hamming_raml.py

def ComputeHammingCDF(len_target, temprature, vocab):
  max_edits = len_target + 1  # we allow between 0 and len_target subs
  a = np.zeros(max_edits)
  for n_subs in range(max_edits):
    count_n_subs = []
    tot_edits = misc.comb(len_target, n_subs)
    a[n_subs] = np.log(tot_edits) + n_subs * np.log(len(vocab) - 1) # number of sequences: tot_edits * (N-1) ^ n_subs
    a[n_subs] += - n_subs / float(temprature) * np.log(len(vocab) - 1) - n_subs / float(temprature)   # tot_edits * (N-1) ^ n_subs * ((N-1)e) ^ (-n_subs / T)
  p_subs = a - np.max(a)
  p_subs = np.exp(p_subs)
  p_subs /= np.sum(p_subs)
  p_hamming_cdf = np.cumsum(p_subs)
  return p_hamming_cdf


# one can precompute hamming CDFs for different sequence lengths
subs_cdf = []
for len_target in range(200):  # assuming maximum length is 200
  p_subs_cdf = ComputeHammingCDF(len_target, temprature = augment_target_prob, vocab = vocab)
  subs_cdf.append(p_subs_cdf)


def SubstitutionSampling(s, temprature, hamming_cdf, vocab):
'''
Sample one sequence from the vicinity of a given target sequence s.
    A string t is sampled proportionally to exp{-hamming_distance(t, s) / temprature}

Args:
      s: numpy array of a sequence which is output of a seq2seq/crf model e.g. POS tag sequence
      temprature: temprature of sampling
      hamming_cdf: precomputed edit CDF
      vocab: the vocabulary elements that are allowed for substitution

    Returns:
      numpy array of a sampled sequence t
'''
  assert(min(vocab) >= 0)
  len_target = len(s) - 1
  p_hamming_cdf = hamming_cdf[len_target]
  # sample
  rand_n_subs = np.sum(np.random.rand() >= p_hamming_cdf)
  # apply changes
  t = copy.copy(s)
  perm = np.random.permutation(len_target)
  subs = perm[:rand_n_subs]
  for i in subs:
    while True:
      rand_char = vocab[np.random.randint(len(vocab))]
      if not t[i] == rand_char:
        break
    t[i] = rand_char
  return t