rotta_rs 0.0.5

a Deep Learning library with rust language
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 std::{ collections::HashSet, sync::{ Arc, Mutex } };

use crate::{
    d_concat,
    rotta_rs_module::{
        d_abs,
        d_add,
        d_broadcasting_tensor,
        d_cel,
        d_divided,
        d_dot,
        d_exp,
        d_index,
        d_ln,
        d_matmul,
        d_mul,
        d_permute,
        d_powf,
        d_powi,
        d_relu,
        d_sign,
        d_slice,
        d_ssresidual,
        d_sub,
        d_sum,
        d_sum_axis,
        d_to_shape,
        BackwardLabel,
        NodeType,
        Tensor,
    },
};

pub struct Backward {
    map: Arc<Mutex<Vec<NodeType>>>,
}

impl Backward {
    pub fn zero_grad(&self) {
        for node in self.map.lock().unwrap().iter() {
            let mut node = node.lock().unwrap();
            if node.auto_zero_grad {
                node.zero_grad();
            }
        }
    }
}

impl Tensor {
    pub fn backward(&self) -> Backward {
        let node = self.node.clone();
        node.lock().as_mut().unwrap().ones_grad();

        let mut graph: Vec<NodeType> = vec![];
        let mut visited: HashSet<u128> = HashSet::new();

        build(node, &mut graph, &mut visited);

        for idx in (0..graph.len()).rev() {
            let node_arc = graph[idx].clone();
            let node = node_arc.lock().unwrap();
            let grad = node.grad.clone();

            // println!("===============");
            // if let Some(label) = &node.label {
            //     match label {
            //         BackwardLabel::Broadcasting(_, _) => (),
            //         _ => {
            //             println!("{}", grad);
            //             println!("{:?}", label);
            //         }
            //     }
            // }

            // println!("===============");

            if let Some(label) = &node.label {
                match label {
                    // opearation
                    BackwardLabel::Dot(a, b) => d_dot(a, b, &grad),
                    BackwardLabel::Matmul(a, b) => d_matmul(a, b, &grad),
                    BackwardLabel::Add(a, b) => d_add(a, b, grad),
                    BackwardLabel::Diveded(a, b) => d_divided(a, b, &grad),
                    BackwardLabel::Mul(a, b) => d_mul(a, b, &grad),
                    BackwardLabel::Sub(a, b) => d_sub(a, b, &grad),

                    // mutation
                    BackwardLabel::Index(x, index) => d_index(x, index.clone(), &grad),
                    BackwardLabel::Broadcasting(tensor_arr, broad_arr) =>
                        d_broadcasting_tensor(tensor_arr, broad_arr.clone(), grad),
                    BackwardLabel::SumAxis(x, d, keep_dim) => d_sum_axis(x, d, *keep_dim, &grad),
                    BackwardLabel::Sum(x) => d_sum(x, &grad),
                    BackwardLabel::Permute(x, order) => d_permute(x, order.clone(), &grad),
                    BackwardLabel::Slice(x, range) => d_slice(x, range.clone(), &grad),
                    BackwardLabel::ToShape(x, to_shape) => d_to_shape(x, to_shape.clone(), &grad),
                    BackwardLabel::Concat(nodes, dim) => d_concat(nodes.clone(), *dim, &grad),

                    // function
                    BackwardLabel::Exp(a, exp_value) => d_exp(a, exp_value, &grad),
                    BackwardLabel::Powi(x, powi) => d_powi(x, *powi, &grad),
                    BackwardLabel::Powf(x, powf) => d_powf(x, *powf, &grad),
                    BackwardLabel::Ln(x) => d_ln(x, &grad),
                    BackwardLabel::Abs(x) => d_abs(x, &grad),
                    BackwardLabel::Sign(x) => d_sign(x),

                    // activation
                    BackwardLabel::Relu(x) => d_relu(x, &grad),

                    // loss
                    BackwardLabel::SSResidual(prediction, actual) =>
                        d_ssresidual(prediction, actual, &grad),
                    BackwardLabel::CEL(prob_prediction, prob_actual) =>
                        d_cel(prob_prediction, prob_actual, &grad),
                }
            }
        }

        let backward = Backward {
            map: Arc::new(Mutex::new(graph)),
        };

        backward
    }
}

pub fn build(node_arc: NodeType, graph: &mut Vec<NodeType>, vitited: &mut HashSet<u128>) {
    let node = node_arc.lock().unwrap();
    if node.requires_grad {
        if let None = vitited.get(&node.id) {
            vitited.insert(node.id);

            let parents = node.parent.clone();

            for parent in parents {
                build(parent, graph, vitited);
            }

            graph.push(node_arc.clone());
        }
    }
}