burn-tensor 0.1.0

This library provides multiple tensor implementations hidden behind an easy to use API that supports reverse mode automatic differentiation. 
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

use crate::graph::node::{ForwardNode, ForwardNodeState};
use crate::graph::ops::{
    BinaryOps, BinaryOpsNodeState, ForwardBinaryRecordedOps, ForwardUnaryRecordedOps, UnaryOps,
    UnaryOpsNodeState,
};
use crate::tensor::backend::autodiff::{ADKind, ADTensor};
use crate::tensor::ops::*;
use crate::tensor::{Element, Tensor};
use std::{ops::Range, sync::Arc};

#[derive(Debug)]
struct ADTensorOpsIndex<P, const D1: usize, const D2: usize> {
    indexes: [Range<usize>; D2],
    _kind: ADKind<P>,
}

impl<P: Default, const D1: usize, const D2: usize> ADTensorOpsIndex<P, D1, D2> {
    pub fn new(indexes: [Range<usize>; D2]) -> Self {
        Self {
            indexes,
            _kind: ADKind::new(),
        }
    }
}

impl<T1, P, const D1: usize, const D2: usize> UnaryOps<T1, T1> for ADTensorOpsIndex<P, D1, D2>
where
    P: Element,
    T1: Tensor<P, D1> + TensorOpsIndex<P, D1, D2>,
{
    fn partial(&self, state: &UnaryOpsNodeState<T1, T1>) -> T1 {
        state
            .input
            .value()
            .zeros()
            .index_assign(self.indexes.clone(), &state.output.grad())
    }
}

#[derive(Debug)]
struct ADTensorOpsIndexAssign<P, const D1: usize, const D2: usize> {
    indexes: [Range<usize>; D2],
    _kind: ADKind<P>,
}

impl<P: Default, const D1: usize, const D2: usize> ADTensorOpsIndexAssign<P, D1, D2> {
    pub fn new(indexes: [Range<usize>; D2]) -> Self {
        Self {
            indexes,
            _kind: ADKind::new(),
        }
    }
}

impl<T, P, const D1: usize, const D2: usize> BinaryOps<T, T, T>
    for ADTensorOpsIndexAssign<P, D1, D2>
where
    P: Element,
    T: Tensor<P, D1> + TensorOpsIndex<P, D1, D2>,
{
    fn partial_left(&self, state: &BinaryOpsNodeState<T, T, T>) -> T {
        state
            .output
            .grad()
            .index_assign(self.indexes.clone(), &state.right.value().zeros())
    }

    fn partial_right(&self, state: &BinaryOpsNodeState<T, T, T>) -> T {
        state.output.grad().index(self.indexes.clone())
    }
}

impl<P, const D1: usize, const D2: usize, T> TensorOpsIndex<P, D1, D2> for ADTensor<P, D1, T>
where
    P: Element,
    T: Tensor<P, D1> + TensorOpsIndex<P, D1, D2>,
{
    fn index(&self, indexes: [Range<usize>; D2]) -> Self {
        let input = self.tensor();
        let out = TensorOpsIndex::index(&input, indexes.clone());
        let shape = out.shape().clone();

        let state = ForwardNodeState::new(out);

        let ops = ADTensorOpsIndex::<P, D1, D2>::new(indexes);
        let ops = Arc::new(ops);
        let ops = ForwardUnaryRecordedOps::new(self.node.clone(), ops);
        let ops = Arc::new(ops);

        let node = ForwardNode::from_unary(&self.node, state, ops);
        let node = Arc::new(node);

        let kind = self.kind.clone();

        Self { node, shape, kind }
    }
    fn index_assign(&self, indexes: [Range<usize>; D2], values: &Self) -> Self {
        let input = self.tensor();
        let out = TensorOpsIndex::index_assign(&input, indexes.clone(), &values.tensor());
        let shape = out.shape().clone();

        let state = ForwardNodeState::new(out);

        let ops = ADTensorOpsIndexAssign::<P, D1, D2>::new(indexes);
        let ops = Arc::new(ops);
        let ops = ForwardBinaryRecordedOps::new(self.node.clone(), values.node.clone(), ops);
        let ops = Arc::new(ops);

        let node = ForwardNode::from_binary(&self.node, &values.node, state, ops);
        let node = Arc::new(node);

        let kind = self.kind.clone();

