Open liuquangao opened 2 weeks ago
The only difference that I could find was that simba will group the sequence into chunks and then only run FFT on those chunks. For example, if I have a sequence of length 6, then I would chunk it into two sequences of length 3, creating a 2x3 grid, and then run two orthogonal FFTs. This would be a 2d FFT, which would make sense for images and other multidimensional data, but simba destroys the spacial structure by flattening the signal, which is even more confusing. Reading the paper I couldn't understand if there was a mathematical reason for why they did what they did with the chunking of a 1D sequence into a 2D sequence. Some insight into that would be nice.
Thank you very much for your excellent work! I have a question regarding the article titled "Frequency-domain MLPs are More Effective Learners in Time Series Forecasting." Could you please explain the difference between the EinFFT and the FreMLP operations mentioned in the paper?