mushin 0.5.0

Computational graphs with reverse automatic differentation in the GPU
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

use crate::graph::{node::Node, tape::Tape};
use crate::tensor::{constant::Constant, Tensor};
use arrayfire::Array;
use std::rc::Rc;

/// A differentiable tensor being tracked in the computation graph
pub struct Variable<const B: u64, const L: u64, const R: u64, const C: u64> {
    tape: Tape,
    node: Rc<Node>,
}

impl<const B: u64, const L: u64, const R: u64, const C: u64> Variable<B, L, R, C> {
    /// Creates a new variable. It assumes its node has already been pushed to the computation graph.
    pub(crate) fn new(mut tape: Tape, node: Node) -> Self {
        let node = Rc::new(node);
        tape.push(node.clone());
        Self { tape, node }
    }

    /// Returns the graph where this tensor is being tracked
    pub(crate) const fn tape(&self) -> &Tape {
        &self.tape
    }

    /// Starting from this tensor node, compute the reverse auto differentiation.
    /// Once called, all the ancestor nodes for which this tensor depends on will have
    /// their gradients filled with the derivative with respect to this tensor
    pub fn backward(&self) {
        // derivative of self wrt to self is one
        self.node.ones_grad();
        for node in self.tape.nodes().rev() {
            node.reverse();
        }
    }

    /// Set all gradients to zero, including this tensor's and all its ancestors
    pub fn reset(&self) {
        for node in self.tape.nodes().rev() {
            node.zero_grad();
        }
    }

    /// Returns as a new `Variable` tensor, the gradients of `Z` with respect to itself.
    /// Where `Z` is the tensor for which `backward` was called. Note that if `backward` was
    /// not called, or this tensor is not a dependency of `Z` the result will always be a
    /// tensor filled with zeros. If on the other hand, this is the same tensor as the one
    /// where `backward` was called, the result will always be filled with ones, because that
    /// is `dz/dz`
    pub fn grad(&self) -> Self {
        Self::new(Tape::default(), Node::declaration(self.node.grad().clone()))
    }

    /// Consume this `Variable` tensor and return a `Constant` that is not tracked in the
    /// computation graph
    pub fn freeze(self) -> Constant<B, L, R, C> {
        Constant::new(self.data())
    }
}

impl<const B: u64, const L: u64, const R: u64, const C: u64> From<&Variable<B, L, R, C>>
    for Rc<Node>
{
    #[inline]
    fn from(tensor: &Variable<B, L, R, C>) -> Self {
        tensor.node.clone()
    }
}

impl<const B: u64, const L: u64, const R: u64, const C: u64> Tensor<B, L, R, C>
    for Variable<B, L, R, C>
{
    fn data(&self) -> Array<f32> {
        self.node.data().clone()
    }
}

#[cfg(test)]
mod tests {
    use super::Variable;
    use crate::graph::{node::Node, tape::Tape};
    use crate::tensor::Tensor;
    use crate::tests::equal_arrays;

    #[test]
    fn new_variable() {
        let tensor = Variable::<3, 4, 2, 1>::new(
            Tape::default(),
            Node::declaration(arrayfire::constant!(5.0; 2,1,4,3)),
        );
        assert!(equal_arrays(
            tensor.data(),
            arrayfire::constant!(5.0; 2,1,4,3)
        ));
        assert_eq!(tensor.node.id(), 0);
    }

    #[test]
    fn variable_freeze() {
        let tensor = Variable::<3, 4, 2, 1>::new(
            Tape::default(),
            Node::declaration(arrayfire::constant!(5.0; 2,1,4,3)),
        )
        .freeze();
        assert!(equal_arrays(
            (&tensor).into(),
            arrayfire::constant!(5.0; 2,1,4,3)
        ));
    }

    #[test]
    fn variable_backward() {
        let tensor = Variable::<3, 4, 2, 1>::new(
            Tape::default(),
            Node::declaration(arrayfire::constant!(5.0; 2,1,4,3)),
        );
        tensor.backward();
        assert!(equal_arrays(
            tensor.grad().data(),
            arrayfire::constant!(1.0; 2,1,4,3)
        ));
    }
}