.dtype.type class object is not preserved over some operations

import numpy as np

for dt in (np.uint8, np.uint16, np.uint32, np.uint64,
           np.int8, np.int16, np.int32, np.int64,
           np.float16, np.float32, np.float64):
    x = np.array([0, 1], dtype=dt)
    if np.issubdtype(dt, np.floating):
        y = np.finfo(dt).max - x
    else:
        y = np.iinfo(dt).max - x

    id_x = hex(id(x.dtype.type))
    id_y = hex(id(y.dtype.type))

    print('{}: {}, {}, {}'.format(dt, id_x, id_y, id_x == id_y))

<class 'numpy.uint8'>: 0x7fa0e529f740, 0x7fa0e529f740, True
<class 'numpy.uint16'>: 0x7fa0e529f5a0, 0x7fa0e529f5a0, True
<class 'numpy.uint32'>: 0x7fa0e529f400, 0x7fa0e529f400, True
<class 'numpy.uint64'>: 0x7fa0e529f260, 0x7fa0e529f0c0, False
<class 'numpy.int8'>: 0x7fa0e529ff60, 0x7fa0e529ff60, True
<class 'numpy.int16'>: 0x7fa0e529fdc0, 0x7fa0e529fdc0, True
<class 'numpy.int32'>: 0x7fa0e529fc20, 0x7fa0e529fc20, True
<class 'numpy.int64'>: 0x7fa0e529fa80, 0x7fa0e529f8e0, False
<class 'numpy.float16'>: 0x7fa0e529ef20, 0x7fa0e529ef20, True
<class 'numpy.float32'>: 0x7fa0e529ed80, 0x7fa0e529ed80, True
<class 'numpy.float64'>: 0x7fa0e529ebe0, 0x7fa0e529ebe0, True

This could lead to some issues in users code heavily relying on dtype.type comparison (e.g. scikit-image/scikit-image#3043).

Linux, x64, numpy 1.14.2.

This is probably related to #10151. There's a fifth integer type that's not in your list there - you should iterate over [np.byte, np.short, np.intc, np.int_, np.longlong] + [np.ubyte, np.ushort, np.uintc, np.uint, np.ulonglong] to hit all the types

A better test now that gh-10151 is in:

import numpy as np

for dt in (np.ubyte, np.ushort, np.uintc, np.uint, np.ulonglong,
           np.byte, np.short, np.intc, np.int_, np.longlong,
           np.half, np.single, np.double):
    x = np.array([0, 1], dtype=dt)
    if np.issubdtype(dt, np.floating):
        y = np.finfo(dt).max - x
    else:
        y = np.iinfo(dt).max - x

    xt = x.dtype.type
    yt = y.dtype.type

    print('{}: {}, {}, {}'.format(dt, xt, yt, xt is yt))

On 64 bit windows, this works as expected:

<class 'numpy.uint8'>: <class 'numpy.uint8'>, <class 'numpy.uint8'>, True
<class 'numpy.uint16'>: <class 'numpy.uint16'>, <class 'numpy.uint16'>, True
<class 'numpy.uintc'>: <class 'numpy.uintc'>, <class 'numpy.uintc'>, True
<class 'numpy.uint32'>: <class 'numpy.uint32'>, <class 'numpy.uint32'>, True
<class 'numpy.uint64'>: <class 'numpy.uint64'>, <class 'numpy.uint64'>, True
<class 'numpy.int8'>: <class 'numpy.int8'>, <class 'numpy.int8'>, True
<class 'numpy.int16'>: <class 'numpy.int16'>, <class 'numpy.int16'>, True
<class 'numpy.intc'>: <class 'numpy.intc'>, <class 'numpy.intc'>, True
<class 'numpy.int32'>: <class 'numpy.int32'>, <class 'numpy.int32'>, True
<class 'numpy.int64'>: <class 'numpy.int64'>, <class 'numpy.int64'>, True
<class 'numpy.float16'>: <class 'numpy.float16'>, <class 'numpy.float16'>, True
<class 'numpy.float32'>: <class 'numpy.float32'>, <class 'numpy.float32'>, True
<class 'numpy.float64'>: <class 'numpy.float64'>, <class 'numpy.float64'>, True

Can someone test this on linux (where int_ is 64-bit not 32-bit)?

Output on 64bit linux:

<class 'numpy.uint8'>: <class 'numpy.uint8'>, <class 'numpy.uint8'>, True
<class 'numpy.uint16'>: <class 'numpy.uint16'>, <class 'numpy.uint16'>, True
<class 'numpy.uint32'>: <class 'numpy.uint32'>, <class 'numpy.uint32'>, True
<class 'numpy.uint64'>: <class 'numpy.uint64'>, <class 'numpy.ulonglong'>, False
<class 'numpy.ulonglong'>: <class 'numpy.ulonglong'>, <class 'numpy.ulonglong'>, True
<class 'numpy.int8'>: <class 'numpy.int8'>, <class 'numpy.int8'>, True
<class 'numpy.int16'>: <class 'numpy.int16'>, <class 'numpy.int16'>, True
<class 'numpy.int32'>: <class 'numpy.int32'>, <class 'numpy.int32'>, True
<class 'numpy.int64'>: <class 'numpy.int64'>, <class 'numpy.longlong'>, False
<class 'numpy.longlong'>: <class 'numpy.longlong'>, <class 'numpy.longlong'>, True
<class 'numpy.float16'>: <class 'numpy.float16'>, <class 'numpy.float16'>, True
<class 'numpy.float32'>: <class 'numpy.float32'>, <class 'numpy.float32'>, True
<class 'numpy.float64'>: <class 'numpy.float64'>, <class 'numpy.float64'>, True

Does replacing np.iinfo(dt).max with 2 give the same results? Would be nice to narrow down the issue

Ah, did not look at it carefully. Yes, that will fix things,

To be clear: Replacing by two gives the correct output (i.e. no False prints).

Just fired up WSL with numpy 1.11 (!), and I think I can pinpoint the bug:

assert np.result_type(np.iinfo(np.int64).max, np.int64).type is np.int64  # fails
assert np.result_type(np.iinfo(np.longlong).max, np.longlong).type is np.longlong # passes

In my opinion, we should swap between these which one passes and which fails.

We're being inconsistent about when we pick long and when we pick long long for int64 on linux. By name, long takes precedence - but it seems that when assigning dtypes to scalars based on their value, we prefer long long.

On windows, it seems we handle int vs long for int32 consistently.

Yeah, the bug is that it falls to the unsigned path, and that fallback does not convert the original typenum to its unsigned version, but instead always uses ulonglong. If you use .min instead of max, it works ;).

The above is still there, I don't want to consider it a serious bug and NEP 50 adoption means that the repro's now all pass, since we ignore the oddity of the Python integer size now. (i.e. the path still exists, but it isn't hit here)

Navigation Menu

Repository navigation

`.dtype.type` class object is not preserved over some operations #11020

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

.dtype.type class object is not preserved over some operations #11020

Description

Activity

eric-wieser commented on Apr 30, 2018

eric-wieser commented on Sep 13, 2019

seberg commented on Sep 13, 2019

eric-wieser commented on Sep 13, 2019

seberg commented on Sep 13, 2019

seberg commented on Sep 13, 2019

eric-wieser commented on Sep 13, 2019

seberg commented on Sep 13, 2019

seberg commented on Nov 3, 2023

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions

`.dtype.type` class object is not preserved over some operations #11020