Research Article
Custom Network Quantization Method for Lightweight CNN Acceleration on FPGAs
Table 5
The comparison between the conventional quantized networks and the optimized networks.
| Network | Top-1 accuracy (%) | XCZU3EG | XCVU13P | Size (M) | Compression ratio (%) | Inference time (ms) | Time reduction (%) | Inference time (ms) | Time reduction (%) |
| MobileNetv1-C | 89.98 | 29.44 | — | 8.04 | — | 3.7 | 70.86 | MobileNetv3-C | 93.90 | 53.97 | — | 14.92 | — | 5.1 | 70.17 | PPLCNet-C | 89.56 | 16.35 | — | 4.30 | — | 2.2 | 71.12 | PPLCNetv2-C | 93.61 | 47.81 | — | 13.79 | — | 6.0 | 74.35 | MobileNetv1-O | 89.77 | 16.44 | 44.16 | 4.32 | 46.27 | 3.5 | 72.44 | MobileNetv3-O | 93.84 | 18.89 | 64.99 | 5.01 | 66.42 | 4.7 | 72.51 | PPLCNet-O | 89.68 | 9.88 | 39.57 | 2.69 | 37.44 | 2.0 | 73.75 | PPLCNetv2-O | 93.46 | 23.01 | 51.87 | 6.10 | 55.76 | 5.7 | 75.64 |
|
|