AI Paper Atlas

Research map / Architectures and efficiency / Quantization

RepQ-ViT: Scale Reparameterization for Post-Training Quantization of Vision Transformers

Zhikai Li, Junrui Xiao, Lianwei Yang, Qingyi Gu

ICCV 2023 · Architectures and efficiency › Quantization · approach: Post-training quantization of ViTs · 157 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
reparameterize channel-wise to layer-wise quantization for post-LayerNorm activations (vs direct layer-wise) (ablation)RepQ-ViT · ImageNet W4/A4Top-1 accuracy: 33.17 -> 69.03 (DeiT-S)Improved
reparameterize channel-wise to layer-wise for post-LayerNorm on Swin-S (ablation)RepQ-ViT · ImageNet W4/A4Top-1 accuracy: 57.63 -> 79.45Improved
use scale reparameterization instead of true channel-wise quantization (ablation)RepQ-ViT · ImageNet W4/A4Top-1 accuracy: 70.28 -> 69.03Hurt
use log sqrt2 quantizer instead of log2 for post-Softmax (ablation)RepQ-ViT · ImageNet W4/A4Top-1 accuracy: 77.87 -> 79.45 (Swin-S)Improved
use log sqrt2 quantizer instead of log2 on DeiT-S (ablation)RepQ-ViT · ImageNet W4/A4Top-1 accuracy: 67.71 -> 69.03Improved
reparameterize log sqrt2 to log2 for inference (ablation)RepQ-ViT post-Softmax · ImageNet W4/A4Top-1 accuracy: 69.03 -> 69.03No clear effect
apply RepQ-ViT instead of APQ-ViT for 4-bit quantizationViT-B · ImageNet W4/A4Top-1 accuracy: 41.41 -> 68.48Improved
apply RepQ-ViT instead of APQ-ViT in COCO W4/A4 with Swin-S Mask R-CNNMask R-CNN Swin-S · COCOAP box: 44.7 -> 44.2Hurt
apply RepQ-ViT instead of APQ-ViT in COCO W4/A4 with Swin-T Cascade Mask R-CNNCascade Mask R-CNN Swin-T · COCOAP box: 27.2 -> 47.0Improved
avoid reconstruction and hyperparameters; use RepQ-ViT instead of PTQ4ViTDeiT-S W4/A4 · ImageNetGPU minutes: 3.2 -> 1.3Improved

Later papers that compared with this method

15 later accepted papers report RepQ-ViT as a baseline in their results tables; 15 of them report a better number on most of the results they share with it.

7 more not listed.

Earlier methods this paper compares with

Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.

Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy