boostr 0.1.0

ML framework built on numr - attention, quantization, model architectures
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

//! Group Normalization module
//!
//! GroupNorm: normalizes over groups of channels independently.
//! Delegates to numr's `var_group_norm` for forward + autograd.

use crate::error::{Error, Result};
use numr::autograd::{Var, var_group_norm};
use numr::ops::{NormalizationOps, ScalarOps, TensorOps};
use numr::runtime::{Runtime, RuntimeClient};
use numr::tensor::Tensor;

/// Group Normalization
///
/// Divides channels into `num_groups` groups and normalizes each group
/// independently. Input shape: `[batch, channels, *spatial]`.
///
/// weight (gamma): `[channels]`
/// bias (beta): `[channels]`
pub struct GroupNorm<R: Runtime> {
    weight: Var<R>,
    bias: Var<R>,
    num_groups: usize,
    eps: f32,
}

impl<R: Runtime> GroupNorm<R> {
    pub fn new(
        weight: Tensor<R>,
        bias: Tensor<R>,
        num_groups: usize,
        eps: f32,
        trainable: bool,
    ) -> Self {
        Self {
            weight: Var::new(weight, trainable),
            bias: Var::new(bias, trainable),
            num_groups,
            eps,
        }
    }

    /// Forward pass with autograd.
    ///
    /// Input: `[batch, channels, *spatial]`
    /// Output: same shape, normalized per group.
    pub fn forward<C>(&self, client: &C, input: &Var<R>) -> Result<Var<R>>
    where
        R: Runtime,
        C: RuntimeClient<R> + NormalizationOps<R>,
        R::Client: TensorOps<R> + ScalarOps<R>,
    {
        var_group_norm(
            input,
            &self.weight,
            &self.bias,
            self.num_groups,
            self.eps,
            client,
        )
        .map_err(Error::Numr)
    }

    pub fn weight(&self) -> &Var<R> {
        &self.weight
    }

    pub fn bias(&self) -> &Var<R> {
        &self.bias
    }

    pub fn num_groups(&self) -> usize {
        self.num_groups
    }
}

#[cfg(test)]
mod tests {
    use super::*;
    use crate::test_utils::cpu_setup;
    use numr::autograd::backward;
    use numr::runtime::cpu::CpuRuntime;

    #[test]
    fn test_groupnorm_output_shape() {
        let (client, device) = cpu_setup();
        let weight = Tensor::<CpuRuntime>::from_slice(&[1.0f32; 4], &[4], &device);
        let bias = Tensor::<CpuRuntime>::from_slice(&[0.0f32; 4], &[4], &device);
        let norm = GroupNorm::new(weight, bias, 2, 1e-5, false);

        let input = Var::new(
            Tensor::<CpuRuntime>::from_slice(&[1.0f32; 24], &[2, 4, 3], &device),
            false,
        );
        let out = norm.forward(&client, &input).unwrap();
        assert_eq!(out.tensor().shape(), &[2, 4, 3]);
    }

    #[test]
    fn test_groupnorm_zero_mean_per_group() {
        let (client, device) = cpu_setup();
        let weight = Tensor::<CpuRuntime>::from_slice(&[1.0f32; 4], &[4], &device);
        let bias = Tensor::<CpuRuntime>::from_slice(&[0.0f32; 4], &[4], &device);
        let norm = GroupNorm::new(weight, bias, 2, 1e-5, false);

        let input = Var::new(
            Tensor::<CpuRuntime>::from_slice(
                &[
                    1.0f32, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0, 9.0, 10.0, 11.0, 12.0,
                ],
                &[1, 4, 3],
                &device,
            ),
            false,
        );
        let out = norm.forward(&client, &input).unwrap();
        let data: Vec<f32> = out.tensor().to_vec();

        // Group 0 (ch 0,1) should have mean ~0
        let g0: f32 = data[0..6].iter().sum();
        assert!(g0.abs() < 1e-4, "group 0 mean should be ~0, sum={g0}");
    }

    #[test]
    fn test_groupnorm_backward() {
        let (client, device) = cpu_setup();
        let weight = Tensor::<CpuRuntime>::from_slice(&[1.0f32; 4], &[4], &device);
        let bias = Tensor::<CpuRuntime>::from_slice(&[0.0f32; 4], &[4], &device);
        let norm = GroupNorm::new(weight, bias, 2, 1e-5, true);

        let input = Var::new(
            Tensor::<CpuRuntime>::from_slice(
                &[1.0f32, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0],
                &[1, 4, 2],
                &device,
            ),
            true,
        );
        let out = norm.forward(&client, &input).unwrap();
        let loss = numr::autograd::var_sum(&out, &[], false, &client).unwrap();
        let grads = backward(&loss, &client).unwrap();

        let d_input: Vec<f32> = grads.get(input.id()).unwrap().to_vec();
        assert_eq!(d_input.len(), 8);
        for v in &d_input {
            assert!(v.is_finite());
        }
    }
}