cupidjay / training-vision-transformers-with-only-2040-images Goto Github PK

Hi, excuse me for a question. The offical flowers dataset have a mat file for data splitting and I found the train and val set size are both 1020, while the test set size is 6149. Do we use this split directly by combing the train and val set to size of 2040 to train and leaving the test set unchanged ? Thank you .

Hi, excuse me for a new question. I try to repeat the DINO acc@1 in table 2, but the results are very different from the paper. I use the code from facebookresearch/dino github repo. Could you share some special training setting ? I have noticed the training details in the end of the paper. Are the /tao parameter means the teacher temp ? And are there any other setting different from the offical repo ? Looking forward to your reply, thank you very much !

Hi,

Many thanks for your excellent work. I am truly interested in your work.

Did you try the cifar10 or cifar100?

I also wonder if it is possible to do some experiments with ImageNet, but it seems that how to select/split the data is a big issue.

cheers.

Hi,

Thank you for your exellent work!

I noticed in your paper that if the W of fc layer is initialized by activations in forward pass, the result will be more regularised and converged. But I didn't see the corresponding code. How do you think about that?

Did you only code the instance discrimination and ema of the parameters?

_, term_width = os.popen('stty size', 'r').read().split()
ValueError: not enough values to unpack (expected 2, got 0)
Why is this error reported