fastnn 0.3.3

A robust deep learning library in Rust with CUDA kernels
Documentation 
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

//! 2-D pooling over `[N, C, H, W]`. Each variant has its own gradient rule.

use std::ops::Range;

use crate::autograd::ops::pool::{AdaptiveAvgPool2dBackward, AvgPool2dBackward, MaxPool2dBackward};
use crate::tensor::Tensor;

use super::conv::Window;

impl Tensor {
    /// Largest value in each window. Saves the winning index for the backward.
    pub fn max_pool2d(&self, window: Window) -> Tensor {
        let (n, c, h, w) = self.expect_nchw("max_pool2d");
        let (out_h, out_w) = window.output_size(h, w);
        let data = self.to_vec();

        let mut out = vec![0.0f32; n * c * out_h * out_w];
        let mut winners = vec![0usize; out.len()];

        for plane in 0..n * c {
            for oh in 0..out_h {
                for ow in 0..out_w {
                    let (mut best, mut best_at) = (f32::NEG_INFINITY, plane * h * w);
                    for (ih, iw) in window.sources(oh, ow, h, w) {
                        let index = plane * h * w + ih * w + iw;
                        if data[index] > best {
                            best = data[index];
                            best_at = index;
                        }
                    }
                    let slot = (plane * out_h + oh) * out_w + ow;
                    out[slot] = best;
                    winners[slot] = best_at;
                }
            }
        }

        let shape = self.shape().to_vec();
        Tensor::from_vec(out, &[n, c, out_h, out_w])
            .to(self.device())
            .with_grad(&[self], || MaxPool2dBackward { shape, winners })
    }

    /// Average of each window, padding not counted.
    pub fn avg_pool2d(&self, window: Window) -> Tensor {
        let (n, c, h, w) = self.expect_nchw("avg_pool2d");
        let (out_h, out_w) = window.output_size(h, w);
        let data = self.to_vec();

        let mut out = vec![0.0f32; n * c * out_h * out_w];
        for plane in 0..n * c {
            for oh in 0..out_h {
                for ow in 0..out_w {
                    let mut total = 0.0f32;
                    let mut count = 0usize;
                    for (ih, iw) in window.sources(oh, ow, h, w) {
                        total += data[plane * h * w + ih * w + iw];
                        count += 1;
                    }
                    out[(plane * out_h + oh) * out_w + ow] = total / count as f32;
                }
            }
        }

        let shape = self.shape().to_vec();
        Tensor::from_vec(out, &[n, c, out_h, out_w])
            .to(self.device())
            .with_grad(&[self], || AvgPool2dBackward { shape, window })
    }

    /// Average into a fixed `(height, width)` output. `(1, 1)` is global pooling.
    pub fn adaptive_avg_pool2d(&self, output: (usize, usize)) -> Tensor {
        let (n, c, h, w) = self.expect_nchw("adaptive_avg_pool2d");
        let (out_h, out_w) = output;
        let data = self.to_vec();

        let mut out = vec![0.0f32; n * c * out_h * out_w];
        for plane in 0..n * c {
            for oh in 0..out_h {
                let rows = adaptive_range(oh, h, out_h);
                for ow in 0..out_w {
                    let cols = adaptive_range(ow, w, out_w);
                    let count = (rows.len() * cols.len()) as f32;
                    let mut total = 0.0f32;
                    for ih in rows.clone() {
                        for iw in cols.clone() {
                            total += data[plane * h * w + ih * w + iw];
                        }
                    }
                    out[(plane * out_h + oh) * out_w + ow] = total / count;
                }
            }
        }

        let shape = self.shape().to_vec();
        Tensor::from_vec(out, &[n, c, out_h, out_w])
            .to(self.device())
            .with_grad(&[self], || AdaptiveAvgPool2dBackward { shape, output })
    }

    fn expect_nchw(&self, op: &str) -> (usize, usize, usize, usize) {
        assert_eq!(
            self.ndim(),
            4,
            "{op} expects [N, C, H, W], got {:?}",
            self.shape()
        );
        (self.dim(0), self.dim(1), self.dim(2), self.dim(3))
    }
}

impl Window {
    /// Input positions feeding output `(oh, ow)`, skipping padding.
    pub(crate) fn sources(
        &self,
        oh: usize,
        ow: usize,
        h: usize,
        w: usize,
    ) -> impl Iterator<Item = (usize, usize)> + '_ {
        let (kh, kw) = self.kernel;
        (0..kh).flat_map(move |ki| {
            let ih = offset(oh, ki, self.stride.0, self.padding.0, h);
            (0..kw).filter_map(move |kj| {
                let iw = offset(ow, kj, self.stride.1, self.padding.1, w);
                match (ih, iw) {
                    (Some(y), Some(x)) => Some((y, x)),
                    _ => None,
                }
            })
        })
    }
}

fn offset(out: usize, k: usize, stride: usize, pad: usize, limit: usize) -> Option<usize> {
    let pos = (out * stride + k) as isize - pad as isize;
    (pos >= 0 && (pos as usize) < limit).then_some(pos as usize)
}

/// The input span that output index `out` averages over.
pub(crate) fn adaptive_range(out: usize, in_size: usize, out_size: usize) -> Range<usize> {
    (out * in_size / out_size)..((out + 1) * in_size).div_ceil(out_size).min(in_size)
}