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

//! Adam optimizer
use crate::ops::gradient_descent_ops::adam;
use crate::tensor::{Input, Tensor, Variable};
use crate::Float;
use crate::Graph;
use crate::NdArray;
use std::sync::{Arc, RwLock};

/// Adam optimizer
///
/// This implementation is based on http://arxiv.org/abs/1412.6980v8.
///
/// ```
/// extern crate autograd as ag;
/// use ag::tensor::Variable;
/// use ag::ndarray_ext::into_shared;
/// use std::sync::{Arc, RwLock};
/// use ag::optimizers::adam;
///
/// // Define parameters to optimize.
/// let rng = ag::ndarray_ext::ArrayRng::<f32>::default();
/// let w: Arc<RwLock<ag::NdArray<f32>>> = into_shared(rng.glorot_uniform(&[28 * 28, 10]));
/// let b: Arc<RwLock<ag::NdArray<f32>>> = into_shared(ag::ndarray_ext::zeros(&[1, 10]));
/// // Make a state of adam.
/// let state = adam::AdamState::new(&[&w, &b]);
///
/// ag::with(|g| {
///     let w_ = g.variable(w.clone());
///     let b_ = g.variable(b.clone());
///
///     // some operations using w_ and b_
///     // let y = ...
///     // let grads = ag::grad(&[y], &[w_, b_]);
///
///     // instantiate an adam optimizer with default setting.
///     let adam = adam::Adam::<f32>::default();
///
///     // Getting update ops of `params` using its gradients and adam.
///     // let update_ops: &[ag::Tensor<f32>] = &adam.compute_updates(&params, &grads, &state, &g);
/// });
/// ```
///
/// See also https://github.com/raskr/rust-autograd/blob/master/examples/
pub struct Adam<F: Float> {
    static_params: StaticParams<F>,
}

impl<T: Float> Default for Adam<T> {
    /// Instantiates `Adam` optimizer with the recommended parameters in the original paper.
    fn default() -> Adam<T> {
        let static_params = StaticParams {
            alpha: T::from(0.001).unwrap(),
            eps: T::from(1e-08).unwrap(),
            b1: T::from(0.9).unwrap(),
            b2: T::from(0.999).unwrap(),
        };
        Adam { static_params }
    }
}

impl<'t, 's: 't, F: Float> Adam<F> {
    /// Instantiates Adam from static params
    pub fn new(static_params: StaticParams<F>) -> Self {
        Adam { static_params }
    }

    /// Creates ops to optimize `params` with Adam.
    ///
    /// Evaluated results of the return values will be `None`.
    pub fn compute_updates(
        &self,
        params: &[Tensor<'s, F>],
        grads: &[Tensor<'s, F>],
        states: &AdamState<F>,
        g: &'s Graph<F>,
    ) -> Vec<Tensor<'s, F>> {
        let num_params = params.len();
        let mut ret = Vec::with_capacity(num_params);
        for i in 0..num_params {
            let param = &params[i];
            let a: *const RwLock<NdArray<F>> = param
                .get_variable_array_ptr()
                .expect("Adam requires *variables* as its inputs.");
            let key = a as usize;
            let state = states
                .var2state
                .get(&key)
                .expect("Adam: state object wasn't fed correctly");
            let m = g.variable(state.m.clone());
            let v = g.variable(state.v.clone());
            let t = g.variable(state.t.clone());

            ret.push(
                Tensor::builder()
                    .set_inputs(&[
                        Input::new_mut(param),
                        Input::new(&grads[i]),
                        Input::new_mut(&m),
                        Input::new_mut(&v),
                        Input::new_mut(&t),
                    ])
                    .build(
                        g,
                        adam::AdamOp {
                            static_params: self.static_params.clone(),
                        },
                    ),
            );
        }
        ret
    }
}

struct StateArrays<F: Float> {
    m: Arc<RwLock<NdArray<F>>>,
    v: Arc<RwLock<NdArray<F>>>,
    t: Arc<RwLock<NdArray<F>>>,
}

/// A state object for an adam optimizer.
pub struct AdamState<F: Float> {
    // map key is the address of a variable array on the heap
    var2state: crate::FxHashMap<usize, StateArrays<F>>,
}

impl<'t, 's: 't, F: Float> AdamState<F> {
    /// Creates a new state object for an adam optimizer.
    ///
    /// `variables` should be variable arrays fed to `autograd::variable`.
    pub fn new(variables: &[&Arc<RwLock<NdArray<F>>>]) -> Self {
        let mut map = crate::FxHashMap::default();
        for &var in variables {
            // Use the address on the heap as a hash key
            let key = ((&**var) as *const RwLock<_>) as usize;
            let var = var.read().unwrap();
            let var_shape = var.shape();
            map.insert(
                key,
                StateArrays {
                    m: Arc::new(RwLock::new(crate::ndarray_ext::zeros(var_shape))),
                    v: Arc::new(RwLock::new(crate::ndarray_ext::zeros(var_shape))),
                    t: Arc::new(RwLock::new(crate::ndarray_ext::from_scalar(F::one()))),
                },
            );
        }
        Self { var2state: map }
    }
}

/// Holds Adam's static parameters (`alpha`, `eps`, `b1`, `b2`).
#[derive(Clone)]
pub struct StaticParams<T: Float> {
    pub alpha: T,
    pub eps: T,
    pub b1: T,
    pub b2: T,
}