levyflux/AViD

What it solves

AViD provides a practical workflow for fine-tuning Grounding DINO, an open-vocabulary object detector, on custom datasets. It allows teams to adapt a large detector to domain-specific objects without the computational cost of updating every model parameter.

How it works

AViD implements parameter-efficient fine-tuning (PEFT) using LoRA (Low-Rank Adaptation), targeting the detector's attention, projection, FFN, bounding-box, and feature-map layers. To stabilize training, it includes optional Exponential Moving Averages (EMA). The system uses a CSV-based dataset reader and provides a complete pipeline including YAML-driven training, checkpointing, and a dedicated evaluation loop for metrics like mAP, precision, and recall.

Who it’s for

It is designed for developers and AI teams who need to train a vision-language detector on specialized objects but want to avoid full-parameter fine-tuning.

Highlights

• Parameter-Efficient Adaptation: Uses LoRA to target specific layers for faster, lighter training.
• End-to-End Workflow: Integrates dataset loading, training, validation, and evaluation in one repository.
• Training Stabilization: Optional EMA support to ensure smoother model trajectories.
• Detection-Focused Evaluation: Provides detailed per-class summaries and visual comparisons.
• Flexible Entry Points: Includes YAML configurations, a Gradio demo, and single-image inference scripts.