Channel number to be pruned must be a pow of 2 ?

[summer-wind-hi]

Hi @liuzhuang13,
There is another issue I met in my experiments: inference time reduced when channel number to be pruned was a pow of 2 , otherwise it increased and longer than baseline, which is not mentioned. Have you ever met same issue in your experiments? and any suggestions?

my hardware and system:
GPU: GeForce RTX 2080 12G
CPU: 58G, 12 cores
System: Ubuntu16.04

regards,
summer, Gao

[HHCorp]

Hi @liuzhuang13, There is another issue I met in my experiments: inference time reduced when channel number to be pruned was a pow of 2 , otherwise it increased and longer than baseline, which is not mentioned. Have you ever met same issue in your experiments? and any suggestions?

my hardware and system: GPU: GeForce RTX 2080 12G CPU: 58G, 12 cores System: Ubuntu16.04

regards, summer, Gao

@summer-wind-hi
It involves memory alignment. The inference time also depends on your processor(CPU, GPU or other chips) and for most processors, the data size that divisible by 8 is aligned to its memory access size, which can make memory access and matrix computation faster. Hope this 'https://stackoverflow.com/questions/381244/purpose-of-memory-alignment' can make you more clear. You can also find explanation in some papers (e.g.https://openaccess.thecvf.com/content_ICCV_2019/papers/Yu_Universally_Slimmable_Networks_and_Improved_Training_Techniques_ICCV_2019_paper.pdf, It mentions the width design divisible by 8,16,32... that matchs the warp of GPU, has better performance.