ruda-optim 0.21.13

Ruda optimizer updates, gradient state, clipping and learning-rate schedules.
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


use super::GradientsParams;
use ruda_model::module::{AutodiffModule, ModuleVisitor, Param, ParamId};
use ruda_model::tensor::{Tensor, backend::AutodiffBackend};
use core::marker::PhantomData;

#[cfg(not(feature = "std"))]
use alloc::vec::Vec;

#[derive(new)]
pub struct GradientsParamsConverter<'a, M: AutodiffModule<B>, B: AutodiffBackend> {
    grads: &'a mut B::Gradients,
    grads_params: &'a mut GradientsParams,
    phatom: PhantomData<M>,
    filter: Option<Vec<ParamId>>,
}

#[derive(new)]
pub struct GradientsParamsChangeDevice<'a, M: AutodiffModule<B>, B: AutodiffBackend> {
    device: &'a B::Device,
    grads: &'a mut GradientsParams,
    phatom: PhantomData<M>,
}

impl<B, M> ModuleVisitor<B> for GradientsParamsConverter<'_, M, B>
where
    B: AutodiffBackend,
    M: AutodiffModule<B>,
{
    fn visit_float<const D: usize>(&mut self, param: &Param<Tensor<B, D>>) {
        if let Some(filter) = self.filter.as_ref()
            && !filter.contains(&param.id)
        {
            return;
        }

        let Some(grad) = param.val().grad_remove(self.grads) else {
            return;
        };

        let grad = match self.grads_params.remove::<B::InnerBackend, D>(param.id) {
            Some(previous) => previous.add(grad),
            None => grad,
        };
        self.grads_params
            .register::<B::InnerBackend, D>(param.id, grad);
    }
}

impl<B, M> ModuleVisitor<B> for GradientsParamsChangeDevice<'_, M, B>
where
    B: AutodiffBackend,
    M: AutodiffModule<B>,
{
    fn visit_float<const D: usize>(&mut self, param: &Param<Tensor<B, D>>) {
        let Some(grad) = self.grads.remove::<B::InnerBackend, D>(param.id) else {
            return;
        };

        self.grads
            .register::<B::InnerBackend, D>(param.id, grad.to_device(self.device));
    }
}