My input has the shape of n rows (time steps) and m columns (attributes). I want to train a convolutional neural network on it to predict a class.
I am currently using 1D-Convolutions. I got a good score, but I know that the relevant patterns in my data are occuring along in just a single (but everytime) different column over time. I dont need to know what the values of the other columns are, because the pattern is not distributed over the other attributes. I just interested in a filter (convolution) for each attribute, because when I learn that filter for my first attribute, I want that it can be used also for the other attributes. Since they can have the same pattern at test time but in training it was in the first column and during test it is in the second. My issue is that this is not the case of a normal 1D-Convolution because it goes through time over all columns.
What I need is the type in keras or tensorflow for a convolution that goes along only one column through time (and not multiple, as the case with 1D-Convolutions).
Thanks in advance for your help!