The type conversion problem of array operations

a = np.array([0,255], dtype='uint8')
output：array([  0, 255], dtype=uint8)

a * 2
output：array([  0, 254], dtype=uint8) // become  the top eight after the operation

b = np.array([1, 255], dtype='uint8')
b * 256
output：array([  256, 65280], dtype=uint16) // Automatically become the type of uint16

The change of the variable type depends on the minimum of the array?

Hmm, that looks like a bug, and maybe a very old one. The array data type should dominate the scalar unless a float is involved. It may be intended behavior, but I cannot think why.

Its the old problem of "oh a scalar, lets see what the minimum type is when considering the other array", and if the arrays type happens to be good enoug, it uses that... I.e. the 256 scalar is too large for uint8, so it decides that uint16 is necessary, but 2 fits uint8, so that is taken in the first example.

Arggghhhh, I didn't know that was the policy, that's awful. I thought that within the same kind, the array type was always dominant, which is simple to understand and has the virtue that the output type only depends on the code, not the code and the input values.

Yeah, that would make more sense, at least for python types (for numpy types they could just be treated like an array too as an option, I find it annoying that if I happen to have a 0d array things might suddenly stop casting as well). And we can put a warning anyway when the python scalar cannot cast safely... There have been a lot of discussions about this IIRC, but they always somewhat died out from a lack of clarity what is right, and then the fact that changing things seems tricky, I think the last time it changed was 1.7, not quite sure what the behaviour was before.

Relevant is #10322, which here looks like:

>>> np.result_type(np.int8, 2)
dtype('int8')
>>> np.result_type(np.int8, 256)
dtype('int16')

Closing, this is effectively fixed by NEP 50 adoption in main/NumPy 2.0.

Navigation Menu

Repository navigation

The type conversion problem of array operations #11485

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

The type conversion problem of array operations #11485

Description

Activity

charris commented on Jul 3, 2018

seberg commented on Jul 3, 2018

charris commented on Jul 3, 2018

seberg commented on Jul 3, 2018

eric-wieser commented on Jul 10, 2018

seberg commented on Jan 30, 2024

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions