I have meant to extract features from a multi-channel time-series Data from the Fourier domain, and after a lot of tries and searching, I have not been able to implement anything. Could anyone help me out with this? The architecture what I am trying to build is something like
Data_input -> FFT -> Convolutional Layers -> Classification Layer
So you would like to extract features from frequency information and not from temporal domain which your data is currently in? I do not see a reason to perform the FFT though, the convolutional layers can behave something like band-pass filters to be sensitive to particular frequencies given the temporal data (of-course only if it helps the classification layers to classify correctly the training data).
Thanks a lot @A_A for replying. So basically what you are saying is that any kind of convolutions are sufficient?.
I was actually going for some kind of a multiScale CNN architechture where I would etract features from the temporal as well the fourier domain and combine them later. Is that kind of redundant ?
Could you please explain what is your end goal and if you have training data to achieve that goal.