altametris.sara.yolo.callbacks¶

YOLO-specific callbacks.

Specialized callbacks for YOLO training with metrics logging, checkpoint management, and early stopping.

Example

>>> callback = YoloMetricsCallback()
>>> trainer = YoloTrainer(model=model, callbacks=[callback])

Attributes¶

logger

Classes¶

`YoloMetricsCallback`	Callback for logging YOLO training metrics.
`YoloCheckpointCallback`	Callback for saving model checkpoints.
`YoloEarlyStoppingCallback`	Callback for early stopping based on metric plateau.
`YoloLoggingCallback`	Callback for structured logging of training progress.

Module Contents¶

altametris.sara.yolo.callbacks.logger¶

class altametris.sara.yolo.callbacks.YoloMetricsCallback(log_file: pathlib.Path | None = None)¶

Bases: altametris.sara.core.base_callback.BaseCallback

Callback for logging YOLO training metrics.

Logs mAP, precision, recall, and loss metrics during training.

Example

>>> callback = YoloMetricsCallback(log_file="metrics.json")
>>> trainer = YoloTrainer(model=model, callbacks=[callback])

log_file = None¶

metrics_history: list[dict[str, Any]] = []¶

on_epoch_end(epoch: int, metrics: dict[str, Any], **kwargs: Any) → None¶: Log metrics at end of epoch.

_save_metrics() → None¶: Save metrics history to JSON file.

class altametris.sara.yolo.callbacks.YoloCheckpointCallback(save_dir: pathlib.Path, monitor_metric: str = 'mAP50', save_every: int = 10)¶

Bases: altametris.sara.core.base_callback.BaseCallback

Callback for saving model checkpoints.

Saves best model based on specified metric and periodic checkpoints.

Parameters:

save_dir – Directory to save checkpoints
monitor_metric – Metric to monitor for best model (default: “mAP50”)
save_every – Save checkpoint every N epochs

Example

>>> callback = YoloCheckpointCallback(
...     save_dir="checkpoints",
...     monitor_metric="mAP50-95",
...     save_every=10
... )

save_dir¶

monitor_metric = 'mAP50'¶

save_every = 10¶

best_metric = 0.0¶

on_epoch_end(epoch: int, metrics: dict[str, Any], **kwargs: Any) → None¶: Save checkpoints at end of epoch.

_save_checkpoint(model: altametris.sara.core.base_model.BaseModel, path: pathlib.Path, checkpoint_type: str) → None¶: Helper to save checkpoint with error handling.

class altametris.sara.yolo.callbacks.YoloEarlyStoppingCallback(monitor_metric: str = 'mAP50', patience: int = 20, min_delta: float = 0.001)¶

Bases: altametris.sara.core.base_callback.BaseCallback

Callback for early stopping based on metric plateau.

Stops training if monitored metric doesn’t improve for N epochs.

Parameters:

monitor_metric – Metric to monitor
patience – Number of epochs without improvement before stopping
min_delta – Minimum change to qualify as improvement

Example

>>> callback = YoloEarlyStoppingCallback(
...     monitor_metric="val_loss",
...     patience=20,
...     min_delta=0.001
... )

monitor_metric = 'mAP50'¶

patience = 20¶

min_delta = 0.001¶

best_metric = 0.0¶

epochs_without_improvement = 0¶

should_stop = False¶

on_epoch_end(epoch: int, metrics: dict[str, Any], **kwargs: Any) → None¶: Check for early stopping condition.

class altametris.sara.yolo.callbacks.YoloLoggingCallback(log_level: str = 'INFO')¶

Bases: altametris.sara.core.base_callback.BaseCallback

Callback for structured logging of training progress.

Logs training events with timestamps and structured data.

Example

>>> callback = YoloLoggingCallback(log_level="INFO")
>>> trainer = YoloTrainer(model=model, callbacks=[callback])

log_level¶

on_train_start(**kwargs: Any) → None¶: Log training start.

on_train_end(**kwargs: Any) → None¶: Log training end.

on_validation_start(**kwargs: Any) → None¶: Log validation start.

on_validation_end(metrics: dict[str, Any], **kwargs: Any) → None¶: Log validation results.