ferrite-dl 0.2.0

Deep learning library written in pure Rust
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

use super::{Graph, Scalar, Value};  // Import Graph from parent module
// scalar.rs
use std::cell::{RefCell};
use std::collections::HashSet;

pub trait Backward {
  fn add_backward(&self, lhs: usize, rhs: usize, grad: f32);
  fn sub_backward(&self, lhs: usize, rhs: usize, grad: f32);
  fn mul_backward(&self, lhs: usize, rhs: usize, grad: f32);
  fn div_backward(&self, lhs: usize, rhs: usize, grad: f32);
  fn exp_backward(&self, lhs: usize, grad: f32);
  fn pow_backward(&self, lhs: usize, rhs: usize, grad: f32);
  fn backward(&self, value: Value);
  fn backward_from(&self, idx: usize);

}

impl Backward for Graph {
  // d(a + b)/da = 1, d(a + b)/db = 1
  fn add_backward(&self, lhs: usize, rhs: usize, grad: f32) {
    let scalars = self.scalars.borrow();
    let lhs_scalar = &scalars[lhs];
    let rhs_scalar = &scalars[rhs];
    lhs_scalar.grad.set(lhs_scalar.grad.get() + grad);
    rhs_scalar.grad.set(rhs_scalar.grad.get() + grad);
  }

  // d(a + b)/da = 1, d(a + b)/db = 1
  fn sub_backward(&self, lhs: usize, rhs: usize, grad: f32) {
    let scalars = self.scalars.borrow();
    let lhs_scalar = &scalars[lhs];
    let rhs_scalar = &scalars[rhs];
    lhs_scalar.grad.set(lhs_scalar.grad.get() + grad);
    rhs_scalar.grad.set(rhs_scalar.grad.get() - grad);
  }

  // d(a * b)/da = b, d(a * b)/db = a
  fn mul_backward(&self, lhs: usize, rhs: usize, grad: f32) {
    let scalars = self.scalars.borrow();
    let lhs_scalar = &scalars[lhs];
    let rhs_scalar = &scalars[rhs];
    lhs_scalar.grad.set(lhs_scalar.grad.get() + rhs_scalar.data * grad);
    rhs_scalar.grad.set(rhs_scalar.grad.get() + lhs_scalar.data * grad);
  }

  // d(a/b)/da = 1/b, d(a/b)/db = (-a/b**2)
  fn div_backward(&self, lhs: usize, rhs: usize, grad: f32) {
    let scalars = self.scalars.borrow();
    let lhs_scalar = &scalars[lhs];
    let rhs_scalar = &scalars[rhs];
    lhs_scalar.grad.set(lhs_scalar.grad.get() + (1./rhs_scalar.data) * grad);
    rhs_scalar.grad.set(rhs_scalar.grad.get() + (-lhs_scalar.data/(rhs_scalar.data*rhs_scalar.data)) * grad);
  }


  // d(e^x)/dx = e^x
  fn exp_backward(&self, lhs: usize, grad: f32) {
    let scalars = self.scalars.borrow();
    let lhs_scalar = &scalars[lhs];
    let exp_x = f32::exp(lhs_scalar.data);
    lhs_scalar.grad.set(lhs_scalar.grad.get() + grad * exp_x);
  }

  // d(x^a)/dx = ax^(a-1), d(x^a)/da = x^a * lnx
  fn pow_backward(&self, lhs: usize, rhs: usize, grad: f32) {
    let scalars = self.scalars.borrow();
    let lhs_scalar = &scalars[lhs];
    let rhs_scalar = &scalars[rhs];
    lhs_scalar.grad.set(lhs_scalar.grad.get() + (rhs_scalar.data * (f32::powf(lhs_scalar.data, rhs_scalar.data - 1.))) * grad);
    rhs_scalar.grad.set(rhs_scalar.grad.get() + (f32::powf(lhs_scalar.data, rhs_scalar.data) * f32::ln(lhs_scalar.data)) * grad);
  }


  fn backward(&self, value: Value) {
    let id = value.idx;
    self.scalars.borrow()[id].grad.set(1.0);

    // Topo sort
    let mut topo = vec![];
    let mut visited = HashSet::new();

    fn build_topo(v: usize, scalars: &RefCell<Vec<Scalar>>, visited: &mut HashSet<usize>, topo: &mut Vec<usize>) {
      if !visited.contains(&v) {
        visited.insert(v);

        let children = {
          let scalars = scalars.borrow();
          scalars[v].prev.clone()
        };

        for &child in &children {
          build_topo(child, scalars, visited, topo);
        }

        topo.push(v);
      }
    }

    build_topo(id, &self.scalars, &mut visited, &mut topo);
    
    for &idx in topo.iter().rev(){
      self.backward_from(idx);
    }

   // 
  }

  fn backward_from(&self, idx: usize) {
    let scalars = self.scalars.borrow();
    let scalar = &scalars[idx];

    if !scalar.requires_grad {
      return;
    }

    let grad = scalar.grad.get();

    match scalar.op.as_str() {
      "+" => {
        let lhs = scalar.prev[0];
        let rhs = scalar.prev[1];
        self.add_backward(lhs, rhs, grad);
      },
      "-" => {
        let lhs = scalar.prev[0];
        let rhs = scalar.prev[1];
        self.sub_backward(lhs, rhs, grad);
      },
      "*" => {
        let lhs = scalar.prev[0];
        let rhs = scalar.prev[1];
        self.mul_backward(lhs, rhs, grad);
      },
      "/" => {
        let lhs = scalar.prev[0];
        let rhs = scalar.prev[1];
        self.div_backward(lhs, rhs, grad);
      },
      "exp" => {
        let lhs = scalar.prev[0];
        self.exp_backward(lhs, grad);
      },
      "pow" => {
        let lhs = scalar.prev[0];
        let rhs = scalar.prev[1];
        self.pow_backward(lhs, rhs, grad);
      },
      
      _ => {}
    }
  }

}