Question

Is it possible to lose any fidelity or precision going from a numpy float32 to a python float (float64) back to a numpy float32?

I can’t seem to find a case where there is data lost but everyone around me is claiming that the world will end because there is data lost. I just need to find a document/example that proves that data is lost before I continue.

Any help pointing me in the right direction is appreciated.

Here is a typical use case I’m seeing:

def serialize(val):
    # val is a np.float32
    return val.astype(float)


def deserialize(msg):
    return np.float32(msg)

message = '1.23456789'
outgoing = serialize(message)
incoming = deserialize(message)

Answer 1

If x is a float32 then float32(float64(x)) == x.

The one exception being if x = nan then nan != nan, although nan is nan, hence if you want to catch all you could use:

float32(float64(x)) == x or x is nan

.

You should take care:

float32(1./3)          # 0.33333334
float64(float32(1./3)) # 0.3333333432674408
float64(0.33333334)    # 0.33333333999999998