371 | + assert_eq!(a.count_eq(&a), 0); |
| 372 | + assert_eq!(a.count_neq(&a), 0); |
| 373 | + assert_eq!(a.sq_l2_dist(&a), 0.); |
| 374 | + assert_eq!(a.l2_dist(&a), 0.); |
| 375 | + assert_eq!(a.l1_dist(&a), 0.); |
| 376 | + assert_eq!(a.linf_dist(&a), 0.); |
| 377 | + |
| 378 | + assert!(a.mean_abs_dev(&a).is_nan()); |
| 379 | + assert_eq!(a.max_abs_dev(&a), 0.); |
| 380 | + assert!(a.mean_sq_dev(&a).is_nan()); |
| 381 | + assert!(a.root_mean_sq_dev(&a).is_nan()); |
| 382 | + assert!(a.peak_signal_to_noise_ratio(&a, 0.).is_nan()); |
| 383 | + |
| 384 | + // TODO: gkl_div |
| 385 | + } |
| 386 | +} |
0 commit comments