Logging Reference

March 24, 2026 ยท View on GitHub

This page summarizes how training logging works in nnU-Net v2.

Logging architecture

Training logs are routed through MetaLogger, which fans out to:

  • LocalLogger: always enabled
  • optional external loggers such as WandbLogger

The local logger is the source of truth for training curves and progress.png.

What is logged by default

Per epoch, nnU-Net stores values such as:

  • mean_fg_dice
  • ema_fg_dice
  • dice_per_class_or_region
  • train_losses
  • val_losses
  • lrs
  • epoch start and end timestamps

These values drive progress.png in the fold output folder.

Resume behavior

When checkpoints are saved or resumed, local logging state is saved and restored too, so training curves remain continuous.

Enable Weights & Biases

  1. Install W&B:
pip install wandb
  1. Set the environment variables:
export nnUNet_wandb_enabled=1
export nnUNet_wandb_project=nnunet
export nnUNet_wandb_mode=online
  1. Run training normally:
nnUNetv2_train DATASET_NAME_OR_ID 3d_fullres 0

nnUNet_wandb_mode can also be offline.

Custom loggers

A custom logger must implement the minimal interface expected by MetaLogger:

  • update_config(self, config: dict)
  • log(self, key, value, step: int)
  • log_summary(self, key, value)

If you add a brand-new logged key, make sure the local logger knows about it as well.

Detailed legacy page