Learning Robust Deep Face Representation about face_verification_experiment HOT 6 CLOSED

For example, we assume the output of conv1 is 12012096. We can use slice layer to divide the conv1 into conv1_1(12012048) and conv1_2(12012048) and then we use eltwise layer to get the output(12012048) by computing the max value between two groups of feature maps.

from face_verification_experiment.

Ok, I understand it now. So if I would like to train the model with ReLU layers instead of MFM, I would replace the layers convolution-pooling-slice-eltwise with convolution-pooling-ReLU, am I right?

from face_verification_experiment.

The order should be convolution-ReLU-pooling in theory, but I think it doesn't influence the results whichever orders you choose.

from face_verification_experiment.

Ok! Yes, it should be the same.

Regarding to the same paper I have another question. "The learning rate is set to 1e-3 initially and reduce to 5e-5 gradually". Did you use any specific learning rate policy (step down, exponential decay...)? Or was it a fixed learning rate (1e-3), manually decreased to 5e-5? Thanks!

from face_verification_experiment.

The policy of learning rate is fixed and I manually changed it if the loss isn't decreased for a long time.

from face_verification_experiment.

Ok, understood! Thank you very much

from face_verification_experiment.