burn-train 0.19.0

Training crate for the Burn framework
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 crate::checkpoint::{Checkpointer, CheckpointingAction, CheckpointingStrategy};
use crate::components::LearnerComponentTypes;
use crate::metric::store::EventStoreClient;
use crate::{CloneEarlyStoppingStrategy, LearnerSummaryConfig, LearningStrategy};
use burn_core::module::Module;
use burn_core::tensor::Device;
use burn_optim::Optimizer;
use burn_optim::lr_scheduler::LrScheduler;
use std::sync::Arc;
use std::sync::atomic::{AtomicBool, Ordering};

/// Learner struct encapsulating all components necessary to train a Neural Network model.
///
/// To create a learner, use the [builder](crate::learner::LearnerBuilder) struct.
pub struct Learner<LC: LearnerComponentTypes> {
    pub(crate) model: LC::Model,
    pub(crate) optim: LC::Optimizer,
    pub(crate) lr_scheduler: LC::LrScheduler,
    pub(crate) num_epochs: usize,
    pub(crate) checkpoint: Option<usize>,
    pub(crate) grad_accumulation: Option<usize>,
    pub(crate) checkpointer: Option<LearnerCheckpointer<LC>>,
    pub(crate) learning_strategy: LearningStrategy<LC::Backend>,
    pub(crate) interrupter: Interrupter,
    pub(crate) early_stopping: Option<EarlyStoppingStrategyRef>,
    pub(crate) event_processor: LC::EventProcessor,
    pub(crate) event_store: Arc<EventStoreClient>,
    pub(crate) summary: Option<LearnerSummaryConfig>,
}

/// Cloneable reference to an early stopping strategy
pub(crate) type EarlyStoppingStrategyRef = Box<dyn CloneEarlyStoppingStrategy>;

#[derive(new)]
pub(crate) struct LearnerCheckpointer<LC: LearnerComponentTypes> {
    model: LC::CheckpointerModel,
    optim: LC::CheckpointerOptimizer,
    lr_scheduler: LC::CheckpointerLrScheduler,
    strategy: LC::CheckpointerStrategy,
}

impl<LC: LearnerComponentTypes> LearnerCheckpointer<LC> {
    pub(crate) fn checkpoint(
        &mut self,
        model: &LC::Model,
        optim: &LC::Optimizer,
        scheduler: &LC::LrScheduler,
        epoch: usize,
        store: &EventStoreClient,
    ) {
        let actions = self.strategy.checkpointing(epoch, store);

        for action in actions {
            match action {
                CheckpointingAction::Delete(epoch) => {
                    self.model
                        .delete(epoch)
                        .expect("Can delete model checkpoint.");
                    self.optim
                        .delete(epoch)
                        .expect("Can delete optimizer checkpoint.");
                    self.lr_scheduler
                        .delete(epoch)
                        .expect("Can delete learning rate scheduler checkpoint.");
                }
                CheckpointingAction::Save => {
                    self.model
                        .save(epoch, model.clone().into_record())
                        .expect("Can save model checkpoint.");
                    self.optim
                        .save(epoch, optim.to_record())
                        .expect("Can save optimizer checkpoint.");
                    self.lr_scheduler
                        .save(epoch, scheduler.to_record())
                        .expect("Can save learning rate scheduler checkpoint.");
                }
            }
        }
    }

    pub(crate) fn load_checkpoint(
        &self,
        model: LC::Model,
        optim: LC::Optimizer,
        scheduler: LC::LrScheduler,
        device: &Device<LC::Backend>,
        epoch: usize,
    ) -> (LC::Model, LC::Optimizer, LC::LrScheduler) {
        let record = self
            .model
            .restore(epoch, device)
            .expect("Can load model checkpoint.");
        let model = model.load_record(record);

        let record = self
            .optim
            .restore(epoch, device)
            .expect("Can load optimizer checkpoint.");
        let optim = optim.load_record(record);

        let record = self
            .lr_scheduler
            .restore(epoch, device)
            .expect("Can load learning rate scheduler checkpoint.");
        let scheduler = scheduler.load_record(record);

        (model, optim, scheduler)
    }
}

#[derive(Clone, Default)]
/// A handle that allows aborting the training/evaluation process early.
pub struct Interrupter {
    state: Arc<AtomicBool>,
}

impl Interrupter {
    /// Create a new instance.
    pub fn new() -> Self {
        Self::default()
    }

    /// Notify the learner that it should stop.
    pub fn stop(&self) {
        self.state.store(true, Ordering::Relaxed);
    }

    /// Reset the interrupter.
    pub fn reset(&self) {
        self.state.store(false, Ordering::Relaxed);
    }

    /// True if .stop() has been called.
    pub fn should_stop(&self) -> bool {
        self.state.load(Ordering::Relaxed)
    }
}