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 Tools annual

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 Tools monthly

Global: Zendy Plus monthly

Presents the access of premium content as premium feature

Premium Content

Global: Zendy Plus annual

Global: Zendy Free Plan

Quantized networks use less computational and memory resources and aresuitable for deployment on edge devices. While quantization-aware training QATis the well-studied approach to quantize the networks at low precision, mostresearch focuses on over-parameterized networks for classification with limitedstudies on popular and edge device friendly single-shot object detection andsemantic segmentation methods like YOLO. Moreover, majority of QAT methods relyon Straight-through Estimator (STE) approximation which suffers from anoscillation phenomenon resulting in sub-optimal network quantization. In thispaper, we show that it is difficult to achieve extremely low precision (4-bitand lower) for efficient YOLO models even with SOTA QAT methods due tooscillation issue and existing methods to overcome this problem are noteffective on these models. To mitigate the effect of oscillation, we firstpropose Exponentially Moving Average (EMA) based update to the QAT model.Further, we propose a simple QAT correction method, namely QC, that takes onlya single epoch of training after standard QAT procedure to correct the errorinduced by oscillating weights and activations resulting in a more accuratequantized model. With extensive evaluation on COCO dataset using various YOLO5and YOLO7 variants, we show that our correction method improves quantized YOLOnetworks consistently on both object detection and segmentation tasks atlow-precision (4-bit and 3-bit).

Reducing the Side-Effects of Oscillations in Training of Quantized YOLO Networks