[BUG] - Possible broadcast issue with empty arrays

[wired8]

Describe the bug In cpython it seems that some binary operations involving an empty numpy array return an empty numpy array.

This difference in behavior is a problem when porting code.

To Reproduce

print(np.array([])+np.array([1.]))
>> array([1.0], dtype=float64)

Expected behavior

print(np.array([])+np.array([1.]))
>> array([], dtype=float64)

Notes: https://stackoverflow.com/questions/58854669/python-list-empty-numpy-array-empty-numpy-array

[v923z]

Describe the bug In cpython it seems that some binary operations involving an empty numpy array return an empty numpy array.

It is some, or all? The empty array is not a very well documented feature of numpy, and it's tripped up in the past.

[wired8]

Looking like some, sorry should have been more specific.

print(np.array([])+np.array([1.]))
>>> array([1.0], dtype=float64)

print(np.array([])-np.array([1.]))
>>> array([-1.0], dtype=float64)

print(np.array([])/np.array([1.]))
>>> array([0.0], dtype=float64)

print(np.array([])*np.array([1.]))
>>> array([0.0], dtype=float64)

a = np.array([])
a *= 1
print(a)
>>> array([], dtype=float64)

b = np.array([])
b /= 1
print(b)
>>> array([], dtype=float64)

c = np.array([])
c += 1
print(c)
>>> array([], dtype=float64)

d = np.array([])
d -= 1
print(d)
>>> array([], dtype=float64)

[v923z]

If I run this with CPython

try:
    from ulab import numpy as np
except:
    import numpy as np

print(np.array([])+np.array([1.]))

print(np.array([])-np.array([1.]))

print(np.array([])/np.array([1.]))

print(np.array([])*np.array([1.]))

a = np.array([])
a *= 1
print(a)

b = np.array([])
b /= 1
print(b)

c = np.array([])
c += 1
print(c)

d = np.array([])
d -= 1
print(d)

I get this:

[]
[]
[]
[]
[]
[]
[]
[]

On the other hand, in ulab, the output is

array([1.0], dtype=float64)
array([-1.0], dtype=float64)
array([0.0], dtype=float64)
array([0.0], dtype=float64)
array([], dtype=float64)
array([], dtype=float64)
array([], dtype=float64)
array([], dtype=float64)

So, I believe, all operators result in an empty array, if one of the operands is empty. Perhaps, the easiest fix is to check, whether the array is empty at the very beginning, and return immediately, if so.

[wired8]

That has been my current approach, but given my use case and the share amount of code I have I'm concerned with having to add this check. What is the overhead for checking this c side?

[v923z]

That has been my current approach, but given my use case and the share amount of code I have I'm concerned with having to add this check. What is the overhead for checking this c side?

Oh, I meant that this should obviously be done in firmware, don't worry! My comment was sort of a question as to whether this would be enough to ensure compatibility.

[wired8]

Thanks @v923z your a champion!

[v923z]

I think, https://github.com/v923z/micropython-ulab/pull/507 should rectify the problem. As soon as you can confirm it, I would merge this.