PerfHD: Efficient ViT Architecture Performance Ranking using Hyperdimensional Computing
- This repository provides a straightforward implementation of PerfHD, which leverage the capability of Hyperdimensional Computing (HDC) to tackle architecture performance ranking problem in ViT neural architecture search.
- Straightforward. PerfHD has straightforward implementation and the entire work can be summarized with one Jupyter notebook with less than 100 lines of Python code.
- Efficient, yet accurate. Using just one GPU, PerfHD can rank nearly 100K ViT models in about just 1 minute. No pretraining or fine-tuning required. PerfHD is up to 100 times faster however still has competitive performance compared with SOTA algorithms.
- The following python packages are required to run the notebook:
json
torch
numpy
sklearn
scipy
tqdm
- Python >= 3.8
- GPU acceleration is strongly recommended however not strictly required. PerfHD is still fast using CPU only.
- The train and test data can be found here (CVPR 2022 NAS competition)
- Dongning Ma (Student), Xun Jiao (Professor), Villanova University
- Pengfei Zhao, Beijing Xiaochuan Technology Co., Ltd.