1 | // Copyright 2018 Developers of the Rand project. |
2 | // |
3 | // Licensed under the Apache License, Version 2.0 <LICENSE-APACHE or |
4 | // https://www.apache.org/licenses/LICENSE-2.0> or the MIT license |
5 | // <LICENSE-MIT or https://opensource.org/licenses/MIT>, at your |
6 | // option. This file may not be copied, modified, or distributed |
7 | // except according to those terms. |
8 | |
9 | //! Weighted index sampling |
10 | |
11 | use crate::distributions::uniform::{SampleBorrow, SampleUniform, UniformSampler}; |
12 | use crate::distributions::Distribution; |
13 | use crate::Rng; |
14 | use core::cmp::PartialOrd; |
15 | use core::fmt; |
16 | |
17 | // Note that this whole module is only imported if feature="alloc" is enabled. |
18 | use alloc::vec::Vec; |
19 | |
20 | #[cfg (feature = "serde1" )] |
21 | use serde::{Serialize, Deserialize}; |
22 | |
23 | /// A distribution using weighted sampling of discrete items |
24 | /// |
25 | /// Sampling a `WeightedIndex` distribution returns the index of a randomly |
26 | /// selected element from the iterator used when the `WeightedIndex` was |
27 | /// created. The chance of a given element being picked is proportional to the |
28 | /// value of the element. The weights can use any type `X` for which an |
29 | /// implementation of [`Uniform<X>`] exists. |
30 | /// |
31 | /// # Performance |
32 | /// |
33 | /// Time complexity of sampling from `WeightedIndex` is `O(log N)` where |
34 | /// `N` is the number of weights. As an alternative, |
35 | /// [`rand_distr::weighted_alias`](https://docs.rs/rand_distr/*/rand_distr/weighted_alias/index.html) |
36 | /// supports `O(1)` sampling, but with much higher initialisation cost. |
37 | /// |
38 | /// A `WeightedIndex<X>` contains a `Vec<X>` and a [`Uniform<X>`] and so its |
39 | /// size is the sum of the size of those objects, possibly plus some alignment. |
40 | /// |
41 | /// Creating a `WeightedIndex<X>` will allocate enough space to hold `N - 1` |
42 | /// weights of type `X`, where `N` is the number of weights. However, since |
43 | /// `Vec` doesn't guarantee a particular growth strategy, additional memory |
44 | /// might be allocated but not used. Since the `WeightedIndex` object also |
45 | /// contains, this might cause additional allocations, though for primitive |
46 | /// types, [`Uniform<X>`] doesn't allocate any memory. |
47 | /// |
48 | /// Sampling from `WeightedIndex` will result in a single call to |
49 | /// `Uniform<X>::sample` (method of the [`Distribution`] trait), which typically |
50 | /// will request a single value from the underlying [`RngCore`], though the |
51 | /// exact number depends on the implementation of `Uniform<X>::sample`. |
52 | /// |
53 | /// # Example |
54 | /// |
55 | /// ``` |
56 | /// use rand::prelude::*; |
57 | /// use rand::distributions::WeightedIndex; |
58 | /// |
59 | /// let choices = ['a' , 'b' , 'c' ]; |
60 | /// let weights = [2, 1, 1]; |
61 | /// let dist = WeightedIndex::new(&weights).unwrap(); |
62 | /// let mut rng = thread_rng(); |
63 | /// for _ in 0..100 { |
64 | /// // 50% chance to print 'a', 25% chance to print 'b', 25% chance to print 'c' |
65 | /// println!("{}" , choices[dist.sample(&mut rng)]); |
66 | /// } |
67 | /// |
68 | /// let items = [('a' , 0), ('b' , 3), ('c' , 7)]; |
