I have a high frequency time series data that I want to fit into a CNN network for identification work. However I need to decide which frequency I would want to look at, so a typical way to do it is :
time_step = 10
A = cumsum(input_data)
B = A[0 : len(A): time_step].
C = B[1:] - B[:len(B)-1]
,which samples at every 10 time points.
However, I wonder if I can make a Keras customer layer that the time_step is a parameter of the network? So that the network would tell me what is the optimal value for time_step parameter.