load-eval

name	load-eval
description	How to load a checkpoint and run evaluation (inference only, no training)

Load and Eval

To evaluate a trained model without training, pass train=False to src/train.py along with a path to the checkpoint(s).

Option 1: Load a specific checkpoint

uv run python src/train.py --config-name=train_your_project \
    train=False \
    paths.restore_ckpts=["./logs/train/runs/2025-01-01_00-00-00/checkpoints/last.ckpt"]

Option 2: Load all checkpoints in a directory

uv run python src/train.py --config-name=train_your_project \
    train=False \
    paths.restore_dir=./logs/train/runs/2025-01-01_00-00-00/checkpoints

All .ckpt files in the directory will be evaluated in sequence.

How it works

• train=False skips the training loop and goes straight to validation.
• paths.restore_ckpts takes precedence over paths.restore_dir if both are set.
• If neither is set, the model runs with random weights (useful for sanity checks and training with random initialization)
• Results are saved to the same logs/ structure as a training run (metrics, visualizations, etc.).

Defaults

Both restore_ckpts and restore_dir default to null (defined in configs/paths/default.yaml). Override them on the command line as shown above, or set them in configs/train_your_project.yaml for repeated use:

train: False
paths:
  restore_dir: ./logs/train/runs/2025-01-01_00-00-00/checkpoints