Questions about communication mechanism

[drowning-in-codes]

Thank you for your great work. I have some questions about the communication computation and the effect of code size.

1. As the paper said, the bandwidth computation involves n_L which is the size of the codebook(I guess). But on the spatial level, the data is H×W× |M|, is n_L needed to compute the transfer data? or does the bandwidth involve the size of the codebook?

the bandwidth given the selection matrix M is calculated as log2(H × W × |M| × log2(nL) × nR/8). Here, log2(nL) signifies the data amount required to represent each code index integer, decided by the codebook size.
If the bandwidth just computes the transmitted code indices, I think log2( H × W × | M | × nR / 8 ) is fair.

2. According to the paper it seems to work best with a codebook size of 64, according to my reproduction the results are not as good as in the paper, does this involve fine-tuning or other training techniques?
   Besides, how does the w/o codebook achieve the transfer? Is it by passing all the data?
   
   Thank you for your work again.

[JuntongPeng]

Thanks for your question.

1. Representing an indice is not simply 1 Byte. It depends on the size of codebook (maximum indice needs to be represented). That's why the nL is wrapped in log.
2. Can you please briefly describe your result? That might help localizing the problem.
3. Yes, it means passing the raw feature.