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
160
161
162
163
164
165
166
|
//! Various optimization algorithms.
use candle::{Result, Tensor, Var};
/// Optimizer for Stochastic Gradient Descent.
///
/// Contrary to the PyTorch implementation of SGD, this version does not support momentum.
#[derive(Debug)]
pub struct SGD {
vars: Vec<Var>,
learning_rate: f64,
}
impl SGD {
pub fn from_slice(vars: &[&Var], learning_rate: f64) -> Self {
let vars: Vec<_> = vars.iter().map(|&v| v.clone()).collect();
Self {
vars,
learning_rate,
}
}
pub fn new(vars: Vec<Var>, learning_rate: f64) -> Self {
Self {
vars,
learning_rate,
}
}
pub fn empty(learning_rate: f64) -> Self {
Self {
vars: vec![],
learning_rate,
}
}
pub fn into_inner(self) -> Vec<Var> {
self.vars
}
pub fn learning_rate(&self) -> f64 {
self.learning_rate
}
pub fn push(&mut self, var: &Var) {
self.vars.push(var.clone())
}
pub fn step(&self, grads: &candle::backprop::GradStore) -> Result<()> {
for var in self.vars.iter() {
if let Some(grad) = grads.get(var) {
var.set(&var.sub(&(grad * self.learning_rate)?)?)?;
}
}
Ok(())
}
pub fn backward_step(&self, loss: &Tensor) -> Result<()> {
let grads = loss.backward()?;
self.step(&grads)
}
}
#[derive(Clone, Debug)]
pub struct ParamsAdamW {
pub lr: f64,
pub beta1: f64,
pub beta2: f64,
pub eps: f64,
pub weight_decay: f64,
}
impl Default for ParamsAdamW {
fn default() -> Self {
Self {
lr: 0.001,
beta1: 0.9,
beta2: 0.999,
eps: 1e-8,
weight_decay: 0.01,
}
}
}
#[derive(Debug)]
struct VarAdamW {
var: Var,
first_moment: Var,
second_moment: Var,
}
#[derive(Debug)]
pub struct AdamW {
vars: Vec<VarAdamW>,
step_t: usize,
params: ParamsAdamW,
}
impl AdamW {
pub fn new(vars: Vec<Var>, params: ParamsAdamW) -> Result<Self> {
let vars = vars
.into_iter()
.map(|var| {
let dtype = var.dtype();
let shape = var.shape();
let device = var.device();
let first_moment = Var::zeros(shape, dtype, device)?;
let second_moment = Var::zeros(shape, dtype, device)?;
Ok(VarAdamW {
var,
first_moment,
second_moment,
})
})
.collect::<Result<Vec<_>>>()?;
Ok(Self {
vars,
params,
step_t: 0,
})
}
pub fn new_lr(vars: Vec<Var>, learning_rate: f64) -> Result<Self> {
let params = ParamsAdamW {
lr: learning_rate,
..ParamsAdamW::default()
};
Self::new(vars, params)
}
pub fn step(&mut self, grads: &candle::backprop::GradStore) -> Result<()> {
self.step_t += 1;
let lr = self.params.lr;
let lambda = self.params.weight_decay;
let lr_lambda = lr * lambda;
let beta1 = self.params.beta1;
let beta2 = self.params.beta2;
let scale_m = 1f64 / (1f64 - beta1.powi(self.step_t as i32));
let scale_v = 1f64 / (1f64 - beta2.powi(self.step_t as i32));
for var in self.vars.iter() {
let theta = &var.var;
let m = &var.first_moment;
let v = &var.second_moment;
if let Some(g) = grads.get(theta) {
// This involves locking 3 RWLocks per params, if the parameters are large this
// should not be an issue but this may be problematic with models with lots of
// small parameters.
let next_m = ((m.as_tensor() * beta1)? + (g * (1.0 - beta1))?)?;
let next_v = ((v.as_tensor() * beta2)? + (g.sqr()? * (1.0 - beta2))?)?;
let m_hat = (&next_m * scale_m)?;
let v_hat = (&next_v * scale_v)?;
let next_theta = (theta.as_tensor() * (1f64 - lr_lambda))?;
let adjusted_grad = (m_hat / (v_hat.sqrt()? + self.params.eps)?)?;
let next_theta = (next_theta - (adjusted_grad * lr)?)?;
m.set(&next_m)?;
v.set(&next_v)?;
theta.set(&next_theta)?;
}
}
Ok(())
}
pub fn backward_step(&mut self, loss: &Tensor) -> Result<()> {
let grads = loss.backward()?;
self.step(&grads)
}
}
|
