luminal 0.2.0

Deep learning at the speed of light.
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 rand::{rngs::StdRng, SeedableRng};

use crate::{
    nn::{activation::ReLU, linear::Linear, transformer::Transformer},
    prelude::*,
};

use super::random_vec_rng;

pub fn matmul() -> (Graph, Vec<GraphTensor<()>>) {
    let mut rng = StdRng::seed_from_u64(0);
    let mut cx = Graph::new();
    let a = cx
        .tensor::<(Dyn<'a'>, Const<3>)>()
        .set_dyn(random_vec_rng(2 * 3, &mut rng), &[2, 3]);
    let b = cx.tensor::<R2<3, 3>>().set(random_vec_rng(3 * 3, &mut rng));
    let c = a.matmul(b).retrieve();
    (cx, vec![c.no_shape()])
}

pub fn batch_matmul() -> (Graph, Vec<GraphTensor<()>>) {
    let mut rng = StdRng::seed_from_u64(0);
    let mut cx = Graph::new();
    let a = cx
        .tensor::<(Dyn<'a'>, Dyn<'b'>, Const<2>)>()
        .set_dyn(random_vec_rng(2 * 3 * 2, &mut rng), &[2, 3, 2]);
    let b = cx.tensor::<R2<2, 4>>().set(random_vec_rng(2 * 4, &mut rng));
    let c = a.matmul(b).retrieve();
    (cx, vec![c.no_shape()])
}

pub fn feedforward() -> (Graph, Vec<GraphTensor<()>>) {
    let mut rng = StdRng::seed_from_u64(0);
    // Test single and batch, unoptimized and optimized
    let mut cx = Graph::new();
    let batch = cx
        .tensor::<(Dyn<'a'>, Const<3>)>()
        .set_dyn(random_vec_rng(2 * 3, &mut rng), &[2, 3]);
    let model: (Linear<3, 4>, ReLU, Linear<4, 2>) = InitModule::initialize(&mut cx);
    model.0.weight.set(random_vec_rng(3 * 4, &mut rng));
    model.2.weight.set(random_vec_rng(4 * 2, &mut rng));
    let batch_out = model.forward(batch).retrieve();

    (cx, vec![batch_out.no_shape()])
}

pub fn transformer() -> (Graph, Vec<GraphTensor<()>>) {
    let mut rng = StdRng::seed_from_u64(0);
    let mut cx = Graph::new();
    let model: Transformer<3, 4, 1, 1, 1, 1> = InitModule::initialize(&mut cx);
    model.decoder.layers[0]
        .self_attention
        .w_k
        .weight
        .set(random_vec_rng(3 * 3, &mut rng));
    model.decoder.layers[0]
        .self_attention
        .w_q
        .weight
        .set(random_vec_rng(3 * 3, &mut rng));
    model.decoder.layers[0]
        .self_attention
        .w_v
        .weight
        .set(random_vec_rng(3 * 3, &mut rng));
    model.decoder.layers[0]
        .self_attention
        .w_o
        .weight
        .set(random_vec_rng(3 * 3, &mut rng));
    model.decoder.layers[0]
        .cross_attention
        .w_k
        .weight
        .set(random_vec_rng(3 * 3, &mut rng));
    model.decoder.layers[0]
        .cross_attention
        .w_q
        .weight
        .set(random_vec_rng(3 * 3, &mut rng));
    model.decoder.layers[0]
        .cross_attention
        .w_v
        .weight
        .set(random_vec_rng(3 * 3, &mut rng));
    model.decoder.layers[0]
        .cross_attention
        .w_o
        .weight
        .set(random_vec_rng(3 * 3, &mut rng));
    model.decoder.layers[0]
        .ff
        .0
        .weight
        .set(random_vec_rng(3 * 4, &mut rng));
    model.decoder.layers[0]
        .ff
        .2
        .weight
        .set(random_vec_rng(3 * 4, &mut rng));
    model.encoder.modules[0]
        .attention
        .w_k
        .weight
        .set(random_vec_rng(3 * 3, &mut rng));
    model.encoder.modules[0]
        .attention
        .w_q
        .weight
        .set(random_vec_rng(3 * 3, &mut rng));
    model.encoder.modules[0]
        .attention
        .w_v
        .weight
        .set(random_vec_rng(3 * 3, &mut rng));
    model.encoder.modules[0]
        .attention
        .w_o
        .weight
        .set(random_vec_rng(3 * 3, &mut rng));
    model.encoder.modules[0]
        .ff
        .0
        .weight
        .set(random_vec_rng(3 * 4, &mut rng));
    model.encoder.modules[0]
        .ff
        .2
        .weight
        .set(random_vec_rng(3 * 4, &mut rng));

    let a = cx.tensor::<(Dyn<'d'>, crate::shape::Const<3>)>();
    let e = cx.tensor::<(Dyn<'e'>, crate::shape::Const<3>)>();
    let b = model.forward((a, e));

    a.set_dyn(random_vec_rng(2 * 3, &mut rng), &[2, 3]);
    e.set_dyn(random_vec_rng(3 * 3, &mut rng), &[3, 3]);
    b.retrieve();

    (cx, vec![b.no_shape()])
}