69 | /// let dist2 = WeightedIndex::new(items.iter().map(|item| item.1)).unwrap(); |
70 | /// for _ in 0..100 { |
71 | /// // 0% chance to print 'a', 30% chance to print 'b', 70% chance to print 'c' |
72 | /// println!("{}" , items[dist2.sample(&mut rng)].0); |
73 | /// } |
74 | /// ``` |
75 | /// |
76 | /// [`Uniform<X>`]: crate::distributions::Uniform |
77 | /// [`RngCore`]: crate::RngCore |
78 | #[derive(Debug, Clone, PartialEq)] |
79 | #[cfg_attr (feature = "serde1" , derive(Serialize, Deserialize))] |
80 | #[cfg_attr (doc_cfg, doc(cfg(feature = "alloc" )))] |
81 | pub struct WeightedIndex<X: SampleUniform + PartialOrd> { |
82 | cumulative_weights: Vec<X>, |
83 | total_weight: X, |
84 | weight_distribution: X::Sampler, |
85 | } |
86 | |
87 | impl<X: SampleUniform + PartialOrd> WeightedIndex<X> { |
88 | /// Creates a new a `WeightedIndex` [`Distribution`] using the values |
89 | /// in `weights`. The weights can use any type `X` for which an |
90 | /// implementation of [`Uniform<X>`] exists. |
91 | /// |
92 | /// Returns an error if the iterator is empty, if any weight is `< 0`, or |
93 | /// if its total value is 0. |
94 | /// |
95 | /// [`Uniform<X>`]: crate::distributions::uniform::Uniform |
96 | pub fn new<I>(weights: I) -> Result<WeightedIndex<X>, WeightedError> |
97 | where |
98 | I: IntoIterator, |
99 | I::Item: SampleBorrow<X>, |
100 | X: for<'a> ::core::ops::AddAssign<&'a X> + Clone + Default, |
101 | { |
102 | let mut iter = weights.into_iter(); |
103 | let mut total_weight: X = iter.next().ok_or(WeightedError::NoItem)?.borrow().clone(); |
104 | |
105 | let zero = <X as Default>::default(); |
106 | if !(total_weight >= zero) { |
107 | return Err(WeightedError::InvalidWeight); |
108 | } |
109 | |
110 | let mut weights = Vec::<X>::with_capacity(iter.size_hint().0); |
111 | for w in iter { |
112 | // Note that `!(w >= x)` is not equivalent to `w < x` for partially |
113 | // ordered types due to NaNs which are equal to nothing. |
114 | if !(w.borrow() >= &zero) { |
115 | return Err(WeightedError::InvalidWeight); |
116 | } |
117 | weights.push(total_weight.clone()); |
118 | total_weight += w.borrow(); |
119 | } |
120 | |
121 | if total_weight == zero { |
122 | return Err(WeightedError::AllWeightsZero); |
123 | } |
124 | let distr = X::Sampler::new(zero, total_weight.clone()); |
125 | |
126 | Ok(WeightedIndex { |
127 | cumulative_weights: weights, |
128 | total_weight, |
129 | weight_distribution: distr, |
130 | }) |
131 | } |
132 | |
133 | /// Update a subset of weights, without changing the number of weights. |
134 | /// |
135 | /// `new_weights` must be sorted by the index. |
136 | /// |
137 | /// Using this method instead of `new` might be more efficient if only a small number of |
138 | /// weights is modified. No allocations are performed, unless the weight type `X` uses |
139 | /// allocation internally. |
140 | /// |
141 | /// In case of error, `self` is not modified. |
142 | pub fn update_weights(&mut self, new_weights: &[(usize, &X)]) -> Result<(), WeightedError> |
143 | where X: for<'a> ::core::ops::AddAssign<&'a X> |
144 | + for<'a> ::core::ops::SubAssign<&'a X> |
145 | + Clone |
146 | + Default { |
147 | if new_weights.is_empty() { |
148 | return Ok(()); |
149 | } |
150 | |
151 | let zero = <X as Default>::default(); |
152 | |
153 | let mut total_weight = self.total_weight.clone(); |
154 | |
