cudn_version_disable_cudnn_effect about temporal-segment-networks

[yjxiong]

Disabling cuDNN won't affect the accuracy. But the memory usage will grow due to the native implementation of convolution in Caffe. We have a local build of OpenCV for DenseFlow, you can also use that to build your Caffe. In this way you no longer need the permission to rebuild the system-wide OpenCV installation.

For the second problem, optical flow is extracted for each time step.

[3DMM-ICME2023]

Thanks for your prompt reply!
I find that if i build of OpenCV for DenseFlow , First i need install OpenCV dependenices such as build-essential, will it update the system-wide dependency?
For the second problem, what each time step means, extract 1 optical flow one second?

[yjxiong]

For the dependencies, there is no need to install them once they already present. Building OpenCV itself won't affect system-wide dependency.

Again, optical flow is extracted from each pair of consecutive frames.

[yjxiong]

@liu666666 , did you manage to build TSN?

[3DMM-ICME2023]

I Disabled cuDNN and make TSN successfully, but when i use the scripts(eval_net.py) to eval a video,i meet a bug which says than no module name pyActionRecog. i did not fix it now and can you tell me how?Thanks for your guidance!

[yjxiong]

Try add . to your PYTHONPATH variable.