I'm battling some floating point problems in Pandas read_csv function. In my investigation, I found this:
In [15]: a = 5.9975
In [16]: a
Out[16]: 5.9975
In [17]: np.float64(a)
Out[17]: 5.9974999999999996
Why is builtin float of Python and the np.float64 type from Python giving different results? I thought they were both C++ doubles?
>>> numpy.float64(5.9975).hex()
'0x1.7fd70a3d70a3dp+2'
>>> (5.9975).hex()
'0x1.7fd70a3d70a3dp+2'
They are the same number. What differs is the textual representation obtained via by their __repr__ method; the native Python type outputs the minimal digits needed to uniquely distinguish values, while NumPy code before version 1.14.0, released in 2018 didn't try to minimise the number of digits output.
__repr__() method or its C-level equivalent, yes.
5.9975.
Numpy float64 dtype inherits from Python float, which implements C double internally. You can verify that as follows:
isinstance(np.float64(5.9975), float) # True
So even if their string representation is different, the values they store are the same.
On the other hand, np.float32 implements C float (which has no analog in pure Python) and no numpy int dtype (np.int32, np.int64 etc.) inherits from Python int because in Python 3 int is unbounded:
isinstance(np.float32(5.9975), float) # False
isinstance(np.int32(1), int) # False
np.float64 at all?np.float64 defines most of the attributes and methods in np.ndarray. From the following code, you can see that np.float64 implements all but 4 methods of np.array:
[m for m in set(dir(np.array([]))) - set(dir(np.float64())) if not m.startswith("_")]
# ['argpartition', 'ctypes', 'partition', 'dot']
So if you have a function that expects to use ndarray methods, you can pass np.float64 to it while float doesn't give you the same.
For example:
def my_cool_function(x):
return x.sum()
my_cool_function(np.array([1.5, 2])) # <--- OK
my_cool_function(np.float64(5.9975)) # <--- OK
my_cool_function(5.9975) # <--- AttributeError
read_csvfunction employs its own super-fast string-to-float conversion that is not correctly rounded. Thus after exporting a value and re-reading it, the recovered value may end up being 1 or 2 ulps different from the original.read_exceltoo?