Struct LearnerBuilder

pub struct LearnerBuilder<B, M, O, S, TI, VI, TO, VO>
where
    B: AutodiffBackend,
    M: AutodiffModule<B> + TrainStep<TI, TO> + Display + 'static,
    <M as AutodiffModule<B>>::InnerModule: ValidStep<VI, VO>,
    O: Optimizer<M, B>,
    S: LrScheduler,
    TI: Send + 'static,
    VI: Send + 'static,
    TO: ItemLazy + 'static,
    VO: ItemLazy + 'static,
{ /* private fields */ }

Struct to configure and create a learner.

The generics components of the builder should probably not be set manually, as they are optimized for Rust type inference.

Implementations

impl<B, M, O, S, TI, VI, TO, VO> LearnerBuilder<B, M, O, S, TI, VI, TO, VO>

where B: AutodiffBackend, M: AutodiffModule<B> + TrainStep<TI, TO> + Display + 'static, <M as AutodiffModule<B>>::InnerModule: ValidStep<VI, VO>, O: Optimizer<M, B>, S: LrScheduler, TI: Send + 'static, VI: Send + 'static, TO: ItemLazy + 'static, VO: ItemLazy + 'static,

pub fn new( directory: impl AsRef<Path>, ) -> LearnerBuilder<B, M, O, S, TI, VI, TO, VO>

Creates a new learner builder.

Arguments

• directory - The directory to save the checkpoints.

pub fn metric_loggers<MT, MV>( self, logger_train: MT, logger_valid: MV, ) -> LearnerBuilder<B, M, O, S, TI, VI, TO, VO>

where MT: MetricLogger + 'static, MV: MetricLogger + 'static,

Replace the default metric loggers with the provided ones.

Arguments

• logger_train - The training logger.
• logger_valid - The validation logger.

pub fn with_checkpointing_strategy<CS>( self, strategy: CS, ) -> LearnerBuilder<B, M, O, S, TI, VI, TO, VO>

where CS: CheckpointingStrategy + 'static,

Update the checkpointing_strategy.

pub fn renderer<MR>( self, renderer: MR, ) -> LearnerBuilder<B, M, O, S, TI, VI, TO, VO>

where MR: MetricsRenderer + 'static,

Replace the default CLI renderer with a custom one.

Arguments

• renderer - The custom renderer.

pub fn metrics<Me>( self, metrics: Me, ) -> LearnerBuilder<B, M, O, S, TI, VI, TO, VO>

where Me: MetricRegistration<B, M, O, S, TI, VI, TO, VO>,

Register all metrics as numeric for the training and validation set.

pub fn metrics_text<Me>( self, metrics: Me, ) -> LearnerBuilder<B, M, O, S, TI, VI, TO, VO>

where Me: TextMetricRegistration<B, M, O, S, TI, VI, TO, VO>,

Register all metrics as numeric for the training and validation set.

pub fn metric_train<Me>( self, metric: Me, ) -> LearnerBuilder<B, M, O, S, TI, VI, TO, VO>

where Me: Metric + 'static, <TO as ItemLazy>::ItemSync: Adaptor<<Me as Metric>::Input>,

Register a training metric.

pub fn metric_valid<Me>( self, metric: Me, ) -> LearnerBuilder<B, M, O, S, TI, VI, TO, VO>

where Me: Metric + 'static, <VO as ItemLazy>::ItemSync: Adaptor<<Me as Metric>::Input>,

Register a validation metric.

pub fn grads_accumulation( self, accumulation: usize, ) -> LearnerBuilder<B, M, O, S, TI, VI, TO, VO>

Enable gradients accumulation.

Notes

When you enable gradients accumulation, the gradients object used by the optimizer will be the sum of all gradients generated by each backward pass. It might be a good idea to reduce the learning to compensate.

The effect is similar to increasing the batch size and the learning rate by the accumulation amount.

pub fn metric_train_numeric<Me>( self, metric: Me, ) -> LearnerBuilder<B, M, O, S, TI, VI, TO, VO>

where Me: Metric + Numeric + 'static, <TO as ItemLazy>::ItemSync: Adaptor<<Me as Metric>::Input>,

Register a numeric training metric.

pub fn metric_valid_numeric<Me>( self, metric: Me, ) -> LearnerBuilder<B, M, O, S, TI, VI, TO, VO>

where Me: Metric + Numeric + 'static, <VO as ItemLazy>::ItemSync: Adaptor<<Me as Metric>::Input>,

Register a numeric validation metric.

pub fn num_epochs( self, num_epochs: usize, ) -> LearnerBuilder<B, M, O, S, TI, VI, TO, VO>

The number of epochs the training should last.

pub fn learning_strategy( self, learning_strategy: LearningStrategy<B>, ) -> LearnerBuilder<B, M, O, S, TI, VI, TO, VO>

Run the training loop with different strategies

pub fn checkpoint( self, checkpoint: usize, ) -> LearnerBuilder<B, M, O, S, TI, VI, TO, VO>

The epoch from which the training must resume.

pub fn interrupter(&self) -> Interrupter

Provides a handle that can be used to interrupt training.

pub fn with_interrupter( self, interrupter: Interrupter, ) -> LearnerBuilder<B, M, O, S, TI, VI, TO, VO>

Override the handle for stopping training with an externally provided handle

pub fn early_stopping<Strategy>( self, strategy: Strategy, ) -> LearnerBuilder<B, M, O, S, TI, VI, TO, VO>

where Strategy: EarlyStoppingStrategy + Clone + Send + Sync + 'static,

Register an early stopping strategy to stop the training when the conditions are meet.

pub fn with_application_logger( self, logger: Option<Box<dyn ApplicationLoggerInstaller>>, ) -> LearnerBuilder<B, M, O, S, TI, VI, TO, VO>

By default, Rust logs are captured and written into experiment.log. If disabled, standard Rust log handling will apply.

pub fn with_file_checkpointer<FR>( self, recorder: FR, ) -> LearnerBuilder<B, M, O, S, TI, VI, TO, VO>

where FR: FileRecorder<B> + 'static + FileRecorder<<B as AutodiffBackend>::InnerBackend>, <O as Optimizer<M, B>>::Record: 'static, <M as Module<B>>::Record: 'static, <S as LrScheduler>::Record<B>: 'static,

Register a checkpointer that will save the optimizer, the model and the scheduler to different files.

pub fn summary(self) -> LearnerBuilder<B, M, O, S, TI, VI, TO, VO>

Enable the training summary report.

The summary will be displayed after .fit(), when the renderer is dropped.

pub fn build( self, model: M, optim: O, lr_scheduler: S, ) -> Learner<LearnerComponentsMarker<B, S, M, O, AsyncCheckpointer<<M as Module<B>>::Record, B>, AsyncCheckpointer<<O as Optimizer<M, B>>::Record, B>, AsyncCheckpointer<<S as LrScheduler>::Record<B>, B>, AsyncProcessorTraining<FullEventProcessorTraining<TO, VO>>, Box<dyn CheckpointingStrategy>, LearningDataMarker<TI, VI, TO, VO>>>

where <M as Module<B>>::Record: 'static, <O as Optimizer<M, B>>::Record: 'static, <S as LrScheduler>::Record<B>: 'static,

Create the learner from a model and an optimizer. The learning rate scheduler can also be a simple learning rate.