search

yangyu12 / squeeze-and-span

7 stars 0 forks source link

Pre-trained GAN checkpoints #1

Open cyq373 opened 1 year ago

cyq373 commented 1 year ago

Nice job! Could you upload pre-trained GAN checkpoints used in the paper, including the checkpoint trained on ImageNet?

Btw, regarding experiments on ImageNet, the top 1 acc of VICReg is 73.2, why it drops to 47.32 when using synthetic data augmentation? On other datasets like CIFAR-10/-100, the performance is improved when using synthetic data.

yangyu12 commented 1 year ago

Hi @cyq373 , Thanks for your interest. We plan to update this repository including the scripts of ImageNet experiments and the pre-trained checkpoints that you required. We will let you know when the update is finished. It is noteworthy that our experiments on ImageNet are at reduced scales, e.g. the input resolution and training duration are 96x96 and 100 epochs instead of 224x224 and 800 or 1000 epochs. That is the main reason why we report lower performance. Other reasons may also be that the GAN checkpoints we use are still of the low capability of generating ImageNet. You could find some discussion about this issue in our paper.

cyq373 commented 1 year ago

Thanks for your kind reply! I have no issues now, and I'm waiting for the project update.