ICheckpointManager<T, TInput, TOutput>

Interfaces · AiDotNet.Interfaces

Defines the contract for checkpoint management systems that save and restore training state.

Known implementations: CheckpointManagerBase<T, TInput, TOutput>

How It Works

A checkpoint manager handles saving and restoring the complete state of model training, allowing you to pause and resume training, recover from failures, and track model evolution.

For Beginners: Think of checkpoints like save points in a video game. They let you:

  • Save your progress so you can come back later
  • Go back to an earlier point if something goes wrong
  • Keep the best version you’ve found so far

Checkpoints typically save:

  • Model parameters (weights and biases)
  • Optimizer state (momentum, learning rate schedule, etc.)
  • Training metadata (epoch number, step count)
  • Performance metrics

Why checkpoint management matters:

  • Training can be interrupted (crashes, time limits)
  • You want to keep the best model even if later training makes it worse
  • Long training runs need progress saved periodically
  • Enables experimentation with different training strategies from same point

Methods

Method Summary
Checkpoint<T, TInput, TOutput> LoadBestCheckpoint(string metricName, MetricOptimizationDirection direction) Loads the checkpoint with the best metric value.
Checkpoint<T, TInput, TOutput> LoadCheckpoint(string checkpointId) Loads a checkpoint and restores the training state.
Checkpoint<T, TInput, TOutput> LoadLatestCheckpoint() Loads the most recent checkpoint.
List<CheckpointMetadata<T>> ListCheckpoints(string sortBy, bool descending) Lists all available checkpoints.
bool ShouldAutoSaveCheckpoint(int currentStep, double? metricValue, bool shouldMinimize) Determines whether an automatic checkpoint should be saved based on current configuration.
int CleanupKeepBest(string metricName, int keepBest, MetricOptimizationDirection direction) Deletes checkpoints except the best N according to a metric.
int CleanupOldCheckpoints(int keepLast) Deletes old checkpoints, keeping only a specified number of the most recent ones.
string GetCheckpointDirectory() Gets the storage path for checkpoints.
string SaveCheckpoint(IModel<TInput, TOutput, TMetadata> model, IOptimizer<T, TInput, TOutput> optimizer, int epoch, int step, Dictionary<string, T> metrics, Dictionary<string, object> metadata) Saves a checkpoint of the current training state.
string TryAutoSaveCheckpoint(IModel<TInput, TOutput, TMetadata> model, IOptimizer<T, TInput, TOutput> optimizer, int epoch, int step, Dictionary<string, T> metrics, double? metricValue, bool shouldMinimize, Dictionary<string, object> metadata) Attempts to save a checkpoint automatically based on configured auto-checkpoint settings.
void ConfigureAutoCheckpointing(AutoCheckpointingOptions options) Sets up automatic checkpointing during training using an AutoCheckpointingOptions object, including optional monotonic wall-clock interval saving and explicit metric optimization direction.
void ConfigureAutoCheckpointing(int saveFrequency, int keepLast, bool saveOnImprovement, string metricName) Sets up automatic checkpointing during training.
void DeleteCheckpoint(string checkpointId) Deletes a specific checkpoint.
void UpdateAutoSaveState(int step, double? metricValue, bool shouldMinimize) Updates the auto-save state after a checkpoint is saved.