A PConv followed by a PWConv well approximates the regular Conv 3 × 3,is it similar in speed

[wsy-yjys]

I read your paper. It was great! And I have a question, the paper says “A PConv followed by a PWConv well approximates the regular Conv 3 × 3” , I'm very curious are they similar in speed, thank you~

[JierunChen]

Thanks for your interest in this work. A PConv followed by a PWConv well approximates and is $2\sim7\times$ faster than a regular Conv 3 × 3. The speedup varies for different feature map sizes and devices, as shown in the following table:

Operator $\times10$ layers	Feature map size	FLOPs (M)	Throughput ($\times1000$ fps) on GPU	Latency (ms) on CPU	Latency (ms) on ARM
Conv $3\times3$	$96\times56\times56$ $192\times28\times28$ $384\times14\times14$ $768\times7\times7$	2601	3.01 4.89 4.56 3.16	35.67 28.41 31.85 62.71	780 620 595 662
PConv 3$\times$3 with $r=\frac{1}{4}$ followed by a PWConv	$96\times56\times56$ $192\times28\times28$ $384\times14\times14$ $768\times7\times7$	451	6.97 9.58 11.43 10.46	7.88 5.42 6.86 8.34	167 133 116 121
PConv 3$\times$3 with $r=\frac{1}{4}$	$96\times56\times56$ $192\times28\times28$ $384\times14\times14$ $768\times7\times7$	162	15.39 20.51 37.21 43.24	4.74 2.38 3.32 3.95	81.70 62.69 48.19 47.52
PWConv (Conv $1\times1$)	$96\times56\times56$ $192\times28\times28$ $384\times14\times14$ $768\times7\times7$	289	12.75 17.98 16.49 13.79	3.14 3.04 3.54 4.39	85.07 70.50 67.38 73.58

[wsy-yjys]

You are very careful. Thank you very much