this network learned 32 bit numbers

outerdim = 32

hiddendim = 3072

save_path='model_weights.pt'

numlayers = 2

batch_size = 100

it seems when we double the outerdim, hidden dim needs to be multiplied with quite bit larger number, at least when 2 layers are used. this scalability may not suffice with 256 bit numbers, but its possible adding more layers improves scalability.