Accessing Training Logs Inside a Custom Callback in model.train() (Ultralytics)

Discussion
, ,
1

I’m using a custom training callback with the model.train() API and would like to access the same training logs/metrics that are printed or logged during training (e.g., loss values, metrics per epoch/iteration).

At the moment, the callback system seems to only expose event hooks (e.g., on_train_epoch_end, on_fit_epoch_end, etc.), but not the actual training logs or metrics dictionary that Ultralytics internally generates and prints.

What I’m trying to do

I want to log or process training statistics (e.g., losses, metrics, MaP scores) inside a custom callback, for example to:

  • Send metrics to an external logger

  • Store custom per-epoch analytics

  • Trigger logic based on loss trends

Current limitation

Inside the callback, I don’t see a clean way to access the same values that appear in the training output, such as:

  • Train/val loss components

  • Metrics per epoch

  • MaP values

It appears that:

  • The callback only receives the trainer object

  • The trainer does not expose a stable/public interface for retrieving the current log dictionary

1 Like
2

You can get the same numbers that show up in the console from inside a callback — Ultralytics already does this in the built-in logger integrations by reading fields off the trainer object and composing a metrics dict.

At epoch end, the “current log dict” is effectively:

  • train losses: trainer.label_loss_items(trainer.tloss, prefix="train")
  • val metrics (mAP, etc.): trainer.metrics
  • learning rates: trainer.lr

This is exactly what the DVCLive callback does in on_fit_epoch_end and what the ClearML/W&B callbacks follow as well (see the patterns in the clearml.py reference and wb.py reference).

from ultralytics import YOLO

def on_fit_epoch_end(trainer):
    # after validation, so trainer.metrics is populated
    metrics = {
        **trainer.label_loss_items(trainer.tloss, prefix="train"),
        **trainer.metrics,
        **trainer.lr,
    }
    # send metrics wherever you want
    print(metrics)

model = YOLO("yolo11n.pt")
model.add_callback("on_fit_epoch_end", on_fit_epoch_end)
model.train(data="coco8.yaml", epochs=3)

If you need per-iteration values, use a batch hook like on_train_batch_end and read whatever the trainer exposes at that point (it’s more “internal” and may change across versions), but for stable logging I’d recommend sticking to on_train_epoch_end / on_fit_epoch_end.

Also worth confirming you’re on the latest package since callback event coverage has improved over time: pip install -U ultralytics.

1 Like