ruda-optim 0.21.51

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


use ruda_model::config::Config;
use ruda_model::record::Record;
use ruda_model::tensor::backend::Backend;
use ruda_model::tensor::{ElementConversion, Tensor};

/// Configuration to create [momentum](Momentum).
#[derive(Config, Debug)]
pub struct MomentumConfig {
    /// Momentum factor
    #[config(default = 0.9)]
    pub momentum: f64,
    /// Dampening factor.
    #[config(default = 0.1)]
    pub dampening: f64,
    /// Enables Nesterov momentum, see [On the importance of initialization and
    /// momentum in deep learning](http://www.cs.toronto.edu/~hinton/absps/momentum.pdf).
    #[config(default = false)]
    pub nesterov: bool,
}

/// State of [momentum](Momentum).
#[derive(Record, Clone, new)]
pub struct MomentumState<B: Backend, const D: usize> {
    velocity: Tensor<B, D>,
}

impl<B:Backend,const D:usize> super::OptimizerCheckpointBuffers<B,D> for MomentumState<B,D> {
    fn visit_checkpoint_buffers<F:FnMut(&Tensor<B,D>)>(&self,visit:&mut F) {visit(&self.velocity);}
    fn map_checkpoint_buffers<F:FnMut(Tensor<B,D>)->Tensor<B,D>>(self,map:&mut F) -> Self {Self {velocity:map(self.velocity)}}
}
impl<B:Backend,const D:usize> super::FlatOptimizerCheckpointState<B,D> for MomentumState<B,D> {
    type FlatState=MomentumState<B,1>;
    fn into_flat_shard(self,shard:&super::FlatOptimizerTensorShard) -> Result<Self::FlatState,super::OptimizerShardError> {
        Ok(MomentumState {velocity:shard.partition(self.velocity)?})
    }
}
impl<B:Backend,const D:usize> super::OptimizerCheckpointScalars for MomentumState<B,D> {
    type Scalars=();
    fn checkpoint_scalars(&self) {}
}

/// Momentum implementation that transforms gradients.
#[derive(Clone)]
pub struct Momentum<B: Backend> {
    momentum: B::FloatElem,
    dampening: f64,
    nesterov: bool,
}

impl<B: Backend> Momentum<B> {
    /// Creates a new [momentum](Momentum) from a [config](MomentumConfig).
    pub fn new(config: &MomentumConfig) -> Self {
        Self {
            momentum: config.momentum.elem(),
            dampening: config.dampening,
            nesterov: config.nesterov,
        }
    }

    /// Transforms a gradient.
    ///
    /// # Arguments
    ///
    /// * `grad` - Gradient to transform.
    /// * `state` - State of the optimizer.
    ///
    /// # Returns
    ///
    /// * `grad` - Transformed gradient.
    /// * `state` - State of the optimizer.
    pub fn transform<const D: usize>(
        &self,
        grad: Tensor<B, D>,
        state: Option<MomentumState<B, D>>,
    ) -> (Tensor<B, D>, MomentumState<B, D>) {
        let velocity = if let Some(state) = state {
            grad.clone()
                .mul_scalar(1.0 - self.dampening)
                .add(state.velocity.mul_scalar(self.momentum))
        } else {
            grad.clone()
        };

        let grad = match self.nesterov {
            true => velocity.clone().mul_scalar(self.momentum).add(grad),
            false => velocity.clone(),
        };

        (grad, MomentumState::new(velocity))
    }
}

impl<B: Backend, const D: usize> MomentumState<B, D> {
    /// Read the momentum buffer without changing its ownership or contents.
    pub fn velocity(&self) -> &Tensor<B, D> {
        &self.velocity
    }

    /// Moves the state to a device.
    ///
    /// # Arguments
    ///
    /// * `device` - Device to move the state to.
    ///
    /// # Returns
    ///
    /// * `self` - Moved state.
    pub fn to_device(mut self, device: &B::Device) -> Self {
        self.velocity = self.velocity.to_device(device);
        self
    }
}