ndarray-conv 0.6.1

N-Dimension convolution (with FFT) lib for ndarray.
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
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

use criterion::{criterion_group, criterion_main, Criterion};

use ndarray::prelude::*;
use ndarray_conv::*;
use ndarray_rand::{rand_distr::Uniform, RandomExt};
use ndarray_vision::processing::ConvolutionExt;
use num::Complex;

/// Benchmark for 1D convolution using `conv_fft` with various libraries.
fn criterion_benchmark(c: &mut Criterion) {
    let x = Array::random(5000, Uniform::new(0f32, 1.).unwrap());
    let k = Array::random(31, Uniform::new(0f32, 1.).unwrap());

    let x_crs = x.to_shape((1, 1, 5000)).unwrap().to_owned();
    let k_crs = k.to_shape((1, 1, 1, 31)).unwrap().to_owned();

    let tensor = tch::Tensor::from_slice(x.as_slice().unwrap())
        .to_dtype(tch::Kind::Float, false, true)
        .reshape([1, 1, 5000]);
    let kernel = tch::Tensor::from_slice(k.as_slice().unwrap())
        .to_dtype(tch::Kind::Float, false, true)
        .reshape([1, 1, 31]);

    // for (a, b) in x
    //     .conv_fft(&k, ConvMode::Same, PaddingMode::Zeros)
    //     .unwrap()
    //     .iter()
    //     .zip(
    //         tensor
    //             .conv1d_padding::<tch::Tensor>(&kernel, None, 1, "same", 1, 1)
    //             .reshape(5000)
    //             .iter::<f64>()
    //             .unwrap(),
    //     )
    // {
    //     // need to div kernel size
    //     assert!((*a as f64 - b).abs() < 1e-5);
    // }

    let mut fft_processor = get_fft_processor();

    /// Benchmark for 1D convolution using `conv_fft`.
    c.bench_function("fft_1d", |b| {
        b.iter(|| x.conv_fft(&k, ConvMode::Same, PaddingMode::Zeros))
    });

    /// Benchmark for 1D convolution using `conv_fft_with_processor`.
    c.bench_function("fft_with_processor_1d", |b| {
        b.iter(|| {
            x.conv_fft_with_processor(&k, ConvMode::Same, PaddingMode::Zeros, &mut fft_processor)
        })
    });

    c.bench_function("torch_1d", |b| {
        b.iter(|| tensor.conv1d_padding::<tch::Tensor>(&kernel, None, 1, "same", 1, 1))
    });

    // c.bench_function("convolution_rs_1d", |b| {
    //     b.iter(|| {
    //         convolutions_rs::convolutions::ConvolutionLayer::new_tf(
    //             k_crs.clone(),
    //             None,
    //             1,
    //             convolutions_rs::Padding::Same,
    //         )
    //         .convolve(&x_crs)
    //     });
    // });

    // c.bench_function("fftconvolve_1d", |b| {
    //     b.iter(|| fftconvolve::fftconvolve(&x, &k, fftconvolve::Mode::Same))
    // });

    let x = Array::random((200, 5000), Uniform::new(0f32, 1.).unwrap());
    let k = Array::random((11, 31), Uniform::new(0f32, 1.).unwrap());

    let x_crs = x.to_shape((1, 200, 5000)).unwrap().to_owned();
    let k_crs = k.to_shape((1, 1, 11, 31)).unwrap().to_owned();

    let x_nvs = x.to_shape((200, 5000, 1)).unwrap().to_owned();
    let k_nvs = k.to_shape((11, 31, 1)).unwrap().to_owned();

    let tensor = tch::Tensor::from_slice(x.as_slice().unwrap())
        .to_dtype(tch::Kind::Float, false, true)
        .reshape([1, 1, 200, 5000]);
    let kernel = tch::Tensor::from_slice(k.as_slice().unwrap())
        .to_dtype(tch::Kind::Float, false, true)
        .reshape([1, 1, 11, 31]);

    let mut fft_processor = get_fft_processor();

    /// Benchmark for 2D convolution using `conv_fft`.
    c.bench_function("fft_2d", |b| {
        b.iter(|| x.conv_fft(&k, ConvMode::Same, PaddingMode::Zeros))
    });

    /// Benchmark for 2D convolution using `conv_fft_with_processor`.
    c.bench_function("fft_with_processor_2d", |b| {
        b.iter(|| {
            x.conv_fft_with_processor(&k, ConvMode::Same, PaddingMode::Zeros, &mut fft_processor)
        })
    });

    c.bench_function("torch_2d", |b| {
        b.iter(|| tensor.conv2d_padding::<tch::Tensor>(&kernel, None, 1, "same", 1, 1))
    });

    // c.bench_function("ndarray_vision_2d", |b| {
    //     b.iter(|| x_nvs.conv2d_with_padding(k_nvs.clone(), &ndarray_vision::core::ZeroPadding))
    // });

    // c.bench_function("convolution_rs_2d", |b| {
    //     b.iter(|| {
    //         convolutions_rs::convolutions::ConvolutionLayer::new_tf(
    //             k_crs.clone(),
    //             None,
    //             1,
    //             convolutions_rs::Padding::Same,
    //         )
    //         .convolve(&x_crs)
    //     });
    // });

    // c.bench_function("fftconvolve_2d", |b| {
    //     b.iter(|| fftconvolve::fftconvolve(&x, &k, fftconvolve::Mode::Same))
    // });

    let x = Array::random((10, 100, 200), Uniform::new(0f32, 1.).unwrap());
    let k = Array::random((5, 11, 31), Uniform::new(0f32, 1.).unwrap());

    let x_crs = x.to_shape((10, 100, 200)).unwrap().to_owned();
    let k_crs = k.to_shape((1, 5, 11, 31)).unwrap().to_owned();

    let tensor = tch::Tensor::from_slice(x.as_slice().unwrap())
        .to_dtype(tch::Kind::Float, false, true)
        .reshape([1, 1, 10, 100, 200]);
    let kernel = tch::Tensor::from_slice(k.as_slice().unwrap())
        .to_dtype(tch::Kind::Float, false, true)
        .reshape([1, 1, 5, 11, 31]);

    let mut fft_processor = get_fft_processor();

    /// Benchmark for 3D convolution using `conv_fft`.
    c.bench_function("fft_3d", |b| {
        b.iter(|| x.conv_fft(&k, ConvMode::Same, PaddingMode::Zeros))
    });

    /// Benchmark for 3D convolution using `conv_fft_with_processor`.
    c.bench_function("fft_with_processor_3d", |b| {
        b.iter(|| {
            x.conv_fft_with_processor(&k, ConvMode::Same, PaddingMode::Zeros, &mut fft_processor)
        })
    });

    c.bench_function("torch_3d", |b| {
        b.iter(|| tensor.conv3d_padding::<tch::Tensor>(&kernel, None, 1, "same", 1, 1))
    });

    // c.bench_function("convolution_rs_3d", |b| {
    //     b.iter(|| {
    //         convolutions_rs::convolutions::ConvolutionLayer::new_tf(
    //             k_crs.clone(),
    //             None,
    //             1,
    //             convolutions_rs::Padding::Same,
    //         )
    //         .convolve(&x_crs)
    //     });
    // });

    // c.bench_function("fftconvolve_3d", |b| {
    //     b.iter(|| fftconvolve::fftconvolve(&x, &k, fftconvolve::Mode::Same))
    // });
}

criterion_group!(benches, criterion_benchmark);
criterion_main!(benches);