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

use crate::{
    Learner, LearningComponentsTypes, MultiDeviceOptim, SupervisedLearningStrategy,
    SupervisedTrainingEventProcessor, TrainLoader, TrainingBackend, TrainingComponents,
    TrainingModel, ValidLoader, multi::epoch::MultiDeviceTrainEpoch,
    single::epoch::SingleDeviceValidEpoch,
};
use burn_core::{data::dataloader::split::split_dataloader, tensor::Device};

pub struct MultiDeviceLearningStrategy<LC: LearningComponentsTypes> {
    devices: Vec<Device<TrainingBackend<LC>>>,
    optim: MultiDeviceOptim,
}
impl<LC: LearningComponentsTypes> MultiDeviceLearningStrategy<LC> {
    pub fn new(devices: Vec<Device<TrainingBackend<LC>>>, optim: MultiDeviceOptim) -> Self {
        Self { devices, optim }
    }
}

impl<LC: LearningComponentsTypes> SupervisedLearningStrategy<LC>
    for MultiDeviceLearningStrategy<LC>
{
    fn fit(
        &self,
        training_components: TrainingComponents<LC>,
        mut learner: Learner<LC>,
        dataloader_train: TrainLoader<LC>,
        dataloader_valid: ValidLoader<LC>,
        starting_epoch: usize,
    ) -> (TrainingModel<LC>, SupervisedTrainingEventProcessor<LC>) {
        let main_device = self.devices.first().unwrap();

        // `MultiDevicesTrainStep` has one worker per device, so we use a fixed device strategy
        // for each (worker) data loader. This matches the expected device on the worker, so we
        // don't have to move the data between devices.
        let dataloader_train = split_dataloader(dataloader_train, &self.devices);
        let dataloader_valid = dataloader_valid.to_device(main_device);

        learner.fork(main_device);
        let mut event_processor = training_components.event_processor;
        let mut checkpointer = training_components.checkpointer;
        let mut early_stopping = training_components.early_stopping;
        let num_epochs = training_components.num_epochs;

        let epoch_train = MultiDeviceTrainEpoch::<LC>::new(
            dataloader_train.clone(),
            num_epochs,
            training_components.grad_accumulation,
        );
        let epoch_valid: SingleDeviceValidEpoch<LC> =
            SingleDeviceValidEpoch::new(dataloader_valid.clone(), num_epochs);

        for epoch in starting_epoch..training_components.num_epochs + 1 {
            epoch_train.run(
                &mut learner,
                epoch,
                &mut event_processor,
                &training_components.interrupter,
                self.devices.to_vec(),
                self.optim,
            );

            if training_components.interrupter.should_stop() {
                let reason = training_components
                    .interrupter
                    .get_message()
                    .unwrap_or(String::from("Reason unknown"));
                log::info!("Training interrupted: {reason}");
                break;
            }

            epoch_valid.run(
                &learner,
                epoch,
                &mut event_processor,
                &training_components.interrupter,
            );

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

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

        (learner.model(), event_processor)
    }
}