Skip to content

CLAUDE.md

Latest commit

 

History

History
55 lines (38 loc) · 1.59 KB

CLAUDE.md

File metadata and controls

55 lines (38 loc) · 1.59 KB
Raw

UTMIST ML Seed

Hydra-driven ML experiment launcher. Supports classification (ResNet/CIFAR-10) and detection (YOLOv8).

Key commands

# Train (default: ResNet50 on CIFAR-10)
python src/train.py

# Override from CLI
python src/train.py model=yolo_v8 training.epochs=20 training.lr=0.0005

# Apply experiment preset
python src/train.py +experiment=long_run

# Evaluate checkpoint
python src/evaluate.py checkpoint=path/to/best_model.pt

# TensorBoard (view training progress)
tensorboard --logdir outputs/

# Lint
ruff check src/ tests/

# Test
pytest tests/ -v

Architecture

  • Single entry point: src/train.py (Hydra)
  • Config composition: configs/config.yaml → model + dataset + experiment overrides
  • Model registry: src/models.py — build_model(cfg) dispatches on cfg.name
  • Dataset registry: src/data.py — build_dataset(cfg) dispatches on cfg.name
  • YOLO uses ultralytics' built-in .train(), not the standard training loop

Adding a model

  1. Add config: configs/model/<name>.yaml with name: field
  2. Add builder: _build_<name>(cfg) in src/models.py
  3. Register in build_model() dispatch

Adding a dataset

  1. Add config: configs/dataset/<name>.yaml with name: field
  2. Add builder: _build_<name>(cfg) returning (train_ds, val_ds) in src/data.py
  3. Register in build_dataset() dispatch

Conventions

  • All hyperparameters live in YAML configs, never hardcoded
  • Functions follow build_*() / _build_*() naming
  • Hydra manages output directories (outputs/<date>/<time>/)
  • Seeds set via src/utils.py:set_seed() — covers python, numpy, torch