Runtime
Image size: (512, 512, 3)
Pytorch==1.12.1, Torchvision==0.13.1
CPU
CPU config: Intel Xeon(R) CPU E5-2640 v3 @ 2.60GHz × 16
| Workflow | GPEN (original repo) | Re-arranged version |
|---|---|---|
| Real-ESRGAN | 1.1215 | 1.0924 |
| Pre-processing | 0.2390 | 0.0388 |
| Restoration | 1.9060 | 1.3512 |
| Post-processing | 1.6701 | 0.0254 |
| Total (without Real-ESRGAN) | 3.8151 | 1.4153 |
| Total | 4.9369 | 2.5078 |
CPU config: 2.8 GHz Quad-Core Intel Core i7
| Workflow | GPEN (original repo) | Re-arranged version |
|---|---|---|
| Real-ESRGAN | 4.2515 | 4.0423 |
| Pre-processing | 0.8013 | 0.0645 |
| Restoration | 6.8322 | 5.3539 |
| Post-processing | 6.3969 | 0.0186 |
| Total (without Real-ESRGAN) | 14.0304 | 5.4370 |
| Total | 18.2923 | 9.4841 |
GPU
GPU config: NVIDIA TITAN V/PCIe/SSE2
| Workflow | GPEN (original repo) | Re-arranged version |
|---|---|---|
| Real-ESRGAN | 0.0421 | 0.0430 |
| Pre-processing | 0.0321 | 0.0281 |
| Restoration | 0.0589 | 0.0558 |
| Post-processing | 0.1389 | 0.0270 |
| Total (without Real-ESRGAN) | 0.2299 | 0.1109 |
| Total | 0.2722 | 0.1539 |
CPU Memory usage
Memory usage was tested by iterating through all 3000 samples of the CelebA-LQ dataset. All images are (512,512,3). PID was used to measure the memory usage of the python script.
CPU config: Intel Xeon(R) CPU E5-2640 v3 @ 2.60GHz × 16
| Model | CPU Memory usage | Visualisation |
|---|---|---|
| GPEN (with Real-ESRGAN) | starts at 5.43GB, slowly increases, can go upto 5.70GB | link |
| GPEN (without Real-ESRGAN) | starts at 5.43GB, slowly increases, can go upto 5.75GB | link |
| Re-arranged (with Real-ESRGAN) | ~5.32GB, no sign of memory leakage | link |
| Re-arranged (without Real-ESRGAN) | ~5.13GB, no sign of memory leakage | link |
Performance
| Input | GPEN (w/o Real-ESRGAN) | GPEN (w/ Real-ESRGAN) | Re-arranged version |
|---|---|---|---|