1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
pub mod kernel;
use std::ptr;
use floaty::Floaty;
use num_cpus;
use thread_scoped as thread;
use univariate::Sample;
use self::kernel::Kernel;
pub struct Kde<'a, A, K> where A: 'a + Floaty, K: Kernel<A> {
bandwidth: A,
kernel: K,
sample: &'a Sample<A>,
}
impl<'a, A, K> Kde<'a, A, K> where A: 'a + Floaty, K: Kernel<A> {
pub fn new(sample: &'a Sample<A>, k: K, bw: Bandwidth<A>) -> Kde<'a, A, K> {
Kde {
bandwidth: bw.estimate(sample),
kernel: k,
sample: sample,
}
}
pub fn bandwidth(&self) -> A {
self.bandwidth
}
pub fn map(&self, xs: &[A]) -> Box<[A]> {
let n = xs.len();
let ncpus = num_cpus::get();
if ncpus > 1 && n > ncpus {
let granularity = n / ncpus + 1;
unsafe {
let mut ys = Vec::with_capacity(n);
ys.set_len(n);
{
let _ = ys.chunks_mut(granularity).enumerate().map(|(i, ys)| {
let offset = i * granularity;
thread::scoped(move || {
for (i, y) in ys.iter_mut().enumerate() {
ptr::write(y, (self)(*xs.get_unchecked(offset + i)))
}
})
}).collect::<Vec<_>>();
}
ys.into_boxed_slice()
}
} else {
xs.iter().map(|&x| (self)(x)).collect::<Vec<_>>().into_boxed_slice()
}
}
}
impl<'a, A, K> Fn<(A,)> for Kde<'a, A, K> where A: 'a + Floaty, K: Kernel<A> {
extern "rust-call" fn call(&self, (x,): (A,)) -> A {
let _0 = A::cast(0);
let slice = self.sample.as_slice();
let h = self.bandwidth;
let n = A::cast(slice.len());
let sum = slice.iter().fold(_0, |acc, &x_i| acc + (self.kernel)((x - x_i) / h));
sum / h / n
}
}
impl<'a, A, K> FnMut<(A,)> for Kde<'a, A, K> where A: 'a + Floaty, K: Kernel<A> {
extern "rust-call" fn call_mut(&mut self, args: (A,)) -> A {
self.call(args)
}
}
impl<'a, A, K> FnOnce<(A,)> for Kde<'a, A, K> where A: 'a + Floaty, K: Kernel<A> {
type Output = A;
extern "rust-call" fn call_once(self, args: (A,)) -> A {
self.call(args)
}
}
pub enum Bandwidth<A> where A: Floaty {
Manual(A),
Silverman,
}
impl<A> Bandwidth<A> where A: Floaty {
fn estimate(self, sample: &Sample<A>) -> A {
match self {
Bandwidth::Silverman => {
let factor = A::cast(4. / 3.);
let exponent = A::cast(1. / 5.);
let n = A::cast(sample.as_slice().len());
let sigma = sample.std_dev(None);
sigma * (factor / n).powf(exponent)
},
Bandwidth::Manual(bw) => bw,
}
}
}
#[cfg(test)]
macro_rules! test {
($ty:ident) => {
mod $ty {
use quickcheck::TestResult;
use univariate::Sample;
use univariate::kde::kernel::Gaussian;
use univariate::kde::{Bandwidth, Kde};
#[quickcheck]
fn integral(size: usize, start: usize) -> TestResult {
const DX: $ty = 1e-3;
if let Some(v) = ::test::vec::<$ty>(size, start) {
let slice = &v[start..];
let data = Sample::new(slice);
let kde = Kde::new(data, Gaussian, Bandwidth::Silverman);
let h = kde.bandwidth();
let (a, b) = (data.min() - 5. * h, data.max() + 5. * h);
let mut acc = 0.;
let mut x = a;
let mut y = kde(a);
while x < b {
acc += DX * y / 2.;
x += DX;
y = kde(x);
acc += DX * y / 2.;
}
TestResult::from_bool(relative_eq!(acc, 1., epsilon = 2e-5))
} else {
TestResult::discard()
}
}
}
}
}
#[cfg(test)]
mod test {
test!(f32);
test!(f64);
}
#[cfg(test)]
macro_rules! bench {
($ty:ident) => {
mod $ty {
use stdtest::Bencher;
use univariate::Sample;
use univariate::kde::kernel::Gaussian;
use univariate::kde::{Bandwidth, Kde};
const KDE_POINTS: usize = 100;
const SAMPLE_SIZE: usize = 100_000;
#[bench]
fn call(b: &mut Bencher) {
let data = ::bench::vec::<$ty>();
let kde = Kde::new(Sample::new(&data), Gaussian, Bandwidth::Silverman);
let x = Sample::new(&data).mean();
b.iter(|| {
kde(x)
})
}
#[bench]
fn map(b: &mut Bencher) {
let data = ::test::vec(SAMPLE_SIZE, 0).unwrap();
let kde = Kde::new(Sample::new(&data), Gaussian, Bandwidth::Silverman);
let xs: Vec<_> = ::itertools::linspace::<$ty>(0., 1., KDE_POINTS).collect();
b.iter(|| {
kde.map(&xs)
})
}
}
}
}
#[cfg(test)]
mod bench {
bench!(f32);
bench!(f64);
}