Some confusion.

[wspba]

Hi, I've run your code and achieved the performance you gave in your paper, it's a very nice work which shocked me a lot. But there are still few things which make me confuse, first, the training accuracy is even lower than the validation accuracy which is quite different from our experience; second, there is only one layer basis for the weight of one layer and it remains unchanged during training, have you tried the filter-wise basis or learnable basis? Third, have you ever tried much lower bit-width such as w2a2 or w1a1?
Hope to receive your reply, thanks very much!