Create wheels and conda tar builds about ignite

[elanmart]

@alykhantejani where should conda build scripts live? Perhaps with the rest of pytorch, so they're on the same conda channel?

pytorch scripts seem to be here

EDIT
As for PyPI I believe we'd need to change the name in setup.py to pytorch-ignite, add a simple setup.cfg, and it would be ready to publish, right?

[alykhantejani]

@elanmart for conda it will live in the pytorch conda channel (I've checked this with soumith).

For PyPI, we do indeed need to change the name. pytorch-ignite is great.

[vfdev-5]

So, I started working on this issue and I have some points to discuss. As @elanmart pointed out pytorch and torchvision conda stuff is built in pytorch/builder. So if we aim to integrate ignite in the pytorch/builder: add build_ignite.sh script and meta.yaml to the repository, I have an impression that we may be limited by pytorch releases and can not make our own releases independently of pytorch. Correct me if I'm wrong.

Another point is on release workflow, following build_vision.sh:

rm -rf torchvision-src
git clone https://github.com/pytorch/vision torchvision-src
pushd torchvision-src
git checkout v$VISION_BUILD_VERSION
popd

we need first to make a tag and then create wheels+tars.

So I created these meta.yaml and build_ignite.sh here but something goes wrong with error:

raise CondaValueError("Empty version string.")

if somebody could assist me with that it would be great !

On PyPI everything is more simple :) just run

# Added setup.cfg with universal = 1
python3 setup.py sdist bdist_wheel
twine upload --repository-url https://test.pypi.org/legacy/ dist/*
# or the following command to upload to real PyPI
twine upload dist/*

so you can checkout a preliminary 0.1.0 version at test.pypi

[alykhantejani]

I think we should have an independent release cycle and process. We can update the wheels on pypi and the tar on conda ourselves.

As for the conda error, I'm not quite sure what could cause this. I can try and reproduce on my side

[alykhantejani]

@vfdev-5 Trying to find out what the problem is with the conda build. From the conda docs it seems like the build command should also contain https://conda.io/docs/user-guide/tasks/build-packages/define-metadata.html#build-section

As for your Empty version string error, not 100% sure but this may be related to the fact that the conda build wont be able to install pytorch (as it's only available in the pytorch channel). This works for the vision build as the ANACONDA_USER=pytorch.

According to this stackoverflow post, its not easy to specify this in the meta.yaml but instead we can add this line before the call to build:

conda config --add channels pytorch