ghostflow-optim 1.0.0

Optimizers for GhostFlow ML framework
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

//! Stochastic Gradient Descent optimizer

use ghostflow_core::Tensor;
use crate::optimizer::Optimizer;

/// SGD optimizer with optional momentum and weight decay
pub struct SGD {
    params: Vec<Tensor>,
    lr: f32,
    momentum: f32,
    weight_decay: f32,
    dampening: f32,
    nesterov: bool,
    velocity: Vec<Vec<f32>>,
}

impl SGD {
    pub fn new(params: Vec<Tensor>, lr: f32) -> Self {
        let velocity = params.iter().map(|p| vec![0.0f32; p.numel()]).collect();
        
        SGD {
            params,
            lr,
            momentum: 0.0,
            weight_decay: 0.0,
            dampening: 0.0,
            nesterov: false,
            velocity,
        }
    }

    pub fn momentum(mut self, momentum: f32) -> Self {
        self.momentum = momentum;
        self
    }

    pub fn weight_decay(mut self, weight_decay: f32) -> Self {
        self.weight_decay = weight_decay;
        self
    }

    pub fn dampening(mut self, dampening: f32) -> Self {
        self.dampening = dampening;
        self
    }

    pub fn nesterov(mut self, nesterov: bool) -> Self {
        self.nesterov = nesterov;
        self
    }
}

impl Optimizer for SGD {
    fn step(&mut self) {
        for (i, param) in self.params.iter_mut().enumerate() {
            if let Some(grad) = param.grad() {
                let mut grad_data = grad.data_f32();
                let param_data = param.data_f32();
                
                // Weight decay (L2 regularization)
                if self.weight_decay != 0.0 {
                    for (g, &p) in grad_data.iter_mut().zip(param_data.iter()) {
                        *g += self.weight_decay * p;
                    }
                }
                
                // Momentum
                if self.momentum != 0.0 {
                    let v = &mut self.velocity[i];
                    
                    for (j, g) in grad_data.iter().enumerate() {
                        v[j] = self.momentum * v[j] + (1.0 - self.dampening) * g;
                    }
                    
                    if self.nesterov {
                        for (j, g) in grad_data.iter_mut().enumerate() {
                            *g += self.momentum * self.velocity[i][j];
                        }
                    } else {
                        grad_data = self.velocity[i].clone();
                    }
                }
                
                // Update parameters
                let new_data: Vec<f32> = param_data.iter()
                    .zip(grad_data.iter())
                    .map(|(&p, &g)| p - self.lr * g)
                    .collect();
                
                *param = Tensor::from_slice(&new_data, param.dims()).unwrap();
            }
        }
    }

    fn zero_grad(&mut self) {
        for param in &mut self.params {
            param.zero_grad();
        }
    }

    fn get_lr(&self) -> f32 {
        self.lr
    }

    fn set_lr(&mut self, lr: f32) {
        self.lr = lr;
    }

    fn parameters(&self) -> &[Tensor] {
        &self.params
    }
}

#[cfg(test)]
mod tests {
    use super::*;

    #[test]
    fn test_sgd_step() {
        let mut param = Tensor::ones(&[3]);
        param.set_requires_grad(true);
        param.set_grad(Tensor::full(&[3], 0.1f32));
        
        let mut sgd = SGD::new(vec![param], 0.1);
        sgd.step();
        
        let updated = &sgd.params[0];
        // 1.0 - 0.1 * 0.1 = 0.99
        assert!((updated.data_f32()[0] - 0.99).abs() < 1e-6);
    }
}