What pre processing should I use on data to feed into a CNN?

Question

I have a dataset of shape 105 x 501 x 266 where 105 is the number of data and 501 x 266 is the shape of 1 data i.e. The labels_dataset is of shape 105 x 1.

Each value of the 501 x 266 matrix is a complex number.

So it essentially becomes 501 * 266 * 2(real and imaginary part of the number)

And now I have to feed this data to a CNN. I 'm new to training networks. So need to know whether my data is in best possible form for the CNN or not.

I've printed out max, min, sd, mean of real part, imaginary part and magnitude of the dataset for more info:

max real = 0.186396, min real = -0.204375
max imag = 0.166608, min imag = -0.159017
max abs = 0.219019, min abs = 2.33527e-10
mean real = 4.01718e-10, complex = 6.79294e-15, abs = 8.82916e-05
std dev real = 0.000442753, complex = 0.000400677, abs = 0.000590573

Is this a good form of data for input to a CNN. What are the options to make it more suitable?

Answer 1

You should normalize your input data for different purposes. As you can read from here, Normalizing data is done for accelerating optimization. If you have features with different scales, it will take too much time for your optimizer function to find optimal points. Suppose you have age feature which can change between 0 to 150 (!) and salary which can be changed from 0 to whatever, like 500,000,000 $. your optimization algorithm used in your ML model will take too much time, if possible, to find appropriate weights for each feature. Moreover, if you don't scale your data, your ML algorithm may take too much care to features with large scales.

You have not explicitly specified your task, if its classification, you may need to turn the labels into one-hot encoded versions or if possible you may need to use word embeding for adding logical distance to different classes. If your task is a regression one, you may need to normalize your outputs. In regression problems it is customary to normalize the output too, because the scale of output and input features may differ.

Answer 2

Usually it helps to have data centered at 0 and with standard deviation 1. I would reescale it such thar its standard deviation is 1. Appart from that, everything looks good.