Menu
How to load a checkpoint and run evaluation (inference only, no training)
Installer avec Codex ou Claude Copiez ce prompt, collez-le dans Codex, Claude ou un autre assistant, puis laissez-le vérifier la page du skill et l'installer pour vous.
Basé sur la classification professionnelle SOC
Git operations: branching, commit message format, pre-commit hook handling, and push policy
Full workflow: environment setup, running experiments, reusing and implementing new components (models, datasets, plmodules, callbacks), and testing
Reference code folder conventions — storing external code and documentation for AI-assisted development context
Coding style guidelines and common pitfalls for this project
Hydra configuration file structure and conventions for models, datasets, callbacks, and training
Data format conventions for datasets and plmodules
| name | load-eval |
| description | How to load a checkpoint and run evaluation (inference only, no training) |
To evaluate a trained model without training, pass train=False to src/train.py along with a path to the checkpoint(s).
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"]
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.
train=False skips the training loop and goes straight to validation.paths.restore_ckpts takes precedence over paths.restore_dir if both are set.logs/ structure as a training run (metrics, visualizations, etc.).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