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
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
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217

use crate::prelude::*;

pub trait Matmul<S: Shape> {
    type Output;
    fn matmul(self, rhs: GraphTensor<S>) -> Self::Output;
}

// ABxBC -> AC
impl<A: Dimension, B: Dimension, C: Dimension> Matmul<(B, C)> for GraphTensor<(A, B)> {
    type Output = GraphTensor<(A, C)>;
    fn matmul(self, rhs: GraphTensor<(B, C)>) -> Self::Output {
        // Broadcasted Multiply
        let mul = self.expand::<(A, C, B), _>()
            * rhs.permute::<_, Axes2<1, 0>>().expand::<(A, C, B), _>();

        // Sum Reduce
        mul.sum_reduce::<_, Axis<2>>()
    }
}

// AxAB -> B
impl<A: Dimension, B: Dimension> Matmul<(A, B)> for GraphTensor<(A,)> {
    type Output = GraphTensor<(B,)>;
    fn matmul(self, rhs: GraphTensor<(A, B)>) -> Self::Output {
        let s: GraphTensor<(Const<1>, A)> = self.expand();

        // Run normal matmul
        let r = s.matmul(rhs);

        // Sum Reduce
        r.reshape()
    }
}

// ABCxCD -> ABD
impl<A: Dimension, B: Dimension, C: Dimension, D: Dimension> Matmul<(C, D)>
    for GraphTensor<(A, B, C)>
{
    type Output = GraphTensor<(A, B, D)>;
    fn matmul(self, rhs: GraphTensor<(C, D)>) -> Self::Output {
        // Reshape
        let w: GraphTensor<(D, C)> = rhs.permute::<_, Axes2<1, 0>>();

        // Broadcasted Multiply
        let mul = self.expand::<(A, B, D, C), _>() * w.expand::<(A, B, D, C), _>();

        // Sum Reduce
        mul.sum_reduce::<_, Axis<3>>()
    }
}

// ABCxACD -> ABD
impl<A: Dimension, B: Dimension, C: Dimension, D: Dimension> Matmul<(A, C, D)>
    for GraphTensor<(A, B, C)>
{
    type Output = GraphTensor<(A, B, D)>;
    fn matmul(self, rhs: GraphTensor<(A, C, D)>) -> Self::Output {
        // Reshape
        let w: GraphTensor<(A, D, C)> = rhs.permute::<_, Axes3<0, 2, 1>>();

        // Broadcasted Multiply
        let mul = self.expand::<(A, B, D, C), _>() * w.expand::<(A, B, D, C), _>();

        // Sum Reduce
        mul.sum_reduce::<_, Axis<3>>()
    }
}

// ABCDxABDE -> ABCE
impl<A: Dimension, B: Dimension, C: Dimension, D: Dimension, E: Dimension> Matmul<(A, B, D, E)>
    for GraphTensor<(A, B, C, D)>
{
    type Output = GraphTensor<(A, B, C, E)>;
    fn matmul(self, rhs: GraphTensor<(A, B, D, E)>) -> Self::Output {
        // Reshape
        let w: GraphTensor<(A, B, E, D)> = rhs.permute::<_, Axes4<0, 1, 3, 2>>();

        // Broadcasted Multiply
        let mul = self.expand::<(A, B, C, E, D), _>() * w.expand::<(A, B, C, E, D), _>();

        // Sum Reduce
        mul.sum_reduce::<_, Axis<4>>()
    }
}

// ABCDExABCEF -> ABCDF
impl<A: Dimension, B: Dimension, C: Dimension, D: Dimension, E: Dimension, F: Dimension>
    Matmul<(A, B, C, E, F)> for GraphTensor<(A, B, C, D, E)>
{
    type Output = GraphTensor<(A, B, C, D, F)>;
    fn matmul(self, rhs: GraphTensor<(A, B, C, E, F)>) -> Self::Output {
        // Reshape
        let w: GraphTensor<(A, B, C, F, E)> = rhs.permute::<_, Axes5<0, 1, 2, 4, 3>>();

        // Broadcasted Multiply
        let mul = self.expand::<(A, B, C, D, F, E), _>() * w.expand::<(A, B, C, D, F, E), _>();

        // Sum Reduce
        mul.sum_reduce::<_, Axis<5>>()
    }
}

impl<A: Dimension> GraphTensor<(A,)> {
    /// Simple dot product of two vectors
    pub fn dot(self, rhs: GraphTensor<(A,)>) -> GraphTensor<R0> {
        (self * rhs).sum_reduce()
    }
}

