setup.py fix for M1 installation

[nleroy917]

Here's what I have locally if you wanted to take a look. To help you out with any potential debugging I just made a PR so you can see it all inline.

[marvinquiet]

Hi Nathan @nleroy917,

Thank you so much for your contribution.

However, I might need to first test if it still works well for MacOS or Linux. Then I will get back and to merge the pull request.

Thank you so much for your help and efforts! Please do not hesitate to let me know if you encounter any issues with running Cellcano.

Sincerely,
Wenjing

[nleroy917]

No problem! I just wanted to share what I got working for me relative to the entire codebase. If you don't think its necessary to merge, that's ok!

[lujiaying]

Hi Nathan and Wenjing,

Thank you for your efforts to make Cellcano setup more robust.
I propose an alternative way to dynamically handle the setup requirement, by looking into current platform and machine types.

Best,
Jiaying

[nleroy917]

Relevant stack overflow per @lujiaying's suggestion.

However, the issue seems resolved by just removing any pinned dependencies - this fixes the issue while still being agnostic to platform. My personal opinion would be that it should be up to the user with the M1 to specifically install tensorflow-macos if they wish to use metal acceleration, but I'll defer to you all to make the decision there since its your package!

If I can help in any way let me know.

[lujiaying]

Gotcha. I agree "removing any pinned dependencies" is a good solution.

Another thing is we want to make numbers reported in the paper replicable, and sometimes certain packages may conduct aggressive upgrade (tensorflow 1.0 -> 2.0 makes a lot of API incompatible). It also impose challenges in terms of maintainess efforts if we need to keep every dependent third-party packages up-to-date.
I would suggest to keep and fix a pinned version, as it is a common practice for many open-source softwares. (e.g. https://github.com/autogluon/autogluon/blob/master/multimodal/setup.py)

[marvinquiet]

Hi Nathan @nleroy917,

Thank you so much for the wait and the willingness to contribute.

After discussing with Jiaying, we think it is better to keep pinned dependencies to make sure all the results are reproducible in the future. I am also worried that if removing those specific versions, people might not be able to run my software.

I would add a reference to this pull request and issue number in the documentation to let people know your solution. Thank you again for your detailed contribution and I really appreciate it!

Sincerely,
Wenjing