lyakaap / landmark2019-1st-and-3rd-place-solution Goto Github PK

Hi , as the issue says, I just wonder the meaning of parameter 'dataset' in function 'explore_exploit'. Could you please explain it? Thank you in advance!

Hi @lyakaap ,

The README mentions ResNet-101-ArcFace results by training on GLD-v1, or GLD-v1 combined with GLD-v2. Would you have results of training on GLD-v2 only? (clean version mainly)

Thanks!

There is no train.pickle file. I tried to use src/prepare_dataset.py for generating respective file but it's taking a lot of time.
Can anyone suggest to me how to generate it?

Hello, thanks for your share. When I use the P2SGradLoss, I got the message "TypeError: P2SGradBackward.forward: expected Variable (got NoneType) for return value 0", do you know what's wrong with me?

Hi lyakaap, thank you for your hard work! In your code v2clean.py, LandmarkNet is initialized with "ep4_augmentation-soft_epochs-5_loss-arcface.pth", and then trained on GLD_v2_clean dataset. To reproduce your result, "ep4_augmentation-soft_epochs-5_loss-arcface.pth" is needed, while it is not provided. Would you please upload this file?
Thanks!

thanks

It would be of great help for many of us

Hi @lyakaap ,

In your paper, you mention that

"p of GeM is set to 3.0 and fixed during the training"

I am wondering if you tried learning the p parameter, and if that produced better/worse results, and what this parameter value converged to.

Thanks!

could you release the pretrained model on the whole clean dataset? thanks

I'm trying to set up to train model. And if you can, I really wanna know how long did it take and can I get your pretrained model to test?

would you give some suggestions deal with the problem of affine transformation ? use RANSAC ?

Thanks