Towards Clip-Free Quantized Super-Resolution Networks: How to Tame Representative Images

TL;DR

This paper introduces a novel post-training quantization pipeline for super-resolution networks that enhances stability and efficiency by eliminating clipped activations, leading to faster inference and improved visual quality without retraining.

Contribution

The proposed clip-free quantization pipeline (CFQP) effectively removes clipped activations in quantized SR models using only outputs of the FP32 model, improving performance without retraining.

Findings

Up to 54% reduction in inference runtime.

Enhanced visual quality over INT8 clipped models.

Outperforms some FP32 models in runtime and quality.

Abstract

Super-resolution (SR) networks have been investigated for a while, with their mobile and lightweight versions gaining noticeable popularity recently. Quantization, the procedure of decreasing the precision of network parameters (mostly FP32 to INT8), is also utilized in SR networks for establishing mobile compatibility. This study focuses on a very important but mostly overlooked post-training quantization (PTQ) step: representative dataset (RD), which adjusts the quantization range for PTQ. We propose a novel pipeline (clip-free quantization pipeline, CFQP) backed up with extensive experimental justifications to cleverly augment RD images by only using outputs of the FP32 model. Using the proposed pipeline for RD, we can successfully eliminate unwanted clipped activation layers, which nearly all mobile SR methods utilize to make the model more robust to PTQ in return for a large overhead in runtime. Removing clipped activations with our method significantly benefits overall increased stability, decreased inference runtime up to 54% on some SR models, better visual quality results compared to INT8 clipped models - and outperforms even some FP32 non-quantized models, both in runtime and visual quality, without the need for retraining with clipped activation.