Incorrect base class detection

[adamjstewart]

Describe the bug

I have a chain of subclasses, like A -> B -> C -> D. When I document class D, it lists the base class as A instead of C.

How to Reproduce

$ git clone https://github.com/microsoft/torchgeo.git
$ cd torchgeo/docs
$ pip install -r requirements.txt
$ make html
$ # open _build/html/api/datasets.html and see that the base class is torch.utils.data.Dataset instead of RasterDataset

Expected behavior

I would expect the base class to be the direct super class.

Your project

https://github.com/microsoft/torchgeo

Screenshots

No response

OS

macOS 10.15.7, Linux (whatever RtD uses)

Python version

3.8.11

Sphinx version

4.0.1, 4.0.2

Sphinx extensions

sphinx.ext.autodoc

Extra tools

No response

Additional context

I don't think this is the same as #9504 because I can reproduce it with 4.0.1 too.

Could be related to #9395.

For a class like Landsat, even though the super class is RasterDataset, the base class is listed as torch.utils.data.Dataset.

@calebrob6

[tk0miya]

It seems Landsat.__orig_bases__ is incorrect. It should be (RasterDataset, abc.ABC). But it returns (Dataset, abc.ABC) instead. It must be a bug of Python interpreter.

$ python
Python 3.8.12 (default, Sep  3 2021, 02:24:44)
[GCC 10.2.1 20210110] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> from torchgeo.datasets import Landsat
>>> Landsat.__mro__
(<class 'torchgeo.datasets.Landsat'>, <class 'torchgeo.datasets.RasterDataset'>, <class 'torchgeo.datasets.GeoDataset'>, <class 'torch.utils.data.Dataset'>, <class 'typing.Generic'>, <class 'abc.ABC'>, <class 'object'>)
>>> Landsat.__orig_bases__
(torch.utils.data.Dataset[typing.Dict[str, typing.Any]], <class 'abc.ABC'>)

I got the same result with 3.9.7.

$ python
Python 3.9.7 (default, Sep  3 2021, 02:02:37)
[GCC 10.2.1 20210110] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> from torchgeo.datasets import Landsat
>>> Landsat.__mro__
(<class 'torchgeo.datasets.Landsat'>, <class 'torchgeo.datasets.RasterDataset'>, <class 'torchgeo.datasets.GeoDataset'>, <class 'torch.utils.data.Dataset'>, <class 'typing.Generic'>, <class 'abc.ABC'>, <class 'object'>)
>>> Landsat.__orig_bases__
(torch.utils.data.Dataset[typing.Dict[str, typing.Any]], <class 'abc.ABC'>)

Note: This is a Dockefile for reproduce.

FROM python:3.8-slim

RUN apt update; apt install -y build-essential curl git make unzip vim
RUN git clone https://github.com/microsoft/torchgeo.git
WORKDIR /torchgeo/docs
RUN pip install -r requirements.txt
RUN make html

[tk0miya]

I found a workaround for Python's behavior (refs: #9654). It will be merged and fixed in 4.3.0.