English (United Kingdom)

https://curated-unify.zendy.io/wp-json/zendy-region/v1/featured_content/oa?rat=en

https://curated-unify.zendy.io/wp-json/zendy-region/v1/highlighted_journal/

Global: Zendy Plus annual

Presents the access of premium content as premium feature

Premium Content

Presents the keyphrase highlighting as premium feature

Keyphrase Highlighting

Presents the summarisation as premium feature

Summarisation

Insights

Presents the pdf analysis as premium feature

PDF Analysis

Presents the zaia usage as premium feature

ZAIA

Global: Zendy Plus monthly

Global: Zendy Tools annual

Global: Zendy Tools monthly

Global: Zendy Free Plan

Quantization is of significance for compressing the over-parameterized deepneural models and deploying them on resource-limited devices. Fixed-precisionquantization suffers from performance drop due to the limited numericalrepresentation ability. Conversely, mixed-precision quantization (MPQ) isadvocated to compress the model effectively by allocating heterogeneousbit-width for layers. MPQ is typically organized into a searching-retrainingtwo-stage process. Previous works only focus on determining the optimalbit-width configuration in the first stage efficiently, while ignoring theconsiderable time costs in the second stage. However, retraining alwaysconsumes hundreds of GPU-hours on the cutting-edge GPUs, thus hinderingdeployment efficiency significantly. In this paper, we devise a one-shottraining-searching paradigm for mixed-precision model compression.Specifically, in the first stage, all potential bit-width configurations arecoupled and thus optimized simultaneously within a set of shared weights.However, our observations reveal a previously unseen and severe bit-widthinterference phenomenon among highly coupled weights during optimization,leading to considerable performance degradation under a high compression ratio.To tackle this problem, we first design a bit-width scheduler to dynamicallyfreeze the most turbulent bit-width of layers during training, to ensure therest bit-widths converged properly. Then, taking inspiration from informationtheory, we present an information distortion mitigation technique to align thebehaviour of the bad-performing bit-widths to the well-performing ones.

Retraining-free Model Quantization via One-Shot Weight-Coupling Learning