yarnn 0.1.0

Yet Another rust Neural Network framework
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

use crate::tensor::{Tensor, TensorShape};
use crate::layer::Layer;
use crate::params::Params;
use crate::backend::{Backend, Conv2dInfo, PaddingKind, BackendBias, BackendConv2d, BackendScale};
use crate::optimizer::{Optimizable, Optimizer};

pub struct Conv2dConfig {
    pub filters: u32,
    pub kernel: (u32, u32),
    pub strides: (u32, u32),
    pub padding: PaddingKind,
    pub biases: bool,
}

impl Default for Conv2dConfig {
    fn default() -> Self {
        Self {
            filters: 1,
            kernel: (3, 3),
            strides: (1, 1),
            padding: PaddingKind::Valid,
            biases: false,
        }
    }
}

pub struct Conv2d<N, B, O> 
    where B: Backend<N>,
          O: Optimizer<N, B>
{
    input_shape: TensorShape,
    units: u32,
    conv_info: Conv2dInfo,
    use_biases: bool,
    filters: Params<N, B, O>,
    biases: Params<N, B, O>,
}

impl <N, B, O> Layer<N, B> for Conv2d<N, B, O> 
    where B: Backend<N> + BackendConv2d<N> + BackendBias<N>,
          O: Optimizer<N, B>
{
    type Config = Conv2dConfig;

    fn name(&self) -> &str {
        "Conv2d"
    }
    
    fn create(input_shape: TensorShape, cfg: Self::Config) -> Self {
        assert!(input_shape.dims == 3);

        Conv2d {
            input_shape,
            units: cfg.filters,
            conv_info: Conv2dInfo {
                kernel: cfg.kernel,
                padding: cfg.padding, 
                strides: cfg.strides,
            },
            use_biases: cfg.biases,
            filters: Params::new((cfg.filters, cfg.kernel.0, cfg.kernel.1)),
            biases: Params::new((cfg.filters, )),
        }
    }
    
    fn init(&mut self, backend: &B) {
        self.filters.init_random(backend, self.conv_info.kernel.0 * self.conv_info.kernel.1 + self.filters.params.shape().get(0));

        if self.use_biases {
            self.biases.init_zero(backend);
        }
    }

    #[inline]
    fn input_shape(&self) -> TensorShape {
        self.input_shape.clone()
    }

    #[inline]
    fn output_shape(&self) -> TensorShape {
        let is = self.input_shape.as_slice();

        // O = (W - K + 2P) / S + 1

        let rows = (is[1] - self.conv_info.kernel.0) / self.conv_info.strides.0 + 1;
        let cols = (is[2] - self.conv_info.kernel.1) / self.conv_info.strides.1 + 1;

        TensorShape::new3d(
            self.units,
            rows,
            cols,
        )
    }
    
    #[inline]
    fn forward(&self, backend: &B, y: &mut B::Tensor, x: &B::Tensor) {
        assert_eq!(y.shape().dims, 4);
        assert_eq!(x.shape().dims, 4);

        backend.conv2d_forward(y, x, &self.filters.params, &self.conv_info);

        if self.use_biases {
            unimplemented!();
            // backend.bias_add(y, &self.biases.params);
        }
    }

    #[inline]
    fn backward(&self, backend: &B, dx: &mut B::Tensor, dy: &B::Tensor, _: &B::Tensor, _: &B::Tensor) {
        assert_eq!(dy.shape().dims, 4);
        assert_eq!(dx.shape().dims, 4);

        backend.conv2d_backward_input(dx, dy, &self.filters.params, &self.conv_info);
    }
}

impl <N, B, O> Optimizable<N, B, O> for Conv2d<N, B, O>
    where B: Backend<N> + BackendConv2d<N> + BackendBias<N> + BackendScale<N>,
          O: Optimizer<N, B>
{
    #[inline]
    fn calc_gradients(&mut self, backend: &B, x: &B::Tensor, dy: &B::Tensor) {
        assert_eq!(dy.shape().dims, 4);
        assert_eq!(x.shape().dims, 4);

        backend.conv2d_backward_filter(&mut self.filters.grads, x, dy, &self.conv_info);
        let prescaler = 1.0 / x.shape().get(0) as f32;

        backend.scale(&mut self.filters.grads, backend.scalar_f32(prescaler));
    }

    #[inline]
    fn optimize(&mut self, backend: &B, optimizer: &O) {
        optimizer.update_params(backend, &mut self.filters.ctx, &mut self.filters.params, &mut self.filters.grads);

        if self.use_biases {
            unimplemented!()
        //     optimizer.update_params(backend, &mut self.biases.ctx, &mut self.biases.params, &self.biases.grads);
        }
    }
}