Super-Resolution Model Quantized in Multi-Precision
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 occu...
| Main Authors: | , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
MDPI AG
2021-09-01
|
| Series: | Electronics |
| Subjects: | |
| Online Access: | https://www.mdpi.com/2079-9292/10/17/2176 |
| _version_ | 1797521532305539072 |
|---|---|
| author | Jingyu Liu Qiong Wang Dunbo Zhang Li Shen |
| author_facet | Jingyu Liu Qiong Wang Dunbo Zhang Li Shen |
| author_sort | Jingyu Liu |
| collection | DOAJ |
| description | 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. |
| first_indexed | 2024-03-10T08:13:49Z |
| format | Article |
| id | doaj.art-9ad1dc4157e6436b997639a65a764d60 |
| institution | Directory Open Access Journal |
| issn | 2079-9292 |
| language | English |
| last_indexed | 2024-03-10T08:13:49Z |
| publishDate | 2021-09-01 |
| publisher | MDPI AG |
| record_format | Article |
| series | Electronics |
| spelling | doaj.art-9ad1dc4157e6436b997639a65a764d602023-11-22T10:31:04ZengMDPI AGElectronics2079-92922021-09-011017217610.3390/electronics10172176Super-Resolution Model Quantized in Multi-PrecisionJingyu Liu0Qiong Wang1Dunbo Zhang2Li Shen3School of Computer, National University of Defense Technology, Changsha 410073, ChinaSchool of Computer, National University of Defense Technology, Changsha 410073, ChinaSchool of Computer, National University of Defense Technology, Changsha 410073, ChinaSchool of Computer, National University of Defense Technology, Changsha 410073, ChinaDeep 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.https://www.mdpi.com/2079-9292/10/17/2176model quantizationsuper-resolutionquantization aware trainingquantization sensitivitiy |
| spellingShingle | Jingyu Liu Qiong Wang Dunbo Zhang Li Shen Super-Resolution Model Quantized in Multi-Precision Electronics model quantization super-resolution quantization aware training quantization sensitivitiy |
| title | Super-Resolution Model Quantized in Multi-Precision |
| title_full | Super-Resolution Model Quantized in Multi-Precision |
| title_fullStr | Super-Resolution Model Quantized in Multi-Precision |
| title_full_unstemmed | Super-Resolution Model Quantized in Multi-Precision |
| title_short | Super-Resolution Model Quantized in Multi-Precision |
| title_sort | super resolution model quantized in multi precision |
| topic | model quantization super-resolution quantization aware training quantization sensitivitiy |
| url | https://www.mdpi.com/2079-9292/10/17/2176 |
| work_keys_str_mv | AT jingyuliu superresolutionmodelquantizedinmultiprecision AT qiongwang superresolutionmodelquantizedinmultiprecision AT dunbozhang superresolutionmodelquantizedinmultiprecision AT lishen superresolutionmodelquantizedinmultiprecision |