@narayanan2004, @lcnature, are you in touch with @mshvartsman about this PR?

@mihaic Sorry, not yet. I will have a look at it after Friday. @mshvartsman I just had a quick glance a the console output of the automatic testing system, it seems that there are quite a few formatting complaint in brainiak/utils/utils.py and a few in your brainiak/matnormal module. You can probably see them if you just run the run-checks.sh

Hi @mshvartsman
Here are some minor comments regarding MN-RSA.ipynb:
Somehow on github (https://github.com/mshvartsman/brainiak/blob/3dcf915ad942b843b3abf6787cf473f8a094d4a6/examples/matnormal/MN-RSA.ipynb), the font of the first line of equation appear extremely small.
The equation after "Now we add a matrix-normal prior on $\beta$, allowing us to marginalize. We parameterize the covariance in terms of its cholesky factor $\L$." seems to be two distribution, one for beta, one for Y conditioning on other variables. So it seems good to separate them to two lines. The same applies for the next two lines of equations.
Is it possible to write transpose as either ' or ^T?

Will continue to read other codes :)

Github rendering of math in notebooks seems pretty bad unfortunately -- it looks fine locally.

Sorry for so many vacuous commits as I get this ship-shape: for some reason run-checks.sh gives different results on my machine vs on the CI box.

Current linux failures look to be due to tensorflow being incompatible with old libc (e.g. tensorflow/tensorflow#177), I'm guessing because the test server runs Centos/Springdale 6. There are various hackarounds possible (easiest might be installing TF from conda) but I don't know how to make any of them work with the CI setup. Any advice?

@mshvartsman, you are right. TensorFlow does appear to require glibc >= 2.17 (tensorflow/tensorflow#53), although this requirement is not explicitly documented, as far as I can tell.

We do want to be able to depend on TensorFlow, so I am looking into a workaround for our Jenkins server.

I added PR #357 as a workaround for the TensorFlow issue.

Great -- so what are next steps? Review / merge #357 and then I need to modify the build script somewhere also to pass --conda-override into the build script? Or does the latter happen on the Travis CI side?

@lcnature As mentioned on slack, the garbled math rendering is an issue with github's jupyter notebook renderer -- it looks fine locally. I don't know of a workaround -- @mihaic do you?

After PR #357 is merged, I will modify the Jenkins configuration to pass the flag. The Jenkins configuration is not part of this repo. Travis is not affected.

As for the GitHub Latex rendering, I am not aware of a workaround, but I don't think it matters. Users can see the rendering locally, as you mentioned; additionally, we should use our HTML docs for displaying examples (see issue #268).

@mshvartsman, please add the commit from PR #357 to your branch, so we can use the Jenkins workaround.

I didn't mean to be obscure. Merging master will add the commit.

@mshvartsman @mihaic Other than the last comment above, I think the PR is good to go!

The conda build fix doesn't look too onerous, we might as well do it. I think I got all the spots where pymanopt is special-cased and did the same for TFP? The one place I'm not sure about is jenkins-server-customization: does that still need to be there, or does that go away with the hackaround I've got in SSSRM?

Latest TF in conda is 2.2 though and TFP requires 2.3 so we can either do the pypi workaround or just wait until TF2.3 makes it into the conda repo, up to you.

@mihaic @mshvartsman Looks good to me too! I am happy to approve! Thank you both for contributing and managing!

@lcnature, thanks a lot for helping with this monumental PR!

@mshvartsman, since you are up fixing the Conda build, could you please try TensorFlow from PyPI as well? Please ignore jenkins-server-customization; we decided to stop testing on Jenkins recently.

Tried it, not sure what happened with the build there. Looks like TFP successfully installed without complaining about missing TF2.3 but then it couldn't find all the TF libs and froze. Any ideas?

The good news is that there is a TensorFlow 2.3 package in Conda (tensorflow-base). The bad news is that it is not compiled for macOS, only Linux.

Does that change our plans for merging?

My comment was a sort of explanation for the failure of commit 5e472a1. Still planning to merge for the upcoming release and still looking for the best way to build the Conda package.

@mshvartsman, I have a new proposal in light of the difficulty in packaging TensorFlow 2.3 for macOS with Conda. What if I make a new brainiak-matnormal package? It would be a pure Python package, so we will be able to publish a binary wheel for it on PyPI. That would be easily installable using either Pip or Conda, with no prerequirement from the operating system.

I also thought about using Setuptools extras_require to make matnormal an optional part of the BrainIAK package, but that does not work in Conda as far as I can tell.

@mshvartsman, another solution I see is to use extras_require and do nothing in Conda. Instead, we would instruct Conda users to install the TensorFlow 2.3 separately. The instructions would be:

No TensorFlow, matnormal cannot be used:

pip install brainiak
or
conda install brainiak

TensorFlow installed, matnormal can be used:

pip install brainiak[matnormal]
or
conda install brainiak && pip install -r requirements-matnormal.txt

What do you think?

@mshvartsman, to make some progress, I will implement the extras_require proposal. We can change it later.

@mshvartsman, thank you very much for your patience! I hope the optional dependencies approach will work well.

Thank you for your help, @narayanan2004, @lcnature, and @szorowi1.