Figure 12.

C3-light network structure. For an input I $\in R^{c\times h\times w}$, PConv applies $c_p$ filter W $\in R^{k\times k}$. This makes PConv have as low Flops as $h\times w\times k^2\times c_p^2$ compared to a regular Conv with $h\times w\times k^2\times c^2$. With a typical partial ratio $\frac{c_p}{c}=\frac{1}{4}$, the Flops of PConv are only $\frac{1}{16}$ of a regular Conv.