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

use super::mha::MultiHeadAttention;
use crate::gradients::{CanUpdateWithGradients, GradientProvider, UnusedTensors};
use crate::prelude::*;

/// **Requires Nightly** A transformer encoder.
///
/// Generics
/// - `MODEL_DIM`: The size of query/key/value tensors. Given to [MultiHeadAttention].
/// - `NUM_HEADS`: The number of heads in [MultiHeadAttention].
/// - `FF_DIM`: The size of the hidden layer in
///   the feedforward network in [TransformerEncoderBlock].
/// - `NUM_LAYERS`: The number of [TransformerEncoderBlock] to use.
/// TODO: Doctests
pub type TransformerEncoder<
    const MODEL_DIM: usize,
    const NUM_HEADS: usize,
    const FF_DIM: usize,
    const NUM_LAYERS: usize,
> = Repeated<TransformerEncoderBlock<MODEL_DIM, NUM_HEADS, FF_DIM>, NUM_LAYERS>;

/// **Requires Nightly** A single transformer encoder block
///
/// Generics
/// - `MODEL_DIM`: The size of query/key/value tensors. Given to [MultiHeadAttention].
/// - `NUM_HEADS`: The number of heads in [MultiHeadAttention].
/// - `FF_DIM`: The size of the hidden layer in the feedforward network.
///
/// **Pytorch equivalent**:
/// ```python
/// encoder = torch.nn.TransformerEncoderLayer(
///    EMBED_DIM, NUM_HEADS, dim_feedforward=FF_DIM, batch_first=True, dropout=0.0
/// )
/// ```
/// TODO: Doctests
#[derive(Clone, Debug, Default)]
pub struct TransformerEncoderBlock<
    const MODEL_DIM: usize,
    const NUM_HEADS: usize,
    const FF_DIM: usize,
> {
    pub self_attn: MultiHeadAttention<MODEL_DIM, NUM_HEADS>,
    pub norm1: LayerNorm1D<MODEL_DIM>,
    pub ff: FF<MODEL_DIM, FF_DIM>,
    pub norm2: LayerNorm1D<MODEL_DIM>,
}

type FF<const M: usize, const F: usize> = Residual<(Linear<M, F>, ReLU, Linear<F, M>)>;

impl<const M: usize, const H: usize, const F: usize> ResetParams
    for TransformerEncoderBlock<M, H, F>
{
    fn reset_params<R: rand::Rng>(&mut self, rng: &mut R) {
        self.self_attn.reset_params(rng);
        self.norm1.reset_params(rng);
        self.ff.reset_params(rng);
        self.norm2.reset_params(rng);
    }
}

impl<const M: usize, const H: usize, const F: usize> CanUpdateWithGradients
    for TransformerEncoderBlock<M, H, F>
{
    fn update<G: GradientProvider>(&mut self, grads: &mut G, unused: &mut UnusedTensors) {
        self.self_attn.update(grads, unused);
        self.norm1.update(grads, unused);
        self.ff.update(grads, unused);
        self.norm2.update(grads, unused);
    }
}

impl<const M: usize, const H: usize, const F: usize, Src> Module<Src>
    for TransformerEncoderBlock<M, H, F>
where
    Src: Tensor<Dtype = f32>,
    MultiHeadAttention<M, H>: Module<(Src, Src::NoTape, Src::NoTape), Output = Src>,
    LayerNorm1D<M>: Module<Src, Output = Src>,
    FF<M, F>: Module<Src, Output = Src>,
{
    type Output = Src;

    fn forward(&self, src: Src) -> Self::Output {
        let (src, tape) = src.split_tape();
        let x = self
            .self_attn
            .forward((src.clone().put_tape(tape), src.clone(), src.clone()));
        let x = add(x, src);
        let x = self.norm1.forward(x);
        let x = self.ff.forward(x);
        self.norm2.forward(x)
    }
}

impl<const M: usize, const H: usize, const F: usize, T> ModuleMut<T>
    for TransformerEncoderBlock<M, H, F>
where
    Self: Module<T>,
{
    type Output = <Self as Module<T>>::Output;

    fn forward_mut(&mut self, t: T) -> Self::Output {
        self.forward(t)
    }
}

