Early stopping for customized training loop

[wangjiawen2013]

Hi,
This is a snippet from burn's guide example, which we can set early_stopping:

    // Model
    let learner = LearnerBuilder::new(artifact_dir)
        .metric_train_numeric(AccuracyMetric::new())
        .metric_valid_numeric(AccuracyMetric::new())
        .with_file_checkpointer(CompactRecorder::new())
        .early_stopping(MetricEarlyStoppingStrategy::new::<LossMetric<B>>(
            Aggregate::Mean,
            Direction::Lowest,
            Split::Valid,
            StoppingCondition::NoImprovementSince { n_epochs: 1 },
        ))
        .devices(vec![device.clone()])
        .num_epochs(config.num_epochs)
        .summary()
        .build(
            config.model.init::<B>(&device),
            config.optimizer.init(),
            config.learning_rate);

But for customized training loop, such as in https://github.com/tracel-ai/burn/blob/main/examples/custom-training-loop/src/lib.rs, how to set early_stopping?

[laggui]

The LearnerBuilder takes an EarlyStoppingStrategy. This is implemented for the MetricEarlyStoppingStrategy.

Then, the Learner simply checks at the end of every epoch if the training should stop based on the strategy.

So if you want to use the same strategy, it should be fairly simple to create the early stopping instance and then check at the end of every epoch in your training loop to break from the loop.