philiptheother / featuredecoupling Goto Github PK

Hi, please tell me where is the prob.dat file?

Thanks.

Hi, congratulations on your work!
I was actually wondering if there is any way to run this code on a custom dataset?

Hi,
Thanks for your work and for sharing the code, the results are very good on ImageNet and other big image datasets.
I was wondering if you ran experiments on smaller images such as the ones in the CIFARs. I tried runing them myself but I am not so sure about how to configure the NCE layer and the results I get are really poor.

Do you have any observations to share over this ?

Thanks

Hi there,

Looking through the code it seems like transforms.RandomResizedCrop is used when training with self-supervision but then it is disabled when training the linear classifier. I was just wondering if there was a reason for the change, and if it makes a big difference in the final performance?

Thanks!

Thanks for the code implementation, but I don't know how to implement the code of "generate the prob.dat file", could you share it? thanks.

Hi, what version of python is used for this project. I am having some issues loading modules like imp. Thanks

Hi,

Many thanks for sharing your work! I am playing with your method in CIFAR-10 and I was wondering if you could help me a bit on the behaviour that I should expect from the losses. Did you observe worse rotation accuracy/loss than when training the rotation method alone? Could this be normal considering that rotation accuracy is not necessarily a good measure and learning together with NPID is just helping on forming more discriminative features that don't help rotation accuracy?

Many thanks in advance!

Best,
Diego.