.

README.md

@@ -1,6 +1,6 @@
 # OFQ: Oscillation-free Quantization for Low-bit Vision Transformers
 
-This repository contains the training code of ViT introduced in our work: "[Oscillation-free Quantization for Low-bit Vision Transformers](https://arxiv.org/abs/2302.02210)"
+This repository contains the training code of ViT introduced in our work: "[Oscillation-free Quantization for Low-bit Vision Transformers](https://arxiv.org/abs/2302.02210)" which has been accepted for ICML 2023.
 
 In this work, we discusses the issue of weight oscillation in quantization-aware training and how it negatively affects model performance. The learnable scaling factor, commonly used in quantization, was found to worsen weight oscillation. The study proposes three techniques to address this issue: statistical weight quantization (StatsQ), confidence-guided annealing (CGA), and query-key reparameterization (QKR). These techniques were tested on the ViT model and were found to improve quantization robustness and accuracy. The proposed 2-bit DeiT-T/DeiT-S algorithms outperform the previous state-of-the-art by 9.8% and 7.7%, respectively.