a error when train with TSN3D about tpn HOT 5 CLOSED

Hi @Tylerjoe
Could you successfully run with config_files/kinetics400/tpn/r50f8s8.py?

from tpn.

Hi @Tylerjoe
Could you successfully run with config_files/kinetics400/tpn/r50f8s8.py?

Thanks for your quick reply, I have run with config_files/kinetics400/tpn/r50f8s8.py successfully.
is there any difference between r50f8s8.py and r101f32s2.py except backbone？

from tpn.

Hi @Tylerjoe ,
Thanks to your issue helping figure out the errors in config files of r101-tpn. Yeah, there exists a difference between the r50 and r101 . The scale of temporal downsample is slightly different, which leads to the shape inconsistency of various features. The performance is almost the same if r101 also takes the same scale for temporal downsample. In order to match the released models, we have to maintain the original one.

I updated the lastest configs for r101. You can try it again. Please let me know if you have any other question, and thank you again!

from tpn.

I would close this issue, you can reopen it if there remains any question.

from tpn.

Thanks for dealing with the problem so quickly!

after update the configure file, it can run successfully

from tpn.