Continuous integration build fails because numpy is missing.

[lorenzocestaro]

Hi @r9y9,

I am building building a project which has pysptk as a dependency and my builds are failing because the setup.py script requires numpy as a dependency.

For my CI, I can use an easy workaround and install numpy previous to running my project's pip install. Anyway this could create problems for production builds (e.g. docker) and usability problems in general.

I think this could be easily fixed using the setup_requires in your setup.py. Search for setup_requires here.

I am willing to look into this and eventually submit a PR if you want.

[r9y9]

Hi, thank you for the suggestion. That sounds good to me. Could you send a PR for it?

[lorenzocestaro]

Sure, I'll get to work

[lorenzocestaro]

Ok, so I looked into this and it appears that setup_requires is incompatible with pip right now and the problem of the missing numpy installation is unavoidable when building from source.
See also this discussion for the exact same issue on scikit-learn.

What is avoidable, though, is having end users being unable to install the package if they do not have numpy in their environment. This can be fixed by publishing a wheel instead of a source bundle (or both) when releasing new versions:

• python setup.py sdist forces end users to build from source and raises an error if numpy is missing.
• python setup.py bdist_wheel will produce a pip wheel that takes care of installing all dependencies under install_requires and wil not produce errors for end users when pip install pysptk is run.

You can build and upload to pypi both of them with python setup.py sdist bdist_wheel.

Hope I've been clear.

[r9y9]

I used to provide wheels for earlier releases (e.g., see https://pypi.org/project/pysptk/0.1.3/#files), but as far as I remember correctly there's a binary compatibility issue(s) for linux and also I didn't want to create wheels for each python/os environments, so I end up with the source distribution. I hope there's a clean and simple solution to this.

[lorenzocestaro]

If that's the case then I am afraid there is no seamless solution. Everyone who needs to install the package will need to have numpy already installed in their environment.

Thanks @r9y9

[r9y9]

I know pyworld (https://github.com/JeremyCCHsu/Python-Wrapper-for-World-Vocoder/blob/823e9bda5f2b5bc4e10e2ff44534aa87073250cb/setup.py#L25-L31) avoids the numpy build-time dependency in a tricky way. I'm not sure if it really works or not, but I will look into it.

EDIT: Sorry for the late reply.

[r9y9]

I've tagged a new release v0.1.15. I believe this should fix the issue. For double-check, @lorenzocestaro Could you please confirm if it works?

[lorenzocestaro]

@r9y9 installation work seamlessly now. Thank you very much!

Unfortunately this new version breaks my tests, I will look into that to see if it is a problem on my side.

[r9y9]

Thanks! If you provide me code to reproduce the test failures, I can take a look. Let me know if you need a help.

[r9y9]

I'm closing this out. Feel free to reopen if the problem persists.