I would like to train a network on input data with variable length (so one sample is 10x2 matrix, one is 8x2, etc.) and variable-length output -> so the length of input and output doesn't have to be the same.

This is all independent from batch dimension, etc.

Is there any architecture already developed that handles this kind of input and output formulation? If not, which ones are the closest to it?

More Helge Hecht's questions See All
Similar questions and discussions