Chengxu Zhuang

Thanks for noticing this. But I believe it is a problem that will not influence the results, the pytorch version automatically broadcast the dimension to make the difference not influence...

Sorry, I am not super familiar with the pytorch implementation, but I think even if for the first batch, the k-means is not done, it can just use the initial...

Hi Zhanxuan, Thanks for your question. We should have made it clearer in the codes. In this "_softmax" function, Z serves as a normalization factor for the non-parametric softmax formulation....

Hi SuiAn, That is possible, although this value does not make a difference in our algorithm. You can also check Instance Discrimination's original implementation.

I believe a recent update of faiss changed its interface. You need to install `faiss_gpu==1.6.1`, see this issue: https://github.com/neuroailab/VIE/issues/7. I will add this to the documentation, thanks!

Hi Marvin, using 2 Titan-Xp gpus, one epoch takes like 2500s to 3000s, so it's less than 1 hour. The whole model training takes usually less than 1 week.

At the begining of the training, which is before any loss computation, Ci would be computed using the saved memory bank.

Hi, please check explanations here: https://github.com/neuroailab/LocalAggregation-Pytorch/issues/1, sorry for the confusion.

The memory bank stores the embeddings of all the training images.

To be clear, the memory bank is in the shape of 1.2M * 128, where 128 is the dimension of the embedding, and the 1.2M is the size of ImageNet.