VisionLLaMA-Base-MAE

With the Masked Autoencoders' paradigm, VisionLLaMA-Base-MAE model is trained on ImageNet-1k without labels. It manifests substantial improvements over classification tasks (SFT, linear probing) on ImageNet-1K and the segmentation task on ADE20K.

Model	ImageNet Acc (SFT)	ImageNet Acc (Linear Probe)	ADE20K Segmentation
VisionLLaMA-Base-MAE (ep800)	84.0	69.7	49.0
VisionLLaMA-Base-MAE (ep1600)	84.3	71.7	50.2

How to Use

Please refer the Github page for usage.

Citation

@article{chu2024visionllama,
  title={VisionLLaMA: A Unified LLaMA Interface for Vision Tasks},
  author={Chu, Xiangxiang and Su, Jianlin and Zhang, Bo and Shen, Chunhua},
  journal={arXiv preprint arXiv:2403.00522},
  year={2024}
}