Table 3 Pruning results of VGG-16 on CIFAR-100. The best pruning results are bolded, and the worst results are underlined.
From: A multi-agent reinforcement learning based approach for automatic filter pruning
\(\textrm{Imp}(\cdot )\) | pruning rate: 50% | |||||
|---|---|---|---|---|---|---|
Acc.(pruned) | Acc.(fine-tuned) | Filters | Params(M) | FLOPs(M) | ||
|  | Base | 0.8961 | – | 4224 | 14.941732 | 314.524672 |
\(\varvec{w}_{i,j}\) | Weight | 0.8672 | 0.8856[98.82%] | 2020 | 3.043907 | 135.09748 |
Taylor | 0.8281 | 0.8755[97.70%] | 2026 | 3.073754 | 100.467832 | |
\(\varvec{F}_{i,j}\) | IB | 0.7969 | 0.8737[97.50%] | 2097 | 3.313371 | 137.484096 |
BN | 0.9062 | 0.8876[99.05%] | 2099 | 3.008191 | 167.201600 | |
Gradient | 0.8047 | 0.8809[98.30%] | 2095 | 3.264122 | 136.683472 | |
\(\varvec{A}_{i,j}\) | APoP | 0.7969 | 0.8786[98.04%] | 2025 | 3.187321 | 160.232956 |
\(\textrm{Imp}(\cdot )\) | pruning rate: 60% | |||||
|---|---|---|---|---|---|---|
Acc.(pruned) | Acc.(fine-tuned) | Filters | Params(M) | FLOPs(M) | ||
|  | Base | 0.8961 | – | 4224 | 14.941732 | 314.524672 |
\(\varvec{w}_{i,j}\) | Weight | 0.7969 | 0.8773[97.90%] | 1611 | 2.117810 | 72.620072 |
Taylor | 0.8594 | 0.8790[98.09%] | 1675 | 2.131325 | 121.722136 | |
\(\varvec{F}_{i,j}\) | IB | 0.6250 | 0.8665[96.69%] | 1673 | 2.582921 | 98.543556 |
BN | 0.8203 | 0.8759[97.74%] | 1619 | 1.911082 | 71.725748 | |
Gradient | 0.7578 | 0.8718[97.28%] | 1660 | 2.069356 | 84.059268 | |
\(\varvec{A}_{i,j}\) | APoP | 0.6016 | 0.8681[98.11%] | 1647 | 2.051498 | 96.258680 |
\(\textrm{Imp}(\cdot )\) | pruning rate: 70% | |||||
|---|---|---|---|---|---|---|
Acc.(pruned) | Acc.(fine-tuned) | Filters | Params(M) | FLOPs(M) | ||
|  | Base | 0.8961 | – | 4224 | 14.941732 | 314.524672 |
\(\varvec{w}_{i,j}\) | Weight | 0.6406 | 0.8657[96.60%] | 1220 | 0.889273 | 51.890312 |
Taylor | 0.7266 | 0.8727[97.39%] | 1193 | 1.075607 | 72.818532 | |
\(\varvec{F}_{i,j}\) | IB | 0.4688 | 0.8626[96.26%] | 1259 | 1.052788 | 88.585496 |
BN | 0.7656 | 0.8707[97.16%] | 1238 | 1.286112 | 50.040904 | |
Gradient | 0.7109 | 0.8730[97.42%] | 1220 | 1.267270 | 92.931200 | |
\(\varvec{A}_{i,j}\) | APoP | 0.3828 | 0.8559[95.51%] | 1246 | 1.282633 | 45.668076 |