155 | // Check for errors first, so we don't modify `self` in case something |
156 | // goes wrong. |
157 | let mut prev_i = None; |
158 | for &(i, w) in new_weights { |
159 | if let Some(old_i) = prev_i { |
160 | if old_i >= i { |
161 | return Err(WeightedError::InvalidWeight); |
162 | } |
163 | } |
164 | if !(*w >= zero) { |
165 | return Err(WeightedError::InvalidWeight); |
166 | } |
167 | if i > self.cumulative_weights.len() { |
168 | return Err(WeightedError::TooMany); |
169 | } |
170 | |
171 | let mut old_w = if i < self.cumulative_weights.len() { |
172 | self.cumulative_weights[i].clone() |
173 | } else { |
174 | self.total_weight.clone() |
175 | }; |
176 | if i > 0 { |
177 | old_w -= &self.cumulative_weights[i - 1]; |
178 | } |
179 | |
180 | total_weight -= &old_w; |
181 | total_weight += w; |
182 | prev_i = Some(i); |
183 | } |
184 | if total_weight <= zero { |
185 | return Err(WeightedError::AllWeightsZero); |
186 | } |
187 | |
188 | // Update the weights. Because we checked all the preconditions in the |
189 | // previous loop, this should never panic. |
190 | let mut iter = new_weights.iter(); |
191 | |
192 | let mut prev_weight = zero.clone(); |
193 | let mut next_new_weight = iter.next(); |
194 | let &(first_new_index, _) = next_new_weight.unwrap(); |
195 | let mut cumulative_weight = if first_new_index > 0 { |
196 | self.cumulative_weights[first_new_index - 1].clone() |
197 | } else { |
198 | zero.clone() |
199 | }; |
200 | for i in first_new_index..self.cumulative_weights.len() { |
201 | match next_new_weight { |
202 | Some(&(j, w)) if i == j => { |
203 | cumulative_weight += w; |
204 | next_new_weight = iter.next(); |
205 | } |
206 | _ => { |
207 | let mut tmp = self.cumulative_weights[i].clone(); |
208 | tmp -= &prev_weight; // We know this is positive. |
209 | cumulative_weight += &tmp; |
210 | } |
211 | } |
212 | prev_weight = cumulative_weight.clone(); |
213 | core::mem::swap(&mut prev_weight, &mut self.cumulative_weights[i]); |
214 | } |
215 | |
216 | self.total_weight = total_weight; |
217 | self.weight_distribution = X::Sampler::new(zero, self.total_weight.clone()); |
218 | |
219 | Ok(()) |
220 | } |
221 | } |
222 | |
223 | impl<X> Distribution<usize> for WeightedIndex<X> |
224 | where X: SampleUniform + PartialOrd |
225 | { |
226 | fn sample<R: Rng + ?Sized>(&self, rng: &mut R) -> usize { |
227 | use ::core::cmp::Ordering; |
228 | let chosen_weight = self.weight_distribution.sample(rng); |
229 | // Find the first item which has a weight *higher* than the chosen weight. |
230 | self.cumulative_weights |
231 | .binary_search_by(|w| { |
232 | if *w <= chosen_weight { |
233 | Ordering::Less |
234 | } else { |
235 | Ordering::Greater |
236 | } |
237 | }) |
238 | .unwrap_err() |
239 | } |
240 | } |
241 | |
242 | #[cfg (test)] |
243 | mod test { |
244 | use super::*; |
245 | |
246 | #[cfg (feature = "serde1" )] |
247 | #[test] |
248 | fn test_weightedindex_serde1() { |
249 | let weighted_index = WeightedIndex::new(&[1, 2, 3, 4, 5, 6, 7, 8, 9, 10]).unwrap(); |
250 | |
251 | let ser_weighted_index = bincode::serialize(&weighted_index).unwrap(); |
252 | let de_weighted_index: WeightedIndex<i32> = |
253 | bincode::deserialize(&ser_weighted_index).unwrap(); |
254 | |
255 | assert_eq!( |
256 | de_weighted_index.cumulative_weights, |
257 | weighted_index.cumulative_weights |
258 | ); |
