Super-Resolution Model Quantized in Multi-Precision

Abstract

Deep learning has achieved outstanding results in various tasks in machine learning under the background of rapid increase in equipment’s computing capacity. However, while achieving higher performance and effects, model size is larger, training and inference time longer, the memory and storage occupancy increasing, the computing efficiency shrinking, and the energy consumption augmenting. Consequently, it’s difficult to let these models run on edge devices such as micro and mobile devices. Model compression technology is gradually emerging and researched, for instance, model quantization. Quantization aware training can take more accuracy loss resulting from data mapping in model training into account, which clamps and approximates the data when updating parameters, and introduces quantization errors into the model loss function. In quantization, we found that some stages of the two super-resolution model networks, SRGAN and ESRGAN, showed sensitivity to quantization, which greatly reduced the performance. Therefore, we use higher-bits integer quantization for the sensitive stage, and train the model together in quantization aware training. Although model size was sacrificed a little, the accuracy approaching the original model was achieved. The ESRGAN model was still reduced by nearly 67.14% and SRGAN model was reduced by nearly 68.48%, and the inference time was reduced by nearly 30.48% and 39.85% respectively. What’s more, the PI values of SRGAN and ESRGAN are 2.1049 and 2.2075 respectively.

Skip-Concatenated Image Super-Resolution Network for Mobile Devices

“…The representative quantization methods include mixed precision [29] and quantization-aware training (QAT) [30]. Mixed precision training can improve the performance by searching the optimized bits per layer.…”