About training detaills about siamdw HOT 2 CLOSED

Hi,

During training for SiamFC+, do you freeze the weights of first 7*7 conv as described in the paper? If yes, why I cannot find the corresponding operations in this code? And does the same operation applies for the SiamRPN+?

Thanks for your time.

There is no difference for results whether you freeze the last layer or not. And yes, I unfreeze the blocks gradually during training. For simplity, I provide the 'unfreezing all with smaller learning rate' version. You can try these two versions, similiar results come. RPN and FC use similiar training strategy.

from siamdw.

Hi,

During training for SiamFC+, do you freeze the weights of first 7*7 conv as described in the paper? If yes, why I cannot find the corresponding operations in this code? And does the same operation applies for the SiamRPN+?

Thanks for your time.

I provide a version with freeze-out strategy. See backbone.py.This issue will be closed.

from siamdw.