Appendix A — The Practitioner’s Toolkit

The skills ladder, collected as a reference handbook. Each section corresponds to the lab that introduces it.

A.1 Experiment tracking with Weights & Biases

A.2 Reproducibility discipline

A.3 Datasets and DataLoaders

A.4 Evaluation by hand

A.5 Annotation

A.6 Checkpointing and W&B artifacts

A.7 Mixed precision and GPU memory

A.8 Hyperparameter optimisation

A.9 Hugging Face fluency

A.10 Inference-time engineering

# The Practitioner's Toolkit {#sec-toolkit}

The skills ladder, collected as a reference handbook. Each section corresponds to the lab that introduces it.

## Experiment tracking with Weights & Biases

<!-- TODO: setup, run config logging, the project conventions used in this course -->

## Reproducibility discipline

<!-- TODO: seeding, config-driven runs, the overfit-a-single-batch ritual -->

## Datasets and DataLoaders

<!-- TODO: writing a Dataset class; transforms; num_workers; common pitfalls -->

## Evaluation by hand

<!-- TODO: IoU and mAP implementations; split hygiene -->

## Annotation

<!-- TODO: CVAT/Roboflow workflow; label-quality checks -->

## Checkpointing and W&B artifacts

<!-- TODO: save/resume; artifact versioning; the pattern Lab 13 resumes from -->

## Mixed precision and GPU memory

<!-- TODO: torch.amp; memory math; gradient accumulation -->

## Hyperparameter optimisation

<!-- TODO: W&B Bayesian sweeps; Optuna and pruning; what to sweep and what to fix -->

## Hugging Face fluency

<!-- TODO: models, processors, model cards; pipelines; Trainer; LoRA/PEFT -->

## Inference-time engineering

<!-- TODO: guidance scale, schedulers, steps-vs-quality as logged experiments -->