I have a dataset, where I simulate 8 different households, and if they walked
outside that day. The next day, other people will leave their houses. They are
impacted by the other people, and the different impacts are stored in a matrix
(W=8x8). The people outside one day (in a 8 element 0/1 vector) is then
multiplied with the matrix, to get the next day. This is then simulated for
many time steps, and I also have many simulations. One simulation is in the
shape (1000 x 8), where there are 1000 time steps.
My goal is to reconstruct the matrix from the time series data. I have thought
of using a transformer based approach, and feed in only chunks from the
simulation data, eg. 32x8 from the simulation.
I have found the
torch.nn.Transformer(), but I don’t understand what src and tgt
are supposed to be.
Also: Are transformer-based methods the correct architecture to attack this problem?