ModelForge

A scaffolded computer vision project inspired by YOLO and DETR. The repository is structured for rapid experimentation with detection models, covering data preparation, model definitions, training, evaluation, and inference.

Project layout

ModelForge/
├── data/                 # Datasets, transforms, and dataloaders
├── models/               # Backbone, heads, YOLO/DETR implementations
├── training/             # Trainers, losses, schedulers, configs
├── scripts/              # CLI entry points for train/eval/test/inference
├── logs/                 # Placeholder for training logs
├── output/               # Model checkpoints and results
├── requirements.txt      # Python dependencies
└── README.md

Key components

• Data pipeline (data/): Includes dataset wrappers, composable transforms, and dataloader creation helpers for COCO-style annotations.
• Models (models/): Minimal YOLO-style and DETR-style architectures with shared ResNet backbones and task-specific heads.
• Training loop (training/): Trainer class, loss dispatching, learning-rate scheduler helpers, and central configuration in config.py.
• Scripts (scripts/): Ready-to-run entrypoints for training, validation, testing, and inference demonstrations.

Getting started

1. Install dependencies:

   pip install -r requirements.txt

2. Prepare your dataset under data/train, data/val, and data/test, and supply annotation dictionaries to the loaders.

3. Adjust hyperparameters in training/config.py.

4. Launch training:

   python scripts/train.py

The provided code uses lightweight stubs for loss and evaluation to illustrate the project flow; extend these components with task-specific logic for real-world use.