#[cfg(test)]
mod tests {
    crate::test_imports!();

    #[test]
    fn test_matrix_vector() {
        let mut cx = Graph::new();
        let (a_vec, b_vec) = (random_vec(3), random_vec(6));
        let a = cx.tensor::<R1<3>>().set(a_vec.clone());
        let b = cx.tensor::<R2<3, 2>>().set(b_vec.clone());
        let mut c = a.matmul(b).retrieve();

        cx.compile(GenericCompiler::default(), &mut c);
        cx.execute();

        let d_dev = Cpu::default();
        let d_a = d_dev.tensor_from_vec(a_vec, (DConst::<3>,));
        let d_b = d_dev.tensor_from_vec(b_vec, (DConst::<3>, DConst::<2>));
        let d_c = d_a.matmul(d_b);

        assert_close(&c.data(), &d_c.as_vec());
    }

    #[test]
    fn test_matmul() {
        let mut cx = Graph::new();
        let (a_data, b_data) = (random_vec(6), random_vec(9));
        let a = cx.tensor::<R2<2, 3>>();
        a.set(a_data.clone());
        let b = cx.tensor::<R2<3, 3>>();
        b.set(b_data.clone());
        let c = a.matmul(b);
        c.retrieve();

        cx.execute();

        let d_dev = Cpu::default();
        let d_a = d_dev.tensor_from_vec(a_data, (DConst::<2>, DConst::<3>));
        let d_b = d_dev.tensor_from_vec(b_data, (DConst::<3>, DConst::<3>));
        let d_c = d_a.matmul(d_b);

        assert_close(&c.data(), &d_c.as_vec());
    }

    #[test]
    fn test_batch_matmul() {
        let mut cx = Graph::new();
        let (a_data, b_data) = (random_vec(12), random_vec(8));
        let a = cx.tensor::<R3<2, 3, 2>>();
        a.set(a_data.clone());
        let b = cx.tensor::<R2<2, 4>>();
        b.set(b_data.clone());
        let c = a.matmul(b);
        c.retrieve();

        cx.execute();

        let d_dev = Cpu::default();
        let d_a = d_dev.tensor_from_vec(a_data, (DConst::<2>, DConst::<3>, DConst::<2>));
        let d_b = d_dev.tensor_from_vec(b_data, (DConst::<2>, DConst::<4>));
        let d_c = d_a.matmul(d_b);

        assert_close(&c.data(), &d_c.as_vec());
    }

    #[test]
    fn test_batch_batch_matmul() {
        let mut cx = Graph::new();
        let (a_data, b_data) = (random_vec(6), random_vec(6));
        let a = cx.tensor::<R3<1, 2, 3>>();
        a.set(a_data.clone());
        let b = cx.tensor::<R3<1, 2, 3>>();
        b.set(b_data.clone());
        let c: GraphTensor<R3<1, 2, 2>> = a.matmul(b.permute::<R3<1, 3, 2>, _>());
        c.retrieve();

        cx.execute();

        let d_dev = Cpu::default();
        let d_a = d_dev.tensor_from_vec(a_data, (DConst::<1>, DConst::<2>, DConst::<3>));
        let d_b = d_dev.tensor_from_vec(b_data, (DConst::<1>, DConst::<2>, DConst::<3>));
        let d_c = d_a.matmul(d_b.permute::<Rank3<1, 3, 2>, DAxes3<0, 2, 1>>());

        assert_close(&c.data(), &d_c.as_vec());
    }

    #[test]
    fn test_batch_batch_matmul2() {
        let mut cx = Graph::new();
        let (a_data, b_data) = (random_vec(4), random_vec(6));
        let a = cx.tensor::<(Dyn<'a'>, Dyn<'b'>)>();
        a.set_dyn(a_data.clone(), &[2, 2]);
        let a = a.expand::<(LConst<1>, Dyn<'a'>, Dyn<'b'>), _>();
        let b = cx.tensor::<(LConst<1>, Dyn<'b'>, LConst<3>)>();
        b.set_dyn(b_data.clone(), &[1, 2, 3]);
        let c = a.matmul(b);
        c.retrieve();

        cx.execute();

        let d_dev = Cpu::default();
        let d_a = d_dev.tensor_from_vec(a_data, (DConst::<1>, DConst::<2>, DConst::<2>));
        let d_b = d_dev.tensor_from_vec(b_data, (DConst::<1>, DConst::<2>, DConst::<3>));
        let d_c = d_a.matmul(d_b);

        assert_close(&c.data(), &d_c.as_vec());
    }
}