np.empty_like not working with zeros array

When an array created using np.zeros is passed as an argument to np.empty_like, the array is initialized to 0 instead of being uninitialized!

Reproducing code example:

a = np.zeros((2, 5), dtype=np.float32)
print(a)

b = np.empty_like(a)
print(b)

Output:

[[0. 0. 0. 0. 0.]
 [0. 0. 0. 0. 0.]]
[[0. 0. 0. 0. 0.]
 [0. 0. 0. 0. 0.]]

Numpy/Python version information:

System: Linux

Numpy Version: 1.18.1

Python Version: 3.6.9

[WarrenWeckesser]

The result can depend on when print() is called. For example:

import numpy as np

a = np.zeros((2, 5), dtype=np.float32)
print(a)

b = np.empty_like(a)
print(b)

Output:

[[0. 0. 0. 0. 0.]
 [0. 0. 0. 0. 0.]]
[[0. 0. 0. 0. 0.]
 [0. 0. 0. 0. 0.]]

But this version

import numpy as np

a = np.zeros((2, 5), dtype=np.float32)
b = np.empty_like(a)

print(a)
print(b)

generates:

[[0. 0. 0. 0. 0.]
 [0. 0. 0. 0. 0.]]
[[0.00e+00 2.00e+00 0.00e+00 2.00e+00 4.20e-45]
 [1.63e-43 1.56e-43 1.60e-43 1.70e-43 0.00e+00]]

So I suspect there is nothing to be concerned about.

This is on macOS 10.12.6, with numpy 1.17.2.

[seberg]

Thanks Warren, nice catch. I will close this. This is the caching mechanism that numpy uses to avoid many small/mid-sized allocations, and thus is simply expected here currently.