Tao Train Mask Auto Label

This handles weakly-supervised segmentation when you have point or box annotations instead of full pixel masks. It's built on ViT-MAE and ships with AutoML enabled by default, tuning learning rate and weight decay in a narrow Bayesian range around the fine-tuning defaults. You'll need COCO-style JSON annotations, at least 24GB VRAM per GPU, and patience with the tiny learning rates that come with vision transformers. The crop size lives under dataset.crop_size, not model.crop_size, which trips people up. If you're prototyping masks from minimal labels and already have the annotation infrastructure, it works. Just know the runtime rejects tiny ViT variants, so stick with base or larger.

Install to Claude Code

npx -y skills add nvidia/skills --skill tao-train-mask-auto-label --agent claude-code