        Self { node, shape, kind }
    }
}
#[cfg(test)]
mod tests {
    use super::*;
    use crate::tensor::{backend::autodiff::helper::TestADTensor, Data};

    #[test]
    fn should_diff_matmul_with_index() {
        let data_1: Data<f64, 2> = Data::from([[1.0, 7.0], [2.0, 3.0]]);
        let data_2: Data<f64, 2> = Data::from([[4.0, 7.0, 100.0], [2.0, 3.0, 15.0]]);

        let tensor_1 = TestADTensor::from_data(data_1.clone());
        let tensor_2 = TestADTensor::from_data(data_2.clone());

        let tensor_3 = tensor_2.index([0..2, 0..2]);
        let tensor_4 = &tensor_1.matmul(&tensor_3);
        let grads = tensor_4.backward();

        let grad_1 = grads.wrt(&tensor_1).unwrap();
        let grad_2 = grads.wrt(&tensor_2).unwrap();

        assert_eq!(grad_1.to_data(), Data::from([[11.0, 5.0], [11.0, 5.0]]));
        assert_eq!(
            grad_2.to_data(),
            Data::from([[3.0, 3.0, 0.0], [10.0, 10.0, 0.0]])
        );
    }

    #[test]
    fn should_diff_matmul_with_index_assign() {
        let data_1: Data<f64, 2> = Data::from([[1.0, 7.0], [2.0, 3.0]]);
        let data_2: Data<f64, 2> = Data::from([[4.0, 7.0], [2.0, 3.0]]);
        let data_assigned: Data<f64, 2> = Data::from([[9.0]]);

        let tensor_1 = TestADTensor::from_data(data_1.clone());
        let tensor_2 = TestADTensor::from_data(data_2.clone());
        let tensor_assigned = TestADTensor::from_data(data_assigned.clone());

        let tensor_3 = tensor_1.matmul(&tensor_2);
        let tensor_4 = tensor_3.index_assign([0..1, 0..1], &tensor_assigned);
        let tensor_5 = &tensor_4.matmul(&tensor_1);

        let grads = tensor_5.backward();

        let grad_1 = grads.wrt(&tensor_1).unwrap();
        let grad_2 = grads.wrt(&tensor_2).unwrap();

        assert_eq!(grad_1.to_data(), Data::from([[58.0, 38.0], [118.0, 82.0]]));
        assert_eq!(grad_2.to_data(), Data::from([[16.0, 15.0], [24.0, 50.0]]));
    }

    #[test]
    fn should_diff_matmul_with_index_assign_complex() {
        let data_1: Data<f64, 2> = Data::from([[1.0, 7.0], [2.0, 3.0]]);
        let data_2: Data<f64, 2> = Data::from([[4.0, 7.0], [2.0, 3.0]]);
        let data_3: Data<f64, 2> = Data::from([[9.0]]);

        let tensor_1 = TestADTensor::from_data(data_1.clone());
        let tensor_2 = TestADTensor::from_data(data_2.clone());
        let tensor_3 = TestADTensor::from_data(data_3.clone());

        let tensor_4 = tensor_1.matmul(&tensor_2);
        let tensor_5 = tensor_2.index([0..1, 0..1]);
        let tensor_6 = tensor_5.mul(&tensor_3);
        let tensor_7 = tensor_4.index_assign([0..1, 0..1], &tensor_6);
        let tensor_8 = &tensor_7.matmul(&tensor_1);

        let grads = tensor_8.backward();

        let grad_1 = grads.wrt(&tensor_1).unwrap();
        let grad_2 = grads.wrt(&tensor_2).unwrap();
        let grad_3 = grads.wrt(&tensor_3).unwrap();

        assert_eq!(grad_3.to_data(), Data::from([[32.0]]));
        assert_eq!(grad_1.to_data(), Data::from([[85.0, 65.0], [118.0, 82.0]]));
        assert_eq!(grad_2.to_data(), Data::from([[88.0, 15.0], [24.0, 50.0]]));
    }
}