Towards Clip-Free Quantized Super-Resolution Networks: How to Tame Representative Images
This work addresses the efficiency and quality issues in mobile super-resolution networks by eliminating the need for clipped activations, which is an incremental improvement for deployment on resource-constrained devices.
The paper tackles the problem of post-training quantization for super-resolution networks by proposing a clip-free quantization pipeline that cleverly augments representative dataset images using only outputs from the FP32 model, resulting in up to 54% faster inference runtime and improved visual quality compared to INT8 clipped models.
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.