xinghaochen / slab Goto Github PK

Thank you for the very nice work.

I’m currently exploring various normalization techniques and noticed that RMSNorm is generally considered a less computationally intensive alternative compared to Layer Norm. Given that I’m not deeply familiar with the intricacies of these normalization methods, I would appreciate some clarification on how your method compares in terms of efficiency with RMSNorm.

Could you provide some insights or benchmarks on this? Any information you can share would be greatly helpful.

Thank you very much for your work, this work should be very scalable!
I have observed that the superparameter step (total step) of linearnorm in the code seems to be only related to the model, but not to the dataset, whether it is because these are fixed by your implementation.
linearnorm = partial(LinearNorm, norm1=ln, norm2=RepBN, step=60000) in slab_swin
linearnorm = partial(LinearNorm, norm1=nn.LayerNorm, norm2=RepBN, step=120000) in slab_cswin
linearnorm = partial(LinearNorm, norm1=ln, norm2=RepBN) in others
Should I set my total step to epoch * len (train_dataloader) in linearnorm when I train it? or any tricks here?

Could you please give me the link of the data set and the processing method of the data, and tell me which directory the data set should be placed in ?