ANet about rtd-action HOT 10 CLOSED

And I try to retrain the model, but the results can not reach the result of your paper, where is the problem?

from rtd-action.

How do you RE-SCALE the feature to 100?

I reproduced the results on more than 2 machines and made the screenshot of the performance available.

from rtd-action.

When I used the feature, there were not so many choices.

There were only initial I3D features, and I had to re-scale them.

from rtd-action.

I use this function to re-scale to 100.

Is this correct?

from rtd-action.

Is it related to batch size? I can't set 128 batch size on one gpu, I can only set 32 or 64 batch

from rtd-action.

Maybe you can try 128 batch size with DP,
but I do not think batch size will affect the evaluation performance.

from rtd-action.

I find in your code, the temporal_scale is not used?

from rtd-action.

https://github.com/MCG-NJU/RTD-Action/blob/ANet/datasets/anet.py#L68-L71

from rtd-action.

I will upload my re-scaled features.

from rtd-action.

Thanks for your reply !

from rtd-action.