training-and-inference

# training-and-inference Use this sub-skill when you need to train, fine-tune, validate, resume, or run inference for this repository's `main.py`-style workflows without reopening the source checkout. ## Use this route for - Fresh training runs and stage-wise fine-tuning. - Checkpoint save/resume, optimizer state recovery, and milestone handling. - Loading pretrained weights with a different final class count. - Validation, multi-clip inference, and result scoring. - DataParallel checkpoint cleanup before reuse or inspection. ## Do not use this route for - Raw video extraction or annotation JSON creation. Use [data-preparation](../data-preparation/SKILL.md). - Repo-wide routing or root skill selection. Use [the root router](../../SKILL.md). - Model conversion or export outside the `main.py` training loop. ## Read first - `references/workflows.md` - `references/model-catalog.md` - `references/troubleshooting.md` - [the root router](../../SKILL.md) - [the root run helper](../../scripts/run_main.py) - [data-preparation](../data-preparation/SKILL.md) when dataset layout is not ready ## Skill-owned scripts - `scripts/evaluate_results.py` — score top-k recognition results against a ground-truth JSON. - `scripts/strip_dataparallel.py` — remove `module.` prefixes from checkpoint state dicts. ## Typical flow 1. Confirm the dataset tree and annotation JSON already exist. 2. Pick a model family and depth from `references/model-catalog.md`. 3. Choose scratch train, resume, or fine-tune flags in `references/workflows.md`. 4. Run the root-level `main.py` wrapper from the repo skill tree. 5. Evaluate results with `scripts/evaluate_results.py` or clean checkpoints with `scripts/strip_dataparallel.py` when needed. ## Cross-links - Dataset preparation and layout prerequisites: [data-preparation](../data-preparation/SKILL.md) - Repo-level routing: [root router](../../SKILL.md) - Root `main.py` wrapper: [run_main helper](../../scripts/run_main.py)