Convert "NaN" -> "nan" to catch up with Numpy 1.6