Pseudo-code for Encoder part (section 4) of http://arxiv.org/pdf/1511.06391v1.pdf

1511.06391-set2vector.py

def set2vector(x, init, f, g, h, RNN, T):
  """
  x: input data w. N samples
  init: function that initializes hidden state of RNN
  f: f(x): function that maps each sample of x to a vector. E.g. the identity function, an MLP, etc.
  g: g(m,q): function that maps vectors m and q to a single number. E.g. the dot product, an MLP, etc.
  h: h(q,r): function that maps vectors q and vector r to a vector, e.g. an MLP or similar.
  T: int, how many steps to run.
  """
  qs[0] = init() # qs[t] is a vector (RNN "post" hidden state?)
  N = len(x)
  for i in 0:N
    m[i] = f(x[i]) # x[i] and m[i] are vectors (input, embedded input)
    
  for t in 1:T
    q[t] = RNN(qs[t-1]) # q[t] is a vector (RNN "pre" hidden state?)
    d = 0 # d is a scalar (softmax denominator)
    
    for i in 0:N
      e[i][t] = exp(g(m[i], q[t])) # e[i] is a scalar ("attention importance")
      d += e[i][t] 
      
    for i in 0:N
      a[i][t] = e[i][t]/d # a[i] is a scalar ("softmaxed attention importance")
      
    r[t] = 0 # r[t] is a vector (softmax sum of embedded input)
    for i in 0:N
      r[t] += a[i][t]*m[i]
    
    qs[t] = h(q[t], r[t])
    
  return qs[T]