#[cfg(feature = "nightly")]
#[cfg(test)]
mod tests {
    use super::*;
    use crate::tests::assert_close;
    use rand::{rngs::StdRng, SeedableRng};

    #[test]
    fn test_encoder_block_forward() {
        let mut rng = StdRng::seed_from_u64(2);

        const BATCH: usize = 3;
        const SEQ_LEN: usize = 5;
        const EMBED_DIM: usize = 9;
        const NUM_HEADS: usize = 3;
        const FF_DIM: usize = 16;

        let mut encoder: TransformerEncoderBlock<EMBED_DIM, NUM_HEADS, FF_DIM> = Default::default();
        encoder.reset_params(&mut rng);

        let x: Tensor3D<BATCH, SEQ_LEN, EMBED_DIM> = TensorCreator::randn(&mut rng);
        let y: Tensor3D<BATCH, SEQ_LEN, EMBED_DIM> = encoder.forward(x);

        // This expected y was generated by:
        // 1. saving `encoder` parameters, `x` and `y` to a npz files
        // 2. Running pytorch with the same values
        // 3. printing out the output
        // See https://github.com/coreylowman/dfdx/wiki/Exporting-MultiHeadAttention-to-pytorch-for-unit-tests
        #[rustfmt::skip]
        assert_close(
            y.data(),
            &[
                [
                    [0.83316803, 0.85057360, 0.37431455, 1.48506296,-0.38405111,-1.89352179,-1.07049453,-0.50913972, 0.31408834],
                    [-0.57205188, 0.64078861,-0.56589824, 0.67155081, 0.65419787, 0.28409126,-1.75282931, 1.68111539,-1.04096484],
                    [-0.01414229, 1.34985816, 0.09684382, 0.13165890,-1.39875984,-1.61741352, 1.28747427, 0.75574619,-0.59126562],
                    [0.12542287, 2.60457349, 0.21064451,-0.81285846,-0.15861531,-0.87273139,-0.81707120,-0.17004849,-0.10931605],
                    [-1.54970682,-0.77183282, 1.37495196,-0.69562960,-0.66684282, 0.24720824, 1.38581741,-0.35962212, 1.03565681],
                ],
                [
                    [-0.15229249,-0.90768278,-0.85165489, 0.12768827, 1.61459768, 1.25826979,-0.46860829, 0.87496787,-1.49528503],
                    [-1.35595357, 1.13305736,-0.08542954, 1.01601434,-0.04678532,-1.69470263, 0.76144469,-0.68443829, 0.95679283],
                    [-1.49877191, 0.64559501, 0.33383703, 1.73698330,-0.14289393, 1.17869902,-1.01659226,-0.61038357,-0.62647283],
                    [0.78263682, 0.78481543,-0.16064386, 1.03396618, 1.49144781,-1.55002558,-1.11833119,-0.62120575,-0.64265978],
                    [-1.58957553, 1.75000548, 0.01272983, 0.11212827,-0.34744453,-1.45086825, 0.95842224, 0.50071126, 0.05389150],
                ],
                [
                    [-1.13160479,-0.21202824, 0.25907388,-0.64313424,-0.76302397,-0.16797650,-0.75345570, 2.01765633, 1.39449334],
                    [-0.16463053,-0.73241645,-0.69120175, 0.13771832, 0.72443259,-2.06525135, 1.02475107, 1.40244913, 0.36414924],
                    [0.38766465,-0.19543301,-1.80767059, 1.11545098, 0.21692322,-1.22834778, 0.13580292, 1.63094711,-0.25533777],
                    [1.22877085, 0.05472810, 0.65142977, 0.73869365,-0.74706972,-1.29277837, 1.07350135, 0.06228387,-1.76955938],
                    [-0.01733636,-1.57447529, 0.79691470, 1.00687420, 1.65637493,-0.75668150,-0.54616517, 0.45799020,-1.02349579],
                ],
            ],
        );
    }
}