How to get the exact memory representation of fp16,fp32,bf16,int8,... and so on in python?

Question

Is there any function that can give me the exact representation (as it is stored in memory) of a number ? What I am looking for is a function with the following behaviour:

memory_repr(3,'int8') = oooooo11

memory_repr(3,'fp16') = 01000000010000000000000000000000

Basically, I want a python api to do what https://www.h-schmidt.net/FloatConverter/IEEE754.html does, that works with any data type

I tried math.frexp comes close but not quite it.

Answer 1

import struct

def memory_repr(number, data_type):
    if data_type == 'int8':
        packed = struct.pack('b', number)  
    elif data_type == 'uint8':
        packed = struct.pack('B', number) 
    elif data_type == 'fp32':
        packed = struct.pack('f', number)  
    elif data_type == 'fp16':
        packed = struct.pack('e', number)  
    else:
        raise ValueError("Unsupported data type")
    packed = struct.pack('>' + format_char, number)
    return ''.join(f'{byte:08b}' for byte in packed)

print(memory_repr(3, 'int8'))  
print(memory_repr(3, 'uint8'))  
print(memory_repr(3.0, 'fp16')) 
print(memory_repr(3.0, 'fp32')) 

Output :

00000011
00000011
0100001000000000
01000000010000000000000000000000

Here I am not sure about bf16.

Corrected for float as rightfully highlighted