乘0比Python中的任何其他乘法快吗？

Question

I want to optimize matrix multiplications (weighs regression) in Python by masking some values I know I don't need in computations as zeros. They will still be there as I don't want to change size of the matrices. Matrices are floats.

Will Python (keras/tensorflow?) treat these multiplications in different way and significantly speed up the process, or it will take similar amount of time, making such masking pointless?

Answer 1

No, the multiplying by zero is the same as multiplying by any other number

>>> def times_zero(x):
...     return x * 0
... 
>>> import dis
>>> dis.dis(times_zero)
  2           0 LOAD_FAST                0 (x)
              3 LOAD_CONST               1 (0)
              6 BINARY_MULTIPLY     
              7 RETURN_VALUE        
>>> def times_four(x):
...     return x * 4
... 
>>> dis.dis(times_four)
  2           0 LOAD_FAST                0 (x)
              3 LOAD_CONST               1 (4)
              6 BINARY_MULTIPLY     
              7 RETURN_VALUE   

Answer 2

I timed them:

from timeit import default_timer as timer
import itertools

my_toggle = itertools.cycle(range(2))

for x in range(20):
    current_number = my_toggle.__next__()
    start = timer()
    y = 1 * current_number
    end = timer()
    print(f"{end - start:.10f} seconds for {current_number}")

But I'm not sure what to make of the results:

0.0000002555 seconds for 0
0.0000002555 seconds for 1
0.0000002555 seconds for 0
0.0000002555 seconds for 1
0.0000002555 seconds for 0
0.0000000000 seconds for 1
0.0000000000 seconds for 0
0.0000002555 seconds for 1
0.0000000000 seconds for 0
0.0000000000 seconds for 1
0.0000000000 seconds for 0
0.0000002555 seconds for 1
0.0000002555 seconds for 0
0.0000000000 seconds for 1
0.0000002555 seconds for 0
0.0000000000 seconds for 1
0.0000000000 seconds for 0
0.0000002555 seconds for 1
0.0000000000 seconds for 0
0.0000000000 seconds for 1