lightly-train

Why we included this project

LightlyTrain keeps pretraining, fine-tuning, and distillation in one framework, so a team can start with a DINOv2 or DINOv3 foundation model trained on its own unlabeled data, then adapt it for detection or segmentation, or distill a large teacher into a smaller student that fits edge hardware. That is useful when the training data is proprietary or the target device cannot host a heavy backbone. It includes ready-to-run recipes for YOLO-style detectors, DETR variants, and smaller models, with ONNX and TensorRT export built in, so trained networks reach production without extra conversion work. Teams that would otherwise glue together separate pretraining, fine-tuning, and export tools get a single place to compare accuracy against latency.