Adds literal node #865

jordan-schneider · 2022-02-26T02:21:24Z

Adds literal node, solving #422

I tried to implement this as a wrapper around EnumNode, but this turned out to be impossible. Literal can accept int, str, bool and Enum values, but enum keys must be strings. So for example

from dataclasses import dataclass

@dataclass
class Config:
  mode: Literal["5", 5]

there is no way to map both of these values to separate enum keys without also causing a collision somewhere else.

This PR validates a LiteralNode value by insisting on both = equality and type equality. Both are needed, otherwise python's promiscuous primitive equalities cause problems e.g. 1.0 is valid under Literal[True].

I tried to add tests wherever EnumNode had tests, but I'm unfamiliar with the codebase. Please let me know if additional tests are needed.

lgtm-com · 2022-02-26T02:27:48Z

This pull request introduces 1 alert when merging 9338386 into fd06224 - view on LGTM.com

new alerts:

1 for `__eq__` not overridden when adding attributes

jordan-schneider · 2022-02-26T02:45:31Z

Okay I clearly missed a bunch of tests, working on that. Not sure what's up with mypy not thinking Literal exists.

…rsion

jordan-schneider · 2022-02-26T07:01:52Z

The literal thing was because setup.cfg sets the mypy version to 3.6 by default. I had a legitimate problem where I need to import from typing_extensions before 3.8, which I now do. I also added a dependency to the pre-commit hooks to fix a bunch of mypy errors locally, although this didn't appear to show up in the last round of circleci results.

lgtm-com · 2022-02-26T07:12:33Z

This pull request introduces 1 alert when merging 6d302ab into fd06224 - view on LGTM.com

new alerts:

1 for `__eq__` not overridden when adding attributes

jordan-schneider · 2022-02-26T07:55:29Z

Looks like I've ironed out all the python version issues.

lgtm-com · 2022-02-26T07:57:16Z

This pull request introduces 1 alert when merging 3c115d7 into fd06224 - view on LGTM.com

new alerts:

1 for `__eq__` not overridden when adding attributes

lgtm-com · 2022-02-26T16:57:35Z

This pull request introduces 1 alert when merging 1592df4 into fd06224 - view on LGTM.com

new alerts:

1 for `__eq__` not overridden when adding attributes

Jasha10 · 2022-03-10T15:20:16Z

Thanks very much @jordan-schneider.
Unfortunately it may be some time before we are able to give this PR a serious review as the Hydra/OmegaConf team has its hands full at the moment.

tmke8

Hi, I'm the one who originally opened #422. So I'm very happy you are doing this!

I'm obviously not an official member of OmegaConf, but I thought I would try to help you with a code review anyway.

Some questions:

What happens if I annotate a node with just Literal without any type arguments for Literal?
Don't we have to add something like typing_extensions;python_version<'3.8' to requirements/base.txt? Or is typing_extensions a dependency anyway?

omegaconf/_utils.py

tmke8 · 2022-03-13T17:33:28Z

omegaconf/_utils.py

+ or (
+ "typing_extensions.Literal" in str(type_) and not isinstance(type_, str)
+ )


Hmm, why is this needed? Isn't Literal always from typing_extensions in this else branch?

So what's happening here is that:

python 3.8 and later has typing.Literal which has an __origin__ attribute
python 3.7 typing_extensions.Literal has an __origin__ attribute
python 3.6 and earlier typing_extensions.Literal does not have an __origin__ attribute or any equivalent attribute

AND

A = Literal[5] A is Literal # False isinstance(A, Literal) # TypeError: isinstance() arg 2 must be a type or tuple of types issubclass(type(A), Literal) # TypeError: issubclass() arg 2 must be a class or tuple of classes

so as far as I can tell, for 3.6 and earlier the only way to check for Literalness is using the string. This is gross but I don't know of a better way.

This logic is unnecessarily complicated though. How do you feel about

def is_literal_annotation(type_: Any) -> bool: origin = getattr(type_, "__origin__", None) return origin is Literal or ("typing_extensions.Literal" in str(type_) and not isinstance(type_, str))

There is an is_literal_type() function here which also work with Python 3.6. The solution to the problem seems very stupid, you have to import a hidden class _Literal:

Python 3.6.13 |Anaconda, Inc.| (default, Jun 4 2021, 14:25:59) [GCC 7.5.0] on linux Type "help", "copyright", "credits" or "license" for more information. >>> from typing_extensions import Literal, _Literal >>> a = Literal[0] >>> type(a) is Literal False >>> type(a) is _Literal True

I have no idea why.

