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
178
179
180
181
182
183
184
185
use crate::{shapes::*, tensor::*, tensor_ops::*};

use super::*;

pub mod builder {
    #[derive(Debug)]
    pub struct LayerNorm1D<const M: usize>;
}
impl<const M: usize, E: Dtype, D: Device<E>> BuildOnDevice<D, E> for builder::LayerNorm1D<M>
where
    LayerNorm1D<M, E, D>: BuildModule<D, E>,
{
    type Built = LayerNorm1D<M, E, D>;
    fn try_build_on_device(device: &D) -> Result<Self::Built, D::Err> {
        Self::Built::try_build(device)
    }
}

/// Implements layer normalization as described in [Layer Normalization](https://arxiv.org/abs/1607.06450).
///
/// This calls [normalize()] on the last axis of the input to normalize to 0 mean and unit std dev, and then does an element-wise
/// affine transform using learnable parameters [Self::gamma] and [Self::beta].
///
/// [Self::epsilon] is passed to [normalize()] and added to the variance to ensure big enough numbers. It defaults to `1e-5`.
///
/// # Generics
/// - `M` The size of the affine transform tensors.
///
/// # Examples
/// ```rust
/// # use dfdx::prelude::*;
/// # let dev: Cpu = Default::default();
/// type Model = LayerNorm1D<5>;
/// let model = dev.build_module::<Model, f32>();
/// let _: Tensor<Rank1<5>, f32, _> = model.forward(dev.zeros::<Rank1<5>>());
/// ```

#[derive(Debug, Clone)]
pub struct LayerNorm1D<const M: usize, E: Dtype, D: Storage<E>> {
    pub gamma: Tensor<Rank1<M>, E, D>,
    pub beta: Tensor<Rank1<M>, E, D>,
    pub epsilon: f64,
}

impl<const M: usize, E: Dtype, D: Storage<E>> NonMutableModule for LayerNorm1D<M, E, D> {}

impl<const M: usize, E: Dtype, D: Device<E>> TensorCollection<E, D> for LayerNorm1D<M, E, D> {
    type To<E2: Dtype, D2: Device<E2>> = LayerNorm1D<M, E2, D2>;

    fn iter_tensors<V: ModuleVisitor<Self, E, D>>(
        visitor: &mut V,
    ) -> Result<Option<Self::To<V::E2, V::D2>>, V::Err> {
        visitor.visit_fields(
            (
                Self::tensor(
                    "gamma",
                    |s| &s.gamma,
                    |s| &mut s.gamma,
                    TensorOptions::reset_to_ones(),
                ),
                Self::tensor(
                    "beta",
                    |s| &s.beta,
                    |s| &mut s.beta,
                    TensorOptions::reset_to_zeros(),
                ),
                Self::scalar(
                    "epsilon",
                    |s| &s.epsilon,
                    |s| &mut s.epsilon,
                    ScalarOptions::from_default(1e-5),
                ),
            ),
            |(gamma, beta, epsilon)| LayerNorm1D {
                gamma,
                beta,
                epsilon,
            },
        )
    }
}

impl<const M: usize, E: Dtype, D: Device<E>, T: Tape<E, D>> Module<Tensor<Rank1<M>, E, D, T>>
    for LayerNorm1D<M, E, D>
{
    type Output = Tensor<Rank1<M>, E, D, T>;
    type Error = D::Err;

    fn try_forward(&self, x: Tensor<Rank1<M>, E, D, T>) -> Result<Self::Output, D::Err> {
        x.try_normalize(self.epsilon)?
            .try_mul(self.gamma.clone())?
            .try_add(self.beta.clone())
    }
}

impl<B: Dim, const M: usize, E: Dtype, D: Device<E>, T: Tape<E, D>>
    Module<Tensor<(B, Const<M>), E, D, T>> for LayerNorm1D<M, E, D>
{
    type Output = Tensor<(B, Const<M>), E, D, T>;
    type Error = D::Err;

    fn try_forward(&self, x: Tensor<(B, Const<M>), E, D, T>) -> Result<Self::Output, D::Err> {
        let shape = *x.shape();
        x.try_normalize::<Axis<1>>(self.epsilon)?
            .try_mul(self.gamma.retaped::<T>().try_broadcast_like(&shape)?)?
            .try_add(self.beta.retaped::<T>().try_broadcast_like(&shape)?)
    }
}

impl<B: Dim, S: Dim, const M: usize, E: Dtype, D: Device<E>, T: Tape<E, D>>
    Module<Tensor<(B, S, Const<M>), E, D, T>> for LayerNorm1D<M, E, D>
{
    type Output = Tensor<(B, S, Const<M>), E, D, T>;
    type Error = D::Err;

    fn try_forward(&self, x: Tensor<(B, S, Const<M>), E, D, T>) -> Result<Self::Output, D::Err> {
        let shape = *x.shape();
        x.try_normalize::<Axis<2>>(self.epsilon)?
            .try_mul(self.gamma.retaped::<T>().try_broadcast_like(&shape)?)?
            .try_add(self.beta.retaped::<T>().try_broadcast_like(&shape)?)
    }
}

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

    #[test]
    fn test_layer_norm_reset() {
        let dev: TestDevice = Default::default();

        let mut m = dev.build_module::<builder::LayerNorm1D<5>, TestDtype>();
        assert_close_to_literal!(m.gamma, [1.0; 5]);
        assert_close_to_literal!(m.beta, [0.0; 5]);

        m.gamma = dev.sample_normal();
        m.beta = dev.sample_normal();

        assert_ne!(m.gamma.array(), [TestDtype::ONE; 5]);
        assert_ne!(m.beta.array(), [TestDtype::default(); 5]);

        m.reset_params();

        assert_close_to_literal!(m.gamma, [1.0; 5]);
        assert_close_to_literal!(m.beta, [0.0; 5]);
    }

    #[test]
    fn test_layer_norm_1d_forward() {
        let dev: TestDevice = Default::default();
        let mut m = dev.build_module::<builder::LayerNorm1D<5>, TestDtype>();
        let x = dev.sample_normal::<Rank1<5>>();
        let r = m.forward_mut(x.leaky_trace());
        assert_close_to_literal!(r, [0.873304, 0.9879816, -1.6083492, 0.44028836, -0.6932247]);
        let g = r.mean().backward();
        assert_close_to_literal!(
            g.get(&m.gamma),
            [0.1746608, 0.19759633, -0.32166985, 0.088057674, -0.13864495]
        );
        assert_close_to_literal!(g.get(&m.beta), [0.2; 5]);
    }

    #[test]
    fn test_layer_norm_2d_forward() {
        let dev: TestDevice = Default::default();
        let m = dev.build_module::<builder::LayerNorm1D<5>, TestDtype>();
        let x = dev.sample_normal::<Rank2<3, 5>>();
        let r = m.forward(x.leaky_trace());
        assert_close_to_literal!(
            r,
            [
                [0.873304, 0.9879816, -1.6083492, 0.44028836, -0.6932247],
                [0.663322, -1.8449169, 0.05217871, 0.056903206, 1.0725129],
                [1.0343355, -1.5559655, -0.40086073, 1.1405537, -0.21806297],
            ]
        );
        let g = r.mean().backward();
        assert_close_to_literal!(
            g.get(&m.gamma),
            [0.1713974, -0.16086, -0.1304687, 0.109183, 0.0107483]
        );
        assert_close_to_literal!(g.get(&m.beta), [0.2; 5]);
    }
}