Reproduce “Distribution Adaptive INT8 Quantization for Training CNNs” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Kang Zhao, Sida Huang, Pan Pan, Yinghan Li, Yingya Zhang, Zhenyu Gu, Yinghui Xu
The authors develop an eight-bit training method for convolutional neural networks that assigns gradient quantization ranges by channel, using gradient patterns and magnitude to guide range selection. On ImageNet, ResNet-50 reaches 76.59% top-1 accuracy, versus its listed FP32 baseline of 76.50%. On an RTX 2080 Ti, its iteration takes 115.6 ms; results are specific to the tested setup.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.