Before 3.6 typing_extensions creates the internal _Literal class using a typing._FinalTypingBase mixin https://github.com/python/typing/blob/master/typing_extensions/src/typing_extensions.py#L344 , and then Literal is an instantiation of that class. So type(a) returns the outer _Literal class instead of the instantiation or something. Anyway this seems less bad than the string hack, so I'll go with this.

tmke8 · 2022-03-13T17:37:43Z

omegaconf/nodes.py

+ if hasattr(self.literal_type, "__args__"):
+ self.fields = list(self.literal_type.__args__) # pragma: no cover
+ elif hasattr(self.literal_type, "__values__"): # pragma: no cover
+ self.fields = list(self.literal_type.__values__) # pragma: no cover


Maybe make these branches explicitly correspond to python versions. So I think

Suggested change

if hasattr(self.literal_type, "__args__"):

self.fields = list(self.literal_type.__args__) # pragma: no cover

elif hasattr(self.literal_type, "__values__"): # pragma: no cover

self.fields = list(self.literal_type.__values__) # pragma: no cover

if sys.version_info >= (3, 7):

self.fields = list(self.literal_type.__args__) # pragma: no cover

else:

self.fields = list(self.literal_type.__values__) # pragma: no cover

but I'm not really sure about this.

Why do you feel this is better? Philosophically I tend to think that directly checking is the right way to do things, in case e.g. the attribute name every switches back or something equally cursed. This also has the benefit that mypy/pylance/other syntax parsers can correctly infer that the attribute exists, which is left implicit if you're going by version.

If this is about readability, I'd be happy to add comments saying why the check is necessary.

Right, I was just thinking that at some point in the future, omegaconf might want to drop support for Python 3.6 and then it would be useful that you can directly see that this if-else check has become unnecessary.

But, yes, a comment is likely already sufficient.

Co-authored-by: Thomas MK <[email protected]>

jordan-schneider · 2022-03-13T20:23:45Z

Hi, I'm the one who originally opened #422. So I'm very happy you are doing this!

I'm obviously not an official member of OmegaConf, but I thought I would try to help you with a code review anyway.

Some questions:
1. What happens if I annotate a node with just `Literal` without any type arguments for `Literal`?

2. Don't we have to add something like `typing_extensions;python_version<'3.8'` to `requirements/base.txt`? Or is `typing_extensions` a dependency anyway?

I don't have time at the moment to directly check this but Literal.__values__/__args__ still exists but is empty, so the LiteralNode is constructed and whatever the value is it fails to validate. This seems like the correct behavior. I'll write a test for this when I have a moment.
I think it's a dependency of one of the other dependencies but it should be added explicitly anyway, in case that changes. Again I'll handle this when I have a moment.

…hon 3.6 and earlier. Adds explicit dependency on typing_extensions. is_literal_type no longer uses string checking hack.

lgtm-com · 2022-03-14T15:48:53Z

This pull request introduces 1 alert when merging a0e2b35 into 1d622e7 - view on LGTM.com

new alerts:

1 for `__eq__` not overridden when adding attributes

jordan-schneider · 2022-03-14T15:52:54Z

All of the CI lints are failing due to

nox > [2022-03-14 15:46:08,598] mypy . --strict --install-types --non-interactive
nox > [2022-03-14 15:46:23,545] Command mypy . --strict --install-types --non-interactive failed with exit code 1:
Collecting types-dataclasses
  Downloading types_dataclasses-0.6.4-py3-none-any.whl (2.7 kB)
Collecting types-setuptools
  Downloading types_setuptools-57.4.10-py3-none-any.whl (27 kB)
Installing collected packages: types-setuptools, types-dataclasses
Successfully installed types-dataclasses-0.6.4 types-setuptools-57.4.10
tests/test_update.py:202: error: Unused "type: ignore" comment

Which I can't replicate locally. If I remove that type: ignore comment locally I do in fact get an error. Anyone know what's going on here?

Jasha10 · 2022-03-14T21:43:14Z

If I remove that type: ignore comment locally I do in fact get an error. Anyone know what's going on here?

@jordan-schneider maybe upgrade the version of mypy that you have installed locally?

Jasha10 · 2022-03-24T05:02:50Z

If I remove that type: ignore comment locally I do in fact get an error. Anyone know what's going on here?

The CI should pass if you rebase this branch against the master branch to incorporate changes from #878.

…rsion

Co-authored-by: Thomas MK <[email protected]>

…hon 3.6 and earlier. Adds explicit dependency on typing_extensions. is_literal_type no longer uses string checking hack.

lgtm-com · 2022-03-26T16:17:17Z

This pull request introduces 1 alert when merging d321a4a into c0197e5 - view on LGTM.com

new alerts:

1 for `__eq__` not overridden when adding attributes

lgtm-com · 2022-03-26T16:37:36Z

This pull request introduces 1 alert when merging b69fd38 into c0197e5 - view on LGTM.com

new alerts:

1 for `__eq__` not overridden when adding attributes

jordan-schneider · 2022-03-26T16:38:19Z

I think it was mostly outdated local versions of tools. The pre-commit version of black and the current version used by CircleCI also disagreed, so I updated the pre-commit version as well.

codekansas · 2022-09-07T22:01:53Z

Are there plans to merge this? It would be super useful!

Jasha10 · 2022-09-22T18:44:41Z

Thanks for the nudge, @codekansas!

Are there plans to merge this?