259 | assert_eq!(de_weighted_index.total_weight, weighted_index.total_weight); |
260 | } |
261 | |
262 | #[test] |
263 | fn test_accepting_nan(){ |
264 | assert_eq!( |
265 | WeightedIndex::new(&[core::f32::NAN, 0.5]).unwrap_err(), |
266 | WeightedError::InvalidWeight, |
267 | ); |
268 | assert_eq!( |
269 | WeightedIndex::new(&[core::f32::NAN]).unwrap_err(), |
270 | WeightedError::InvalidWeight, |
271 | ); |
272 | assert_eq!( |
273 | WeightedIndex::new(&[0.5, core::f32::NAN]).unwrap_err(), |
274 | WeightedError::InvalidWeight, |
275 | ); |
276 | |
277 | assert_eq!( |
278 | WeightedIndex::new(&[0.5, 7.0]) |
279 | .unwrap() |
280 | .update_weights(&[(0, &core::f32::NAN)]) |
281 | .unwrap_err(), |
282 | WeightedError::InvalidWeight, |
283 | ) |
284 | } |
285 | |
286 | |
287 | #[test] |
288 | #[cfg_attr (miri, ignore)] // Miri is too slow |
289 | fn test_weightedindex() { |
290 | let mut r = crate::test::rng(700); |
291 | const N_REPS: u32 = 5000; |
292 | let weights = [1u32, 2, 3, 0, 5, 6, 7, 1, 2, 3, 4, 5, 6, 7]; |
293 | let total_weight = weights.iter().sum::<u32>() as f32; |
294 | |
295 | let verify = |result: [i32; 14]| { |
296 | for (i, count) in result.iter().enumerate() { |
297 | let exp = (weights[i] * N_REPS) as f32 / total_weight; |
298 | let mut err = (*count as f32 - exp).abs(); |
299 | if err != 0.0 { |
300 | err /= exp; |
301 | } |
302 | assert!(err <= 0.25); |
303 | } |
304 | }; |
305 | |
306 | // WeightedIndex from vec |
307 | let mut chosen = [0i32; 14]; |
308 | let distr = WeightedIndex::new(weights.to_vec()).unwrap(); |
309 | for _ in 0..N_REPS { |
310 | chosen[distr.sample(&mut r)] += 1; |
311 | } |
312 | verify(chosen); |
313 | |
314 | // WeightedIndex from slice |
315 | chosen = [0i32; 14]; |
316 | let distr = WeightedIndex::new(&weights[..]).unwrap(); |
317 | for _ in 0..N_REPS { |
318 | chosen[distr.sample(&mut r)] += 1; |
319 | } |
320 | verify(chosen); |
321 | |
322 | // WeightedIndex from iterator |
323 | chosen = [0i32; 14]; |
324 | let distr = WeightedIndex::new(weights.iter()).unwrap(); |
325 | for _ in 0..N_REPS { |
326 | chosen[distr.sample(&mut r)] += 1; |
327 | } |
328 | verify(chosen); |
329 | |
330 | for _ in 0..5 { |
331 | assert_eq!(WeightedIndex::new(&[0, 1]).unwrap().sample(&mut r), 1); |
332 | assert_eq!(WeightedIndex::new(&[1, 0]).unwrap().sample(&mut r), 0); |
333 | assert_eq!( |
334 | WeightedIndex::new(&[0, 0, 0, 0, 10, 0]) |
335 | .unwrap() |
336 | .sample(&mut r), |
337 | 4 |
338 | ); |
339 | } |
340 | |
341 | assert_eq!( |
342 | WeightedIndex::new(&[10][0..0]).unwrap_err(), |
343 | WeightedError::NoItem |
344 | ); |
345 | assert_eq!( |
346 | WeightedIndex::new(&[0]).unwrap_err(), |
347 | WeightedError::AllWeightsZero |
348 | ); |
349 | assert_eq!( |
350 | WeightedIndex::new(&[10, 20, -1, 30]).unwrap_err(), |
351 | WeightedError::InvalidWeight |
352 | ); |
353 | assert_eq!( |
354 | WeightedIndex::new(&[-10, 20, 1, 30]).unwrap_err(), |
355 | WeightedError::InvalidWeight |
356 | ); |
357 | assert_eq!( |
358 | WeightedIndex::new(&[-10]).unwrap_err(), |
359 | WeightedError::InvalidWeight |
360 | ); |
361 | } |
362 | |
363 | #[test] |
364 | fn test_update_weights() { |
365 | let data = [ |
366 | ( |
367 | &[10u32, 2, 3, 4][..], |
368 | &[(1, &100), (2, &4)][..], // positive change |
369 | &[10, 100, 4, 4][..], |
370 | ), |
371 | ( |
