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

use std::sync::{ Arc, Mutex };

use crate::rotta_rs_module::{ arrayy::Arrayy, Backward, NodeType };

pub struct AdaGrad {
    parameters: Arc<Mutex<Vec<NodeType>>>,
    pub lr: Arrayy,
    pub g: Vec<Arrayy>,
    pub eps: f64,
    pub auto_zero_grad_execute: bool,
}

impl AdaGrad {
    pub fn init(parameters: Arc<Mutex<Vec<NodeType>>>, lr: f64) -> AdaGrad {
        let lr = Arrayy::from_vector(vec![1], vec![lr]);
        AdaGrad {
            parameters,
            lr,
            g: vec![],
            eps: 1e-8,
            auto_zero_grad_execute: true,
        }
    }

    // zero
    pub fn zero_grad(&self) {
        for node_type in self.parameters.lock().unwrap().iter() {
            node_type.lock().as_mut().unwrap().zero_grad();
        }
    }

    // optimazer
    pub fn optim(&mut self, backward: Backward) {
        for (i, node_type) in self.parameters.lock().unwrap().iter().enumerate() {
            let mut node = node_type.lock().unwrap();
            if let None = self.g.get(i) {
                self.g.push(Arrayy::arrayy_from_element(node.value.shape.clone(), 0.0));
            }

            // ada grad
            // g_n = g_n - 1 + g_n^2
            // w_n + 1 = w_n - (lr/(g_n^0.5 + e).sqrt) * grad(w_n)

            let eps = self.eps;
            let grad = &node.grad;
            let g_n = &self.g[i] + grad.powi(2);
            let new = &node.value - (&self.lr / (g_n.powf(0.5) + eps)) * grad;

            node.update_value(new);
            // update g
            self.g[i] = g_n;
        }

        // auto_grad_zero
        if self.auto_zero_grad_execute {
            backward.zero_grad();
        }
    }
}