Yes. We haven't been able to finalize this PR yet as the OmegaConf team's bandwidth is limited. It is on the roadmap.

indigoviolet · 2023-08-14T07:15:11Z

@Jasha10 nudge again to enquire about plans/status since some time has passed

Queuecumber · 2023-08-31T13:08:34Z

@omry @Jasha10 would appreciate some eyes on this, it's been in limbo for a year and a half and would be a nice way to produce clean structured configs

daturkel · 2023-09-27T19:03:02Z

+1 this feature would be great and feels more pythonic than Enums.

spfrommer · 2024-03-05T03:28:29Z

+1

gsq7474741 · 2024-03-05T19:28:03Z

+1

jbohnslav · 2024-03-06T14:54:47Z

+1, still relevant

SheaML · 2024-04-04T21:01:12Z

+1

jkulhanek · 2024-04-30T08:10:23Z

+1

dkmiller · 2024-05-24T17:55:58Z

Another gentle nudge, this would be very nice to have!

jordan-schneider added 2 commits February 25, 2022 17:54

Adds literal node

3d3ae1d

Missed a not

9338386

Adds additional LiteralNode tests. Imports Literal based on python ve…

b78d4dc

…rsion

Adds no cover pragma to typing import branch

6d302ab

Adds additional fixes for earlier python versions.

3c115d7

Used the wrong type

1592df4

Jasha10 linked an issue Mar 7, 2022 that may be closed by this pull request

Allow the use of typing.Literal as a pythonic alternative to Enums #422

Open

tmke8 reviewed Mar 13, 2022

View reviewed changes

Jordan Schneider and others added 2 commits March 13, 2022 14:52

Update omegaconf/_utils.py

3b8a82f

Co-authored-by: Thomas MK <[email protected]>

Update omegaconf/_utils.py

80f9861

Co-authored-by: Thomas MK <[email protected]>

jordan-schneider added 3 commits March 14, 2022 10:36

Throws validation error when bare Literal annotation is given for pyt…

0737ec5

…hon 3.6 and earlier. Adds explicit dependency on typing_extensions. is_literal_type no longer uses string checking hack.

Merge remote-tracking branch 'upstream/master'

eb4b20c

Fixes typo in typing_extensions requirement.

a0e2b35

jordan-schneider added 4 commits March 26, 2022 11:06

Adds literal node

f3b3f75

Missed a not

694a032

Adds additional LiteralNode tests. Imports Literal based on python ve…

1bd9c93

…rsion

Adds no cover pragma to typing import branch

fb32fa5

jordan-schneider and others added 7 commits March 26, 2022 11:10

Adds additional fixes for earlier python versions.

89414a8

Used the wrong type

c86cec5

Update omegaconf/_utils.py

59eb792

Co-authored-by: Thomas MK <[email protected]>

Update omegaconf/_utils.py

5d1b801

Co-authored-by: Thomas MK <[email protected]>

Throws validation error when bare Literal annotation is given for pyt…

e2d950f

…hon 3.6 and earlier. Adds explicit dependency on typing_extensions. is_literal_type no longer uses string checking hack.

Fixes typo in typing_extensions requirement.

4fa9593

Merge branch 'master' of github.com:jordan-schneider/omegaconf

d321a4a

Formats according to new black version

b69fd38

daturkel mentioned this pull request Sep 27, 2023

Allow the use of typing.Literal as a pythonic alternative to Enums #422

Open

Adds literal node #865

Are you sure you want to change the base?

Adds literal node #865

Conversation

jordan-schneider commented Feb 26, 2022

lgtm-com bot commented Feb 26, 2022

jordan-schneider commented Feb 26, 2022

jordan-schneider commented Feb 26, 2022

lgtm-com bot commented Feb 26, 2022

jordan-schneider commented Feb 26, 2022

lgtm-com bot commented Feb 26, 2022

lgtm-com bot commented Feb 26, 2022

Jasha10 commented Mar 10, 2022

tmke8 left a comment

Choose a reason for hiding this comment

tmke8 Mar 13, 2022

Choose a reason for hiding this comment

jordan-schneider Mar 13, 2022

Choose a reason for hiding this comment

tmke8 Mar 13, 2022

Choose a reason for hiding this comment

jordan-schneider Mar 14, 2022

Choose a reason for hiding this comment

tmke8 Mar 13, 2022

Choose a reason for hiding this comment

jordan-schneider Mar 13, 2022

Choose a reason for hiding this comment

tmke8 Mar 13, 2022

Choose a reason for hiding this comment

jordan-schneider commented Mar 13, 2022 • edited

lgtm-com bot commented Mar 14, 2022

jordan-schneider commented Mar 14, 2022

Jasha10 commented Mar 14, 2022

Jasha10 commented Mar 24, 2022

lgtm-com bot commented Mar 26, 2022

lgtm-com bot commented Mar 26, 2022

jordan-schneider commented Mar 26, 2022

codekansas commented Sep 7, 2022

Jasha10 commented Sep 22, 2022

indigoviolet commented Aug 14, 2023

Queuecumber commented Aug 31, 2023

daturkel commented Sep 27, 2023

spfrommer commented Mar 5, 2024

gsq7474741 commented Mar 5, 2024

jbohnslav commented Mar 6, 2024

SheaML commented Apr 4, 2024

jkulhanek commented Apr 30, 2024

dkmiller commented May 24, 2024

jordan-schneider commented Mar 13, 2022 •

edited