I have a large 2D grid having 30k rows and 35k columns, so a total of 30x35k grid cells. Each grid cell is represented by a unique integer number (identity of grid cell). I have several trajectories that passes through these grid cells. Each trajectory is represented by a sequence of numbers (that are grid cells through which the trajectory passes through).
I want to solve the problem of trajectory prediction by giving the partial trajectory as input and predict the full trajectory. This becomes a sequence to sequence problem, where all sequences are integer values by default.
I am trying to solve this problem through encoder-decoder LSTM architecture. Most tutorials/examples regarding sequence to sequence on net are on machine translations in which vocabularies or characters are one-hot-encoded to represent the text as integer values. When I hot-encode my sequence values the one-hot vector becomes very large because there are (30x35)k grid cells, the program has given memory overflow error (because each vector has of size 1 million).
I am confused here, do I need to treat grid identity as categorical variable? because all grid identities are numeric numbers but these identities are not comparable (like prices).
Do I need to hot-encode my integer values in my sequence? Or is there any other alternative to solve this problem? I also appreciate if you suggest me the similar tutorials having the sequence to sequence prediction problem.