Experiments and Outputs

Training entry points

• train.py: primary continual SSL training flow

• pretrain.py: pretraining/evaluation script variant

• tune.py: Optuna-based hyperparameter search

Logs and artifacts

Typical output locations:

• logs/: linear, kNN, and NCM metrics

• checkpoints/: model checkpoints

• lightning_logs/: PyTorch Lightning artifacts

• wandb/: local Weights and Biases files

Common output naming pattern

Metrics folders often follow patterns such as:

• <Model>_linear_<dataset>_<num_tasks>

• <Model>_knn_<dataset>_<num_tasks>

• <Model>_ncm_<dataset>_<num_tasks>

When plugins are active, folder names can include plugin identifiers.