This is an approximate but simple pytorch implementation of HetConv: Heterogeneous Kernel-Based Convolutions for Deep CNNs. The main purpose is to reduce the number of FLOPs and parameters without sacrificing speed.
| Model | FLOPs | Params | Acc% |
|---|---|---|---|
| vgg16_bn_P1 | 313.47M | 14.99M | 93.88 |
| vgg16_bn_P2 | 192.36M | 9.27M | 94.09 |
| vgg16_bn_P4 | 114.50M | 5.59M | 93.74 |
| vgg16_bn_P8 | 75.57M | 3.75M | 93.62 |
| vgg16_bn_P16 | 56.11M | 2.83M | 93.53 |
| vgg16_bn_P32 | 46.38M | 2.37M | 93.06 |
| vgg16_bn_P64 | 41.51M | 2.14M | 92.69 |