372 | &[1u32, 2, 3, 0, 5, 6, 7, 1, 2, 3, 4, 5, 6, 7][..], |
373 | &[(2, &1), (5, &1), (13, &100)][..], // negative change and last element |
374 | &[1u32, 2, 1, 0, 5, 1, 7, 1, 2, 3, 4, 5, 6, 100][..], |
375 | ), |
376 | ]; |
377 | |
378 | for (weights, update, expected_weights) in data.iter() { |
379 | let total_weight = weights.iter().sum::<u32>(); |
380 | let mut distr = WeightedIndex::new(weights.to_vec()).unwrap(); |
381 | assert_eq!(distr.total_weight, total_weight); |
382 | |
383 | distr.update_weights(update).unwrap(); |
384 | let expected_total_weight = expected_weights.iter().sum::<u32>(); |
385 | let expected_distr = WeightedIndex::new(expected_weights.to_vec()).unwrap(); |
386 | assert_eq!(distr.total_weight, expected_total_weight); |
387 | assert_eq!(distr.total_weight, expected_distr.total_weight); |
388 | assert_eq!(distr.cumulative_weights, expected_distr.cumulative_weights); |
389 | } |
390 | } |
391 | |
392 | #[test] |
393 | fn value_stability() { |
394 | fn test_samples<X: SampleUniform + PartialOrd, I>( |
395 | weights: I, buf: &mut [usize], expected: &[usize], |
396 | ) where |
397 | I: IntoIterator, |
398 | I::Item: SampleBorrow<X>, |
399 | X: for<'a> ::core::ops::AddAssign<&'a X> + Clone + Default, |
400 | { |
401 | assert_eq!(buf.len(), expected.len()); |
402 | let distr = WeightedIndex::new(weights).unwrap(); |
403 | let mut rng = crate::test::rng(701); |
404 | for r in buf.iter_mut() { |
405 | *r = rng.sample(&distr); |
406 | } |
407 | assert_eq!(buf, expected); |
408 | } |
409 | |
410 | let mut buf = [0; 10]; |
411 | test_samples(&[1i32, 1, 1, 1, 1, 1, 1, 1, 1], &mut buf, &[ |
412 | 0, 6, 2, 6, 3, 4, 7, 8, 2, 5, |
413 | ]); |
414 | test_samples(&[0.7f32, 0.1, 0.1, 0.1], &mut buf, &[ |
415 | 0, 0, 0, 1, 0, 0, 2, 3, 0, 0, |
416 | ]); |
417 | test_samples(&[1.0f64, 0.999, 0.998, 0.997], &mut buf, &[ |
418 | 2, 2, 1, 3, 2, 1, 3, 3, 2, 1, |
419 | ]); |
420 | } |
421 | |
422 | #[test] |
423 | fn weighted_index_distributions_can_be_compared() { |
424 | assert_eq!(WeightedIndex::new(&[1, 2]), WeightedIndex::new(&[1, 2])); |
425 | } |
426 | } |
427 | |
428 | /// Error type returned from `WeightedIndex::new`. |
429 | #[cfg_attr (doc_cfg, doc(cfg(feature = "alloc" )))] |
430 | #[derive(Debug, Clone, Copy, PartialEq, Eq)] |
431 | pub enum WeightedError { |
432 | /// The provided weight collection contains no items. |
433 | NoItem, |
434 | |
435 | /// A weight is either less than zero, greater than the supported maximum, |
436 | /// NaN, or otherwise invalid. |
437 | InvalidWeight, |
438 | |
439 | /// All items in the provided weight collection are zero. |
440 | AllWeightsZero, |
441 | |
442 | /// Too many weights are provided (length greater than `u32::MAX`) |
443 | TooMany, |
444 | } |
445 | |
446 | #[cfg (feature = "std" )] |
447 | impl std::error::Error for WeightedError {} |
448 | |
449 | impl fmt::Display for WeightedError { |
450 | fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result { |
451 | f.write_str(match *self { |
452 | WeightedError::NoItem => "No weights provided in distribution" , |
453 | WeightedError::InvalidWeight => "A weight is invalid in distribution" , |
454 | WeightedError::AllWeightsZero => "All weights are zero in distribution" , |
455 | WeightedError::TooMany => "Too many weights (hit u32::MAX) in distribution" , |
456 | }) |
457 | } |
458 | } |
459 | |