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
155
156
157
158
159

extern crate ndarray;

use ndarray_ext::NdArray;
use op;
use std::collections::BTreeMap;
use std::collections::btree_map::Entry;
use tensor::Tensor;

struct AdamOp
{
    static_params: StaticParams,
}

impl ::op::Op for AdamOp
{
    fn name(&self) -> &str
    {
        "Adam"
    }

    fn compute(&self, mut ctx: ::runtime::OpComputeContext) -> op::ComputeResult
    {
        let StaticParams { alpha, eps, b1, b2 } = self.static_params;
        let xs = unsafe { ctx.grab_assignable_inputs() };

        // Make new m
        let new_m = {
            let mut new_m = (xs[2] as &NdArray) * b1;
            let tmp = 1. - b1;
            new_m.zip_mut_with(xs[1], move |a, &g| *a += tmp * g);
            new_m
        };

        // Make new v
        let new_v = {
            let mut new_v = (xs[3] as &NdArray) * b2;
            let tmp = 1. - b2;
            new_v.zip_mut_with(xs[1], move |a, &g| *a += tmp * g * g);
            new_v
        };

        // Make hat
        let m_hat = {
            let t = xs[4][ndarray::IxDyn(&[])];
            let v_hat = new_v * (1. / (1. - b2.powf(t)));
            let mut m_hat = new_m * (1. / (1. - b1.powf(t)));
            m_hat.zip_mut_with(&v_hat, move |a, &b| (*a) /= b.sqrt() + eps);
            m_hat
        };

        // Update t and param
        xs[4][ndarray::IxDyn(&[])] += 1.;
        xs[0].scaled_add(-alpha, &m_hat);
        vec![Err(::op::ComputeError::NoOutput)]
    }

    fn grad(&self, _: &Tensor, _: &[&Tensor], _: &Tensor) -> Vec<Option<Tensor>>
    {
        vec![None]
    }
}

/// Adam optimizer
///
/// This implementation is based on http://arxiv.org/abs/1412.6980v8
pub struct Adam<'a>
{
    pub alpha:           f32,
    pub eps:             f32,
    pub b1:              f32,
    pub b2:              f32,
    pub stateful_params: BTreeMap<super::StateKey<'a>, StatefulParams>,
}

impl<'a> Default for Adam<'a>
{
    fn default() -> Adam<'a>
    {
        Adam {
            alpha:           0.001,
            eps:             1e-08,
            b1:              0.9,
            b2:              0.999,
            stateful_params: BTreeMap::new(),
        }
    }
}

#[derive(Copy, Clone)]
pub struct StaticParams
{
    pub alpha: f32,
    pub eps:   f32,
    pub b1:    f32,
    pub b2:    f32,
}

pub struct StatefulParams
{
    pub m: Tensor,
    pub v: Tensor,
    pub t: Tensor, // shape: []
}

impl<'a> super::Optimizer<'a> for Adam<'a>
{
    fn compute_updates<T: AsRef<Tensor>>(
        &mut self,
        params: &[&'a Tensor],
        grads: &[T],
    ) -> Vec<Tensor>
    {
        params
            .into_iter()
            .zip(grads)
            .map(|(param, grad)| {
                let op = AdamOp {
                    static_params: StaticParams {
                        alpha: self.alpha,
                        eps:   self.eps,
                        b1:    self.b1,
                        b2:    self.b2,
                    },
                };

                if let Some(ref param_arr) = param.persistent_array {
                    match self.stateful_params.entry(super::StateKey(param)) {
                        Entry::Vacant(ent) => {
                            let StatefulParams {
                                ref m,
                                ref v,
                                ref t,
                            } = *ent.insert(StatefulParams {
                                m: ::ops::variable(NdArray::zeros(param_arr.shape())),
                                v: ::ops::variable(NdArray::zeros(param_arr.shape())),
                                t: ::ops::variable(::ndarray_ext::from_scalar(1.)),
                            });
                            Tensor::builder()
                                .set_inputs(vec![param, grad.as_ref(), m, v, t])
                                .build(op)
                        }
                        Entry::Occupied(ent) => {
                            let StatefulParams {
                                ref m,
                                ref v,
                                ref t,
                            } = *ent.get();
                            Tensor::builder()
                                .set_inputs(vec![param, grad.as_ref(), m, v, t])
                                .build(op)
                        }
                    }
                } else {
                    panic!("Can't optimize non-variable.")
                }
            })
            .collect()
    }
}