Installation instructions

[bwohlberg]

The installation instructions need to be updated to reflect the new recommended way of installing jaxlib. Updating of the astra-toolbox installation instructions should also be considered since the conda version (2.0.0) is now substantially more recent than the pip version (1.8b5).

Also note https://scico.readthedocs.io/en/latest/contributing.html#scico-dev-contributing does not mention dev_requirements.txt and it probably should.

[smajee]

I would also add cached_property in example_requirements since the PyPi version of xdesign doesn't seem to include this and I always get a missing cached_property error whenever I do a fresh installation.
Has anyone else faced this as well?

[bwohlberg]

It would be best to also report a bug at the xdesign repo. I see that cached-property is listed in their requirements.txt file, so the problem may be due to the absence of any install_requires section in their setup.py.

[smajee]

Good idea. Added an issue (AdvancedPhotonSource/xdesign#77) in the xdesign repo.

[Michael-T-McCann]

Specifically what kind of install has incorrect instructions? From pip or from github? GPU or CPU? Developer or user?

[bwohlberg]

For jaxlib, we currently recommend something like

pip install --upgrade "jaxlib==0.1.70+cuda110" -f https://storage.googleapis.com/jax-releases/jax_releases.html

but it seems that

pip install --upgrade "jax[cuda]" -f https://storage.googleapis.com/jax-releases/jax_releases.html

is now a cleaner way of installing, no? Or does this only upgrade jax, and not the jaxlib dependency?

For astra-toolbox, we currently recommend

conda install -c astra-toolbox/label/dev astra-toolbox

but the version available via pip is now a lot more recent.

[Michael-T-McCann]

The cleaner way of installing jax will (I believe) give you the latest version of jax and jaxlib, not the pinned one needed by SCICO.

For astra, we currently recommend the conda version and conda version is more recent. This is fine, right?

[bwohlberg]

You're right. I had just assumed that we were recommending pip installation since that's the default for everything else.

OK, then I guess nothing to do here other than to add a note on xdesign requirements as pointed out by @smajee.

[smajee]

Wouldn't we would also have to modify the GPU install command to reflect the newer version of jax in scico : pip install --upgrade "jaxlib==0.1.75+cuda110" -f https://storage.googleapis.com/jax-releases/jax_releases.html (0.1.75 from 0.1.70)?

[smajee]

It seems that we do recommend installation of the example_requirements via pip in docs/source/contributing.rst. This includes the astra-toolbox which should be installed via conda.

The instructions in https://scico.readthedocs.io/en/latest/install.html#additional-dependencies seem to be correct.

Should I add a note in docs/source/contributing.rst that astra should be installed via the conda command and the rest via pip using example_requirements.txt?

Another way to handle this would be to 1) split example_requirements.txt into example_requirements_pip.txt and example_requirements_conda.txt or 2) put it in a conda environment.yml file and specify the different channels.

[bwohlberg]

For now, let's just add the note to docs/source/contributing.rst.

When possible, though, we should discuss

• trying to avoid the partial duplication of installation instructions between the main docs and contributing docs, and
• whether it's worth implementing the splitting of the pip requirements files or using a conda environment.yml file.

[smajee]

Added a note in docs/source/contributing.rst for now.

There seem to be more duplications regarding building of docs. We should discuss how to consolidate these and avoid duplications. My inclination is to:

1. Move docs build instructions to the contributing/developer section since it is unlikely that normal users will do this
2. Move the example specific dependency installation to the separate example section and a link in the installation section that points to this

[bwohlberg]

Agreed.

[smajee]

Made the changes in pull request #154

[smajee]

Fixed in pull request #154