AI Paper Atlas

Research map / Architectures and efficiency / Quantization

QPP: Real-Time Quantization Parameter Prediction for Deep Neural Networks

Vladimir Kryzhanovskiy, Gleb Balitskiy, Nikolay Kozyrskiy, Aleksandr Zuruev

CVPR 2021 · Architectures and efficiency › Quantization · approach: Post-training weight quantization · 16 citations (OpenAlex)

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
Predict per-sample QPs with QPP instead of using averaged static QPsD+S quantization · ImageNet, CityScapes, COFW, WFLW, 300W, Vimeo, Set5, Set14task performance (Top-1 accuracy, mIoU, NME, or PSNR): 69.08 -> 69.14; 72.60 -> 72.67; 69.99 -> 70.0; 3.52 -> 3.55; 4.68 -> 4.69; 3.92 -> 3.93; 34.0 -> 34.01; 30.70 -> 30.69; Mixed
Predict QPs with QPP rather than averaged static QPsD+S quantization · All evaluated task datasetsFR(1): 3.4 / 16.1 -> 0.78 / 7.7; 4.5 / 23.9 -> 0.5 / 7.1; 0.7 / 13.9 -> 0.06 / 2.2; 3.5 / 64.2 -> 0.2 / 56.7; 8.5 / 93.4 -> 1.4Improved
Use D+S quantization instead of standard dynamic quantizationPost-training 4-bit activation quantization · ImageNet (ResNet34)Top-1 accuracy (%): 70.34 -> 72.72Improved
Replace dynamic quantile calculation with QPPD+S quantization inference · Kirin 980 convolution benchmarkinference time: 23% overhead; decreased by 11%Improved
Use D+S QPP rather than full-precision inferenceConvolution inference on Kirin 980 · Kirin 980 convolution benchmarkacceleration: x1.46Improved

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