There were several issues with the code.

1. `Tensor.prototype.transpose` symbol no longer exist. Replaced it with
  `tf.transpose` instead.
2. KNN reuse was flawed as reused version would return a different
  number of outputs than what it would return normally.
3. `tf.split` expects all returned members to have the equal size where
it previously could not guarantee that. We now pad the matrix so the
  equal size can be formed even if we discard useless padded elements.

In the future, we may want to disable GPU based computation acceleration
as it is actually slower than just doing it on CPU.