pairwise functions for TF

tf_pairwise.py

def pdist(X):
  """
  Computes pairwise distance between each pair of points
  Args:
    X - [N,D] matrix representing N D-dimensional vectors 
  Returns:
    [N,N] matrix of (squared) Euclidean distances
  """
  x2 = tf.reduce_sum(X * X, 1, True)
  return x2 - 2 * tf.matmul(X, tf.transpose(X)) + tf.transpose(x2)

def cdist(X, Y):
  """
  Computes pairwise distance between each pair of points in two sets
  Args:
    X - [N,D] matrix representing N D-dimensional vectors 
    Y - [M,D] matrix representing N D-dimensional vectors     
  Returns:
    [N,M] matrix of (squared) Euclidean distances
  """  
  return (tf.reduce_sum(X * X, 1, True) -
          2 * tf.matmul(X, tf.transpose(Y)) + 
          tf.transpose(tf.reduce_sum(Y * Y, 1, True)))

def contrastive_loss(D, labels, margin):
  """
  computes contrastive loss:
  L = 0.5 * (1 - labels) * D + 0.5 * labels * {max(0, margin - D^0.5)}^2
  where D is a matrix of (squared) euclidean distances
  Args:
    D - [N,N] tensor, euclidean distances
    labels - [N,N] tensor, labels {0,1}
    margin - scalar, contrastive margin
  Returns:
    scalar, the average loss
  """
  # TODO: we definitely need some kind of weight/mask for the samples
  loss = (tf.mul((1.0 - labels), D) + 
          tf.mul(labels, tf.square(tf.maximum(0.0, margin - tf.sqrt(D)))))
  loss = 0.5 * tf.reduce_mean(loss)
  return loss