fscore_noodling.py 文件源码

def expected_fscore(n_predicted, n=100, prob_per_instance=.01):
  #assert n_predicted in (1, n)
  ppi = prob_per_instance
  ppibar = (1 - prob_per_instance)
  def fscore(fn, tp, fp):
    num = 2*tp
    denom = 2*tp + fp + fn
    return num / denom

  parts = []
  for ntrue in range(1, n+1):
    # ntrue := this many were actually true in the gold standard
    if n_predicted == 1:
      tp = 1
      fn = ntrue - 1
      fp = 0
      # Prob. of our 1 predicted label being true
      prob_weight = ppi
      # Prob of ntrue-1 other labels being true out of n-1
      prob_weight *= binom(n-1, ntrue-1) * ppi**(ntrue-1) * ppibar**(n-ntrue)
    elif n_predicted == n:
      tp = ntrue
      fn = 0
      fp = n - ntrue
      prob_weight = binom(n, ntrue) * ppi**ntrue * ppibar**(n-ntrue)
    else:
      assert False, "fuck it"
      for tp in range(1, ntrue):
        fp = n_predicted - tp
        fn = ntrue - tp


    fs = fscore(fn, tp, fp)
    dbg('fscore = {}'.format(fs))
    dbg('prob weight = {}'.format(prob_weight))
    res = fs * prob_weight
    parts.append(res)
  return parts