Hanxun Huang

Try`loss.backward()` Before using`param.grad.data`

Hi, Thanks for your interest in our work. I did not try this on the segmentation dataset, but I'm very interested in the results. I don't think the current implementation...

Hi, The dataset is randomly splited into 80% for training and 20% for testing, according to each identity.

Have a look at this thread. https://github.com/happynear/AMSoftmax/issues/18

Yes, we use 112*112 for both training and testing.

Apologies, I lost track of where I downloaded the dataset. Here is another pre-processed version with 112*112. I believe I downloaded the preprocessed version. https://github.com/yule-li/CosFace/blob/master/README.md

Yes, the train_clean.sh is the baseline that is trained with the original clean dataset.

This part is for evaluating face verifications on LFW, depending if you need the results or not.

Yes, the noise should be added first. This could improve the protection performance. Keep me updated if you notice any difference in the results.

Good catch, thanks!