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
use alloc::vec::Vec;
use zyx_core::backend::Backend;
use zyx_core::tensor::Tensor;

/// # Adaptive momentum estimation optimizer
pub struct Adam<B: Backend> {
    /// learning rate (default: 1e-3)
    pub learning_rate: f32,
    /// coefficients used for computing running averages of gradient and its square (default: (0.9, 0.999))
    pub betas: (f32, f32),
    /// term added to the denominator to improve numerical stability (default: 1e-8)
    pub eps: f32,
    /// weight decay (L2 penalty) (default: 0)
    pub weight_decay: f32,
    /// whether to use the AMSGrad variant of this algorithm from the paper On the Convergence of Adam and Beyond (default: false)
    pub amsgrad: bool,
    /// maximize the objective with respect to the params, instead of minimizing (default: false)
    pub maximize: bool,
    /// m
    pub m: Vec<Tensor<B>>,
    /// v
    pub v: Vec<Tensor<B>>,
    /// vm
    pub vm: Vec<Tensor<B>>,
    /// t
    pub t: usize,
}

impl<B: Backend> Default for Adam<B> {
    fn default() -> Self {
        Self {
            learning_rate: 0.001,
            betas: (0.9, 0.999),
            eps: 1e-8,
            weight_decay: 0.0,
            amsgrad: false,
            maximize: false,
            m: Vec::new(),
            v: Vec::new(),
            vm: Vec::new(),
            t: 0,
        }
    }
}

impl<B: Backend> Adam<B> {
    /// Updates parameters with gradients.
    /// Number of parameters must be the same as number of gradients.
    /// Gradients can be None, those are simply skipped.
    pub fn update<'a>(
        &mut self,
        parameters: impl IntoIterator<Item = &'a mut Tensor<B>>,
        gradients: impl IntoIterator<Item = Option<Tensor<B>>>,
    ) where
        B: 'a,
    {
        let params: Vec<&mut Tensor<B>> = parameters.into_iter().collect();
        let grads: Vec<Option<Tensor<B>>> = gradients.into_iter().collect();

        assert_eq!(
            params.len(),
            grads.len(),
            "Number of parameters != number of gradients."
        );

        for (i, (param, grad)) in params.into_iter().zip(grads).enumerate() {
            if let Some(mut grad) = grad {
                if self.maximize {
                    grad = -grad;
                }
                if self.weight_decay != 0.0 {
                    grad = grad + &*param * self.weight_decay;
                }
                if let Some(m) = self.m.get_mut(i) {
                    *m = &*m * self.betas.0 + &grad * (1.0 - self.betas.0);
                } else {
                    self.m.push(&grad * (1.0 - self.betas.0));
                }
                if let Some(v) = self.m.get_mut(i) {
                    *v = &*v * self.betas.1 + &grad * &grad * (1.0 - self.betas.1);
                } else {
                    self.v.push(&grad * &grad * (1.0 - self.betas.1));
                }
                let mh = &self.m[i] / (1.0 - self.betas.0.powi(self.t as i32));
                let vh = &self.v[i] / (1.0 - self.betas.1.powi(self.t as i32));
                if self.amsgrad {
                    if let Some(vm) = self.vm.get_mut(i) {
                        *vm = vm.cmplt(&vh).where_(vh, &*vm);
                    } else {
                        self.vm.push(vh);
                    }
                    // Cast since learning_rate is f32, but parameters can have different precision.
                    // Can this cast be somehow avoided? Is it better to always work with original dtype?
                    *param = (&*param - mh / ((self.vm[i].sqrt() + self.eps) * self.learning_rate))
                        .cast(param.dtype());
                } else {
                    *param = (&*param - mh / ((vh.sqrt() + self.eps) * self.learning_rate))
                        .cast(param.dtype());
                }
            }
        }
    }
}