Research Article
Custom Network Quantization Method for Lightweight CNN Acceleration on FPGAs
Table 4
The comparison between the FP networks and the optimized networks.
| | Network | Size (M) | Top-1 accuracy (%) | Accuracy loss (%) | Compression ratio (%) |
| | MobileNetv1-F | 12.7 | 90.72 | — | — | | MobileNetv3-F | 17.1 | 94.75 | — | — | | PPLCNet-F | 7.62 | 91.51 | — | — | | PPLCNetv2-F | 23.4 | 94.42 | — | — | | MobileNetv1-O | 3.5 | 89.77 | 0.95 | 72.44 | | MobileNetv3-O | 4.7 | 93.84 | 0.91 | 72.51 | | PPLCNet-O | 2 | 89.68 | 1.83 | 73.75 | | PPLCNetv2-O | 5.7 | 93.46 | 0.96 | 75.64 |
|
|
