All the batches are not cyclic consistent when using my own dataset about squeeze-and-recursion-temporal-gates

[alexandrosstergiou]

Hi @zzwei1,
I think in cases that the data are not cyclic consistent it would make more sense to simply use the Squeeze and Recursion blocks without the cyclic-consistency gating mechanism (i.e. setting the gates parameter to False). This way you can regularise/calibrate the activation maps based on information from the entire video.

I do not know if it would be sensible for your data, but If you still want to use cyclic-consistency, you could make something like a hybrid training scheme where the first x epochs you do not use the gates and then for the last y epochs you enable the gates.

Hope that any of the above ideas are proven useful, and thanks for using the code.

Best,
Alex

[alexandrosstergiou]

Will close this issue for now due to inactivity