some test about the inference time

[ccsvd]

hi, @ycszen , i test the native-hrnet and lite-hrnet mnn model on my pc, although theirs flops is 309M vs 203M,but their inference time is almost the same. i thank is the reason lite-hrnet has more memory read or write than native-hrnet when inference. is that right? here is some info: parameters infer memory(MB) MAdd Flops MemRead(B) MemWrite(B) duration native-hrnet: lite-hrnet ：

[YaqiLYU]

@ccsvd Hi,

1. From the info you provided, the Flops, MemRead(B) and MemWrite(B) of native-hrnet are all larger than lite-hrnet, but you said "lite-hrnet has more memory read or write than native-hrnet when inference". Is there some mistake?
2. the inference time of lite-hrnet is almost 2.5 times than native-hrnet, but you said "their inference time is almost the same". Is there some mistake?

[clover978]

I did similar experments. The conclusion is consistent.

I use onnxruntime to profile exported two onnx files on rtx3090. Inference code is exactly the same except for the onnx files to be tested.

for hrnet_w32_384x288, avg infer time is 11.7ms for lite_hrnet_30_384x288, avg infer time is 35.3ms

[tucachmo2202]

Hi @clover978, @ccsvd Could you give me file to inference an input image. I am not familiar with mmpose so I don't know how to use. Thanks!

[fire717]

so why lite-hrnet cost more time than hrnet? then why we need lite-hrnet? @clover978

[clover978]

so why lite-hrnet cost more time than hrnet? then why we need lite-hrnet? @clover978

Maybe lite-hrnet result in better efficiency on specific hardware (e.g. CPU, FPGA)，since it does have smaller FLOPS in theory.