BnCfg in PrunePose_HrNet

[tuongtranngoc]

Hi, I see bncfg param in purnpose_hrnet.py but i do'nt see it. PLease explain!

[vvhj]

For APRCP HRNet you can get our prtrain model in : https://drive.google.com/file/d/1-EXl9dSatzmUSGpWGuBFlcPPM9T8Gcfr/view?usp=drivesdk

For a purned model, there are two main file:

pruneXXX.txt // to build model XXXXXXXX.pth // weight of model

pruneXXX.txt is bncfg

[tuongtranngoc]

Thank @vvhj , I test model w32_extreme and w32_best on RTX2080ti inference time of two model (~30ms) are same original model. it's no change!!

[vvhj]

CPU is work, The effect of GPU is not obvious because of parallel computing, but I am trying to speed up GPU by pruning without skipping fusion layer (another reason mybe that RTX2080ti is too powerful).

[vvhj]

Some new work will open after cvpr2021, and we will response you as soon as possible.

[tuongtranngoc]

I understand, thank you very much.

[sunmengnan]

why self.bnindex += 2 in class PosePurnHighResolutionNet？WHy isn't +1?

[vvhj]

We prune 2 bn layer in one basic modules, so self.bnindex += 2. If you only prune 1 bn layer self.bnindex += 1. You can add a breakpoint on the "modules" to see the network structure.

[sunmengnan]

Thx whj, I saw in the end, you measure the idx by adding the mean of ap+acc and the index of percent, why add the index of percent? I don't quite understand.

[sunmengnan]

BTW, Have you ever taking not only the percent, but also the absolute sum value of the all the weights in each channel? It could be more reasonable.

[ghgggg]

I modified it with fuse layer pruning。this will reduce about 1/3 flops and the infertime will be 4/5 of the original。parameters could reach 1/4 of the src model size ，however it is not realtime

[vvhj]

BTW, Have you ever taking not only the percent, but also the absolute sum value of the all the weights in each channel? It could be more reasonable.

Thx, BTW I will try it.