wyrm 0.9.1

A low-overhead, define-by-run autodifferentiation library.
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

use super::Optimizer;
use std::ops::DerefMut;
use std::sync::Arc;
use {numerics, Arr, HogwildParameter, ParameterNode, Variable};

use ndarray::Axis;

struct AdamParameters<'params> {
    value: &'params mut Arr,
    m: &'params mut Arr,
    v: &'params mut Arr,
    t: &'params mut i32,
}

/// ADAM optimizer.
pub struct Adam {
    learning_rate: f32,
    l2: f32,
    beta_m: f32,
    beta_v: f32,
    eps: f32,
    clamp: Option<(f32, f32)>,
}

impl Default for Adam {
    fn default() -> Self {
        Self::new()
    }
}

impl Adam {
    /// Build new optimizer object.
    pub fn new() -> Self {
        Self {
            learning_rate: 0.05,
            l2: 0.0,
            beta_m: 0.9,
            beta_v: 0.999,
            eps: 1.0e-8,
            clamp: None,
        }
    }

    /// Set the learning rate.
    pub fn learning_rate(mut self, learning_rate: f32) -> Self {
        self.learning_rate = learning_rate;
        self
    }

    /// Set the clamp bounds.
    pub fn clamp(mut self, min: f32, max: f32) -> Self {
        self.clamp = Some((min, max));
        self
    }

    /// Set the L2 penalty.
    pub fn l2_penalty(mut self, l2_penalty: f32) -> Self {
        self.l2 = l2_penalty;
        self
    }

    fn param_fields<'par>(&self, value: &'par Arc<HogwildParameter>) -> AdamParameters<'par> {
        AdamParameters {
            value: unsafe { value.value_mut() },
            m: unsafe { value.squared_gradient_mut() },
            v: unsafe { value.moments_mut() },
            t: unsafe { value.num_updates_mut() },
        }
    }

    #[inline(always)]
    fn update(&self, value: &mut f32, gradient: f32, m: &mut f32, v: &mut f32, t: &i32) {
        // Apply L2 to gradient.
        let gradient = gradient + *value * self.l2;

        // Update m and v.
        *m = self.beta_m * *m + (1.0 - self.beta_m) * gradient;
        *v = self.beta_v * *v + (1.0 - self.beta_v) * numerics::pow2(gradient);

        let m_hat = *m / (1.0 - self.beta_m.powi(*t));
        let v_hat = *v / (1.0 - self.beta_v.powi(*t));

        *value -= self.learning_rate / (v_hat.sqrt() + self.eps) * m_hat;
    }

    fn inner_step<T: DerefMut<Target = ::nodes::GradientAccumulator>>(
        &self,
        param: &Arc<HogwildParameter>,
        mut sink: T,
    ) {
        if let Some((min, max)) = self.clamp {
            sink.clamp(min, max);
        }

        let param = self.param_fields(param);

        // Increment number of updates
        *param.t = param.t.saturating_add(1);

        if sink.has_dense() {
            for (value, &gradient, m, v) in izip!(
                param.value.as_slice_mut().unwrap(),
                sink.gradient().as_slice().unwrap(),
                param.m.as_slice_mut().unwrap(),
                param.v.as_slice_mut().unwrap(),
            ) {
                self.update(value, gradient, m, v, param.t);
            }
        } else {
            for (row_idx, ref grad) in sink.sparse_iter() {
                let mut value_row = param.value.subview_mut(Axis(0), row_idx);
                let mut m_row = param.m.subview_mut(Axis(0), row_idx);
                let mut v_row = param.v.subview_mut(Axis(0), row_idx);

                for (value, &gradient, m, v) in izip!(
                    value_row.as_slice_mut().unwrap(),
                    grad.iter(),
                    m_row.as_slice_mut().unwrap(),
                    v_row.as_slice_mut().unwrap(),
                ) {
                    self.update(value, gradient, m, v, param.t);
                }
            }
        }
    }
}

impl Optimizer for Adam {
    /// Perform a single SGD step.
    fn step(&self, parameters: &[Variable<ParameterNode>]) {
        for parameter in parameters {
            self.inner_step(&parameter.node.value, parameter.node.gradient.borrow_mut());
            parameter.node.zero_gradient();
        }
    }
}