lightly-ai/lightly-train

Provides end-to-end vision model training with support for self-supervised pretraining (DINOv2/v3), distillation from foundation models, and task-specific fine-tuning for detection, segmentation, and panoptic tasks using transformer backbones optimized for edge deployment via ONNX/TensorRT export. Built on PyTorch with unified APIs across vision tasks, enabling seamless transitions from unlabeled data pretraining to production-ready inference on embedded devices with models ranging from 1M to 315M parameters.

1,359 stars. Actively maintained with 44 commits in the last 30 days.