---

name: data-training description: "Validate datasets, understand augmentation, and plan SSD.PyTorch training commands." metadata: disco-role: operating disable-model-invocation: true license: MIT

data-training

Use this operating sub-skill when a Researcher needs to prepare or sanity-check SSD.PyTorch training inputs before launching a long training run. It covers VOC/COCO dataset layout, annotation transforms, SSD data augmentation, collation, training command planning, checkpoints, base VGG weights, Visdom, and CPU/CUDA handling.

Fast routing

• For dataset directory checks, first read references/datasets-and-augmentation.md, then run the bundled validator:
  • python scripts/validate_dataset_layout.py --dataset voc --root VOCDEVKIT_ROOT --require-train
  • python scripts/validate_dataset_layout.py --dataset coco --root COCO_ROOT --require-train
• For training command construction, read references/training-workflow.md, then generate templates with:
  • python scripts/plan_training_command.py --dataset VOC --dataset-root VOCDEVKIT_ROOT --cuda true
  • python scripts/plan_training_command.py --dataset COCO --dataset-root COCO_ROOT --cuda true
• For startup failures, missing files, import errors, empty targets, or CUDA surprises, read references/troubleshooting.md.

Operating boundaries

This sub-skill owns:

• VOC and COCO training dataset skeletons and split-file expectations.
• VOC/COCO annotation transform outputs and class-count implications.
• BaseTransform, SSDAugmentation, and detection_collate behavior.
• train.py option planning, required files, checkpoints, learning-rate schedule, optional Visdom, and CUDA/CPU flags.
• Safe preflight commands and JSON-producing helper scripts.

Route out of this sub-skill when the request is not about data or training setup:

• Model construction, SSD priors, detection heads, loss internals, and inference model behavior: ../model-inference/SKILL.md.
• Evaluation, test-time detection, demo notebook/webcam usage, and mAP reporting: ../evaluation-demos/SKILL.md.

Non-negotiable safety

• Do not execute dataset download scripts or bundle network download logic. The original dataset scripts are reference-only because they perform large downloads, extraction, copies, and deletes.
• Treat full training as external, data-heavy, and long-running. Only plan commands unless the user explicitly asks to run training in an appropriate Researcher session.
• Keep commands explicit about --dataset_root, --save_folder, --basenet, --resume, --cuda, --batch_size, and --num_workers rather than relying on hidden local defaults.
• Remember the COCO label-map import caveat: importing the package-level data module can fail before argument parsing if the default COCO label map is absent at the home-derived default COCO location.

Minimal workflow

1. Identify target dataset: VOC or COCO.
2. Validate the layout with scripts/validate_dataset_layout.py; fix missing split files, annotation files, images, COCO JSON files, or label-map placement before training.
3. Confirm required training files:
   • Fresh training needs vgg16_reducedfc.pth under the configured save folder.
   • Resume training needs a checkpoint compatible with the selected dataset/model configuration.
4. Generate a command plan with scripts/plan_training_command.py and review the emitted notes.
5. If the user asks to execute training, warn about runtime/data cost, CUDA preference, and dependency/version constraints before running anything.