local-vision-trainer

What problem does it solve?

Train and fine-tune computer vision models locally on CUDA GPUs, removing cloud training costs and data transfer bottlenecks.

Core Features & Use Cases

• Local training of object detection and image classification on a single workstation with a CUDA GPU.
• MegaDetector fine-tuning support (MD1000-larch, MD1000-sorrel, MDV6-rtdetr-c) via train_megadetector.py.
• Support for diverse architectures: YOLOv8 detection, RTDETRv2, YOLOS, DETR for detection; ViT, DINOv2, MobileViT, ResNet for classification; Mask2Former and SegFormer for segmentation.
• Experiment logging to WandB and TensorBoard, with deterministic training options and GPU-friendly configurations.
• Practical workflow: dataset YAML format (nc, names) and a robust three-phase training strategy to preserve MegaDetector knowledge while fine-tuning.

Quick Start

Run the training with a dataset YAML and pretrained weights to start local vision model fine-tuning on your CUDA GPU.