ruda-tensor 0.21.10

Ruda tensor operations, backend contracts, and framework-facing primitives.
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

use super::{ConvOptions, UnfoldOptions};
use crate::tensor::FloatTensor;
use crate::{Backend, TensorData, TensorMetadata, element::ElementConversion};
use alloc::vec;
use alloc::vec::Vec;
use ruda_core::tensor::{DType, Shape};

/// Constructs a special weight tensor used for unfolding.
///
/// # Notes
///
/// The idea behind using convolution for unfolding is to leverage the sliding window mechanism of
/// convolution. By creating a weight tensor with ones in a particular pattern, we are able to borrow
/// the convolution operation's mechanism as it moves across the input tensor, picking up the desired
/// values in the pattern of the unfolding operation.
pub(crate) fn create_unfolding_weight<B: Backend>(
    in_channels: usize,
    kernel_size: [usize; 2],
    device: &B::Device,
    dtype: DType,
) -> FloatTensor<B> {
    let shape = Shape::new([
        in_channels * kernel_size[0] * kernel_size[1],
        in_channels,
        kernel_size[0],
        kernel_size[1],
    ]);

    let mut strides = [0; 4];
    let mut current = 1;
    shape.iter().enumerate().rev().for_each(|(index, val)| {
        strides[index] = current;
        current *= val;
    });

    let num_elements = shape.num_elements();

    let mut weight: Vec<B::FloatElem> = vec![0.0.elem(); num_elements];

    for k in 0..in_channels {
        for i in 0..kernel_size[0] {
            for j in 0..kernel_size[1] {
                let output_channel = k * kernel_size[0] * kernel_size[1] + i * kernel_size[1] + j;
                let index =
                    output_channel * strides[0] + k * strides[1] + i * strides[2] + j * strides[3];

                weight[index] = 1.elem();
            }
        }
    }

    B::float_from_data(TensorData::new(weight, shape).convert_dtype(dtype), device)
}

/// Compute the unfold4d operation using the conv2d operations.
pub(crate) fn unfold4d_using_conv2d<B: Backend>(
    x: FloatTensor<B>,
    kernel_size: [usize; 2],
    options: UnfoldOptions,
) -> FloatTensor<B> {
    let [_batch_size, in_channels, _in_height, _in_width] = x.shape().dims();
    let weight =
        create_unfolding_weight::<B>(in_channels, kernel_size, &B::float_device(&x), x.dtype());
    let unfolded = B::conv2d(
        x,
        weight,
        None,
        ConvOptions::new(options.stride, options.padding, options.dilation, 1),
    );

    let [batch_size, channels_out, out_height, out_width] = unfolded.shape().dims();

    B::float_reshape(
        unfolded,
        Shape::new([batch_size, channels_out, out_height * out_width]),
    )
}

pub use ruda_core::tensor::spatial::{calculate_unfold_windows, calculate_unfold_shape};

#[cfg(test)]
mod tests {
    use super::*;

    #[test]
    fn test_calculate_unfold_windows() {
        assert_eq!(calculate_unfold_windows(2, 5, 1), 0);

        assert_eq!(calculate_unfold_windows(2, 3, 1), 0);
        assert_eq!(calculate_unfold_windows(3, 3, 1), 1);
        assert_eq!(calculate_unfold_windows(4, 3, 1), 2);
        assert_eq!(calculate_unfold_windows(5, 3, 1), 3);

        assert_eq!(calculate_unfold_windows(2, 3, 2), 0);
        assert_eq!(calculate_unfold_windows(3, 3, 2), 1);
        assert_eq!(calculate_unfold_windows(4, 3, 2), 1);
        assert_eq!(calculate_unfold_windows(5, 3, 2), 2);
    }

    #[test]
    fn test_calculate_unfold_shape() {
        assert_eq!(
            calculate_unfold_shape([2, 6, 6], 1, 3, 2),
            Shape::new([2, 2, 6, 3])
        );
    }
}