acyclib 0.3.0

ML library for directed acyclic tensor graphs.
Documentation 
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

use std::{collections::HashMap, sync::Arc};

use crate::device::{Device, OperationError, tensor::DenseMatrix};

use super::{
    OptimiserState, WrapOptimiser,
    radam::{RAdam, RAdamParams},
    utils,
};

#[derive(Clone, Debug)]
pub struct RangerLookaheadParams<T> {
    pub inner: T,
    pub alpha: f32,
    pub k: usize,
}

impl<T: Default> Default for RangerLookaheadParams<T> {
    fn default() -> Self {
        Self { inner: T::default(), alpha: 0.5, k: 6 }
    }
}

pub struct RangerLookahead<D: Device, S> {
    inner: S,
    slow_params: DenseMatrix<D>,
    alpha: f32,
    k: usize,
    step: usize,
}

impl<D: Device, S: OptimiserState<D>> OptimiserState<D> for RangerLookahead<D, S> {
    type Params = RangerLookaheadParams<S::Params>;

    fn new(device: Arc<D>, size: usize, params: Self::Params) -> Result<Self, D::DeviceError> {
        Ok(Self {
            inner: S::new(device.clone(), size, params.inner.clone())?,
            slow_params: DenseMatrix::zeroed(device, size, None)?,
            alpha: params.alpha,
            k: params.k,
            step: 0,
        })
    }

    fn update(
        &mut self,
        weights: &mut DenseMatrix<D>,
        grads: &mut DenseMatrix<D>,
        gradient_factor: f32,
        learning_rate: f32,
    ) -> Result<(), OperationError<D::DeviceError>> {
        self.step += 1;
        self.inner.update(weights, grads, gradient_factor, learning_rate)?;

        if self.step % self.k == 0 {
            assert_eq!(weights.single_size, self.slow_params.single_size);
            assert!(self.slow_params.batch_size().is_none());

            self.slow_params.lerp(self.alpha, weights)?;
            weights.copy_from(&self.slow_params)?;
        }

        Ok(())
    }

    fn reset(&mut self) -> Result<(), D::DeviceError> {
        self.inner.reset()?;
        self.step = 0;
        Ok(())
    }

    fn set_params(&mut self, params: Self::Params) {
        self.inner.set_params(params.inner);
        self.alpha = params.alpha;
        self.k = params.k;
        self.step = 0;
    }

    fn load_from_checkpoint(
        map: &mut HashMap<String, &mut Self>,
        path: &str,
        old_format: bool,
    ) -> Result<(), OperationError<D::DeviceError>> {
        let slow_params = utils::load_weights_from_file(&format!("{path}/slow.bin"), old_format);

        for (id, par) in &slow_params {
            let single = map.get_mut(id).unwrap();
            single.slow_params.load_from_slice(None, par)?;
        }

        let mut map = map.iter_mut().map(|(id, single)| (id.clone(), &mut single.inner)).collect();
        S::load_from_checkpoint(&mut map, path, old_format)
    }

    fn write_to_checkpoint(map: &HashMap<String, &Self>, path: &str) -> Result<(), D::DeviceError> {
        let slow_params: Vec<_> = map.iter().map(|(id, single)| (id, &single.slow_params)).collect();
        utils::write_weights_to_file(&slow_params, &format!("{path}/slow.bin"))?;

        let map = map.iter().map(|(id, single)| (id.clone(), &single.inner)).collect();
        S::write_to_checkpoint(&map, path)
    }
}

pub type Ranger<D> = WrapOptimiser<RangerLookahead<D, RAdam<D>>, RangerParams>;

#[derive(Clone, Copy, Debug)]
pub struct RangerParams {
    pub decay: f32,
    pub beta1: f32,
    pub beta2: f32,
    pub min_weight: f32,
    pub max_weight: f32,
    pub alpha: f32,
    pub k: usize,
}

impl Default for RangerParams {
    fn default() -> Self {
        RangerParams { decay: 0.01, beta1: 0.99, beta2: 0.999, min_weight: -1.98, max_weight: 1.98, alpha: 0.5, k: 6 }
    }
}

impl From<RangerParams> for RangerLookaheadParams<RAdamParams> {
    fn from(value: RangerParams) -> Self {
        RangerLookaheadParams {
            inner: RAdamParams {
                beta1: value.beta1,
                beta2: value.beta2,
                n_sma_threshold: 5.0,
                decay: value.decay,
                clip: Some((value.min_weight, value.max_weight)),
            },
            alpha: value.alpha,
            k: value.k,
        }
    }
}