ENH: Changed repr of np.bool_ #17592

ganesh-k13 · 2020-10-20T10:02:24Z

Changes:

Changed repr of np.bool_

After Fix:

>>> str(np.bool_(False))
'numpy.False_'
>>> str(np.bool_(True))
'numpy.True_'
>>>

Note:

I expect a lot of UT and Doc failures, will fix as we go along.

cc: @eric-wieser , @seberg

numpy/core/src/multiarray/scalartypes.c.src

ganesh-k13 · 2020-10-20T10:57:08Z

Some tests are forming cores, will send the mail to the mailing list, once that's fixed

numpy/core/src/multiarray/scalartypes.c.src

ganesh-k13 · 2020-10-21T04:53:14Z

Do macs have a different set of tests? I noticed this is failing.

ganesh-k13 · 2020-10-21T09:11:09Z

It's working in linux:

(Pdb) p numpy_f
<function any at 0x7f78f15e7310>
(Pdb) p numpy_f(a, axis=1, keepdims=True)[...,:-1]
masked_array(
  data=[[[True, False, True]],

        [[True, False, True]]],
  mask=[[[False, False, False]],

        [[False, False, False]]],
  fill_value=numpy.True_)

> /home/ganesh/open-source/numpy/build/testenv/lib/python3.8/site-packages/numpy/ma/tests/test_core.py(4948)testkeepdims()
-> assert_equal(ma_f(a, keepdims=True).shape,
(Pdb) p ma_f(a, axis=1, keepdims=True)[...,:-1]
masked_array(
  data=[[[True, False, True]],

        [[True, False, True]]],
  mask=[[[False, False, False]],

        [[False, False, False]]],
  fill_value=numpy.True_)
(Pdb) p f
'any'

numpy/numpy/ma/tests/test_core.py

Lines 4958 to 4960 in 32f1359

    
           for f in ['sum', 'prod', 'mean', 'var', 'std']: 
        
               testaxis(f, a, d) 
        
               testkeepdims(f, a, d)

Any idea why it's failing only on Mac?

eric-wieser · 2020-10-21T09:30:22Z

The test that is failing is a documentation test, and I guess the documentation tests are only run on mac? You need to update the docstrings to use the new printing.

mattip · 2020-10-21T10:23:59Z

The incantation is python runtests.py -g --refguide-check

ganesh-k13 · 2020-10-24T06:06:19Z

I have emailed the mailing list (nabble link, pipermail link)for their opinion @eric-wieser, as mentioned in the issue.

eric-wieser

Can you add an "improvements" release note, like we did when we improved void.__repr__ with https://numpy.org/doc/stable/release/1.14.0-notes.html#void-datatype-elements-are-now-printed-in-hex-notation?

ganesh-k13 · 2020-11-01T22:04:17Z

@eric-wieser , any more changes needed?

doc/release/upcoming_changes/17592.improvement.rst

eric-wieser

Looks good to me, and the mailing list seems not to care. Will leave this for another maintainer to push the button on if they agree.

I reworded the release note slightly to make clear also what did not change.

ganesh-k13 · 2020-11-01T22:22:00Z

Thanks Eric, sure

seberg · 2020-11-02T14:58:23Z

Happy to put this in as such, although I wonder a bit whether this has the potential to be a slightly noisy transition, and 1.20 is already a big release.

hameerabbasi · 2020-11-04T19:34:41Z

This is different from every other NumPy scalar, np.int*, np.float*, ...

Also reprs don't usually include the module.

seberg · 2020-11-04T19:37:10Z

We are a bit unsure whether this is a good idea at all, could you propose this on the mailing list, since it is an API change?

eric-wieser · 2020-11-04T20:54:57Z

This already hit the mailing list and was meet with silence

mattip · 2020-11-18T19:44:24Z

In a triage discussion we seem to prefer fixing this across all scalars and not just np._bool, if we are to change anything at all

ganesh-k13 · 2020-11-19T04:53:51Z

Sounds good @mattip , I'll use this PR to change the repr of all classes. When time comes, we can merge all together.

ganesh-k13 · 2020-11-20T10:01:03Z

I have made some change to reflect like this:

>>> np.int64(42)
numpy.int64(42)

Only for ints as of now. If it's breaking too much stuff let's put a hold on this.

eric-wieser · 2020-11-20T10:03:22Z

numpy/core/src/multiarray/scalartypes.c.src

+    if ((dtype = PyObject_GetAttrString(self, "dtype")) == NULL) {
+        return NULL;
+    }
+
+    if ((dtype_name = PyObject_GetAttrString(dtype, "name")) == NULL) {
+        return NULL;
+    }
+
+    item_str = PyUString_FromFormat("numpy.%S(%S)", dtype_name, item);


This feels like overkill - does the following work?

Suggested change

if ((dtype = PyObject_GetAttrString(self, "dtype")) == NULL) {

return NULL;

}

if ((dtype_name = PyObject_GetAttrString(dtype, "name")) == NULL) {

return NULL;

}

item_str = PyUString_FromFormat("numpy.%S(%S)", dtype_name, item);

item_str = PyUString_FromFormat("%s(%S)", Py_TYPE(self)->tp_name, item);

This is much better 😄 . Didn't think of this, thanks.

@eric-wieser , so just a doubt regarding debugging, how do you gdb into that code? currently, I put raise(SIGINT) and do gdb python, then do run code.py that hits that. Is there a better way?
[edit] or any numpy specific tools/best practice to debug

I haven't had to jump into GDB in a while, so I'm afraid I don't remember my usual trick. Are you asking because this is giving the wrong result?

Oh I see, I was just curious about how numpy folks debug in general, some uncommon techniques/tools, etc to simplify stuff than just raw gdb. The results are as expected, thanks

charris · 2021-04-12T22:23:53Z

This needs a rebase.

charris · 2021-04-12T22:25:02Z

Lot of test failures, but the logs have expired.

mattip · 2022-04-18T18:09:22Z

Summarizing the current status:

this PR should be retitled to reflect the pivot to changing the repr of all scalars
tests are failing and it is still marked as a draft
@eric-wieser could you remove your approval of the PR, it is confusing when looking at PR statuses.

github-actions bot added the 01 - Enhancement label Oct 20, 2020