The accuracy is so high, so will this paper be the end of NTU-60? Or will there be any potential for promotion? about ms-g3d

[ZZYuting]

from ptflops import get_model_complexity_info
flops, params = get_model_complexity_info(model, (3, 10, 25, 2), as_strings=True, print_per_layer_stat=True)
print("%s |%s |%s" % ('MSG3D', flops, params))

[ZZYuting]

I think the directions are how to light the model. It seems cost much time.

[desertfex]

@ZZYuting The number of layers has dropped to 3 comparing with the ST-GCN with 10 layers. I have no idea how to light the model without loss of accuracy.

[ZZYuting]

@ZZYuting The number of layers has dropped to 3 comparing with the ST-GCN with 10 layers. I have no idea how to light the model without loss of accuracy.
Although it has dropped to 3 compared with the st-gcn, the flops i caculate the model still up to 20G+flops, you can read some paper in cvpr 2020

[kenziyuliu]

My personal opinions:

• It seems that the performance on NTU RGB+D 60 dataset is getting saturated (e.g. suggested by this preprint, section 4.1) even with model ensembles, which are used by most work on this topic now; I would suggest working on NTU RGB+D 120 and Kinetics Skeleton 400 for future work.
• One future direction is to look at how to make MS-G3D more efficient (as @ZZYuting has pointed out). Some components that slow down training include the spatial-temporal window graph/feature construction and the use of graph residual masks. Please also check our supplementary material where we discuss some future directions.

[desertfex]

@kenziyuliu Thank you very much indeed.

[kenziyuliu]

I'll close this issue now but please feel free to add more comments to this thread.

[Lyndsey-Xing]

@ZZYuting The number of layers has dropped to 3 comparing with the ST-GCN with 10 layers. I have no idea how to light the model without loss of accuracy.
Although it has dropped to 3 compared with the st-gcn, the flops i caculate the model still up to 20G+flops, you can read some paper in cvpr 2020

Hello, how do you calculate flops in this paper? Is there any way to reduce flops？
Thanks.

[ZZYuting]

@ZZYuting The number of layers has dropped to 3 comparing with the ST-GCN with 10 layers. I have no idea how to light the model without loss of accuracy.
Although it has dropped to 3 compared with the st-gcn, the flops i caculate the model still up to 20G+flops, you can read some paper in cvpr 2020

Hello, how do you calculate flops in this paper? Is there any way to reduce flops？
Thanks.
I use the ptflops to calculate flops coarsely. In fact, you can calculate flops on your hand which is more precise. I think it's a good way to learn how to reduce flops if you know which parts is complex.

[desertfex]

@ZZYuting Could you please provide the codes calculating the flops? I know its may be only two or three codes, but I still get "error" from the compiler. Thank you.

[desertfex]

@ZZYuting Thank you very much.