burn-train 0.20.1

Training crate for the Burn framework
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

use crate::ddp::epoch::{DdpTrainEpoch, DdpValidEpoch};
use crate::ddp::strategy::WorkerComponents;
use crate::{
    Learner, LearningCheckpointer, LearningComponentsTypes, SupervisedTrainingEventProcessor,
    TrainLoader, TrainingBackend, ValidLoader,
};
use burn_collective::{self, CollectiveConfig, PeerId};
use burn_core::tensor::Device;
use burn_core::tensor::backend::AutodiffBackend;
use std::sync::{Arc, Mutex};
use std::thread::JoinHandle;

/// A worker runs the model, syncing gradients using collective operations.
/// Event processing and validation is optional too.
pub(crate) struct DdpWorker<LC>
where
    LC: LearningComponentsTypes + Send + 'static,
{
    peer_id: PeerId,
    device: Device<TrainingBackend<LC>>,
    learner: Learner<LC>,
    event_processor: Arc<Mutex<SupervisedTrainingEventProcessor<LC>>>,
    components: WorkerComponents,
    checkpointer: Option<LearningCheckpointer<LC>>,
    dataloader_train: TrainLoader<LC>,
    dataloader_valid: Option<ValidLoader<LC>>,
    collective_config: CollectiveConfig,
    starting_epoch: usize,
    peer_count: usize,
    is_main: bool,
}

impl<LC> DdpWorker<LC>
where
    LC: LearningComponentsTypes + Send + 'static,
{
    /// Starts a worker that runs the model in a data distributed parallel
    #[allow(clippy::too_many_arguments)]
    pub fn start(
        peer_id: PeerId,
        device: Device<TrainingBackend<LC>>,
        learner: Learner<LC>,
        event_processor: Arc<Mutex<SupervisedTrainingEventProcessor<LC>>>,
        components: WorkerComponents,
        checkpointer: Option<LearningCheckpointer<LC>>,
        dataloader_train: TrainLoader<LC>,
        dataloader_valid: Option<ValidLoader<LC>>,
        collective_config: CollectiveConfig,
        starting_epoch: usize,
        peer_count: usize,
        is_main: bool,
    ) -> JoinHandle<<LC as LearningComponentsTypes>::TrainingModel> {
        let worker = Self {
            peer_id,
            device,
            learner,
            event_processor,
            components,
            checkpointer,
            dataloader_train,
            dataloader_valid,
            collective_config,
            starting_epoch,
            peer_count,
            is_main,
        };

        std::thread::spawn(|| worker.fit())
    }

    /// Fits the model,
    pub fn fit(mut self) -> <LC as LearningComponentsTypes>::TrainingModel {
        burn_collective::register::<<TrainingBackend<LC> as AutodiffBackend>::InnerBackend>(
            self.peer_id,
            self.device.clone(),
            self.collective_config.clone(),
        )
        .expect("Couldn't register for collective operations!");

        let num_epochs = self.components.num_epochs;
        let interrupter = self.components.interrupter;

        // Changed the train epoch to keep the dataloaders
        let epoch_train = DdpTrainEpoch::<LC>::new(
            self.dataloader_train.clone(),
            num_epochs,
            self.components.grad_accumulation,
        );
        let epoch_valid = self
            .dataloader_valid
            .map(|dataloader| DdpValidEpoch::<LC>::new(dataloader, num_epochs));
        self.learner.fork(&self.device);

        for epoch in self.starting_epoch..num_epochs + 1 {
            epoch_train.run(
                &mut self.learner,
                epoch,
                self.event_processor.clone(),
                &interrupter,
                self.peer_id,
                self.peer_count,
                self.is_main,
            );

            if interrupter.should_stop() {
                break;
            }

            // Validation
            if let Some(runner) = &epoch_valid {
                let mut event_processor = self.event_processor.lock().unwrap();
                runner.run(
                    &self.learner.model(),
                    epoch,
                    &mut event_processor,
                    &interrupter,
                );
            }

            if let Some(checkpointer) = &mut self.checkpointer {
                checkpointer.checkpoint(&self.learner, epoch, &self.components.event_store);
            }

            if let Some(early_stopping) = &mut self.components.early_stopping
                && early_stopping.should_stop(epoch, &self.components.event_store)
            {
                break;
            }
        }

        self.learner.model()
    }
}