Enhance Eval/Predict with Auto Checkpoint Mode (Best / Last)

[willyfh]

Overview

Currently, eval.py and predict.py require users to manually provide a checkpoint path. To improve usability, add support for automatic checkpoint selection when no path is provided.

Proposed Behavior

If ckpt_path is explicitly provided → use it (manual mode).

If not, use checkpoint_mode to decide:

• best: Load best checkpoint from the latest experiment run.

• last: Load most recent (latest epoch/step) checkpoint from the latest run.

Example Usage

python eval.py experiment=example_experiment checkpoint_mode=best
# OR with manual override (existing behavior)
python eval.py experiment=example_experiment ckpt_path=path/to/checkpoint.ckpt