Issue installing ffnet with python 3.12

[mathomp4]

I recently tried to install ffnet from this repo and I got:

7366   │ Collecting git+https://github.com/mrkwjc/ffnet
7367   │   Cloning https://github.com/mrkwjc/ffnet to /tmp/pip-req-build-r4u80rit
7368   │   Running command git clone --filter=blob:none --quiet https://github.com/mrkwjc/ffnet /tmp/pip-req-build-r4u80rit
7369   │   Resolved https://github.com/mrkwjc/ffnet to commit 0759200adb0ca6544874fa233da4f12f04dee228
7370   │   Preparing metadata (setup.py): started
7371   │   Preparing metadata (setup.py): finished with status 'done'
7372   │ Building wheels for collected packages: ffnet
7373   │   Building wheel for ffnet (setup.py): started
7374   │   Building wheel for ffnet (setup.py): finished with status 'error'
7375   │   error: subprocess-exited-with-error
7376   │
7377   │   × python setup.py bdist_wheel did not run successfully.
7378   │   │ exit code: 1
7379   │   ╰─> [6 lines of output]
7380   │       Traceback (most recent call last):
7381   │         File "<string>", line 2, in <module>
7382   │         File "<pip-setuptools-caller>", line 34, in <module>
7383   │         File "/tmp/pip-req-build-r4u80rit/setup.py", line 26, in <module>
7384   │           from numpy.distutils.core import setup
7385   │       ModuleNotFoundError: No module named 'numpy.distutils'
7386   │       [end of output]
7387   │
7388   │   note: This error originates from a subprocess, and is likely not a problem with pip.
7389   │   ERROR: Failed building wheel for ffnet
7390   │   Running setup.py clean for ffnet
7391   │ Failed to build ffnet
7392   │ ERROR: Could not build wheels for ffnet, which is required to install pyproject.toml-based projects

If I try this command with python3.11 I see:

$ python3 --version
Python 3.11.6
$ python3 -c 'from numpy.distutils.core import setup'
<string>:1: DeprecationWarning:

  `numpy.distutils` is deprecated since NumPy 1.23.0, as a result
  of the deprecation of `distutils` itself. It will be removed for
  Python >= 3.12. For older Python versions it will remain present.
  It is recommended to use `setuptools < 60.0` for those Python versions.
  For more details, see:
    https://numpy.org/devdocs/reference/distutils_status_migration.html

and with python3.12:

$ python3 --version
Python 3.12.2
mathomp4@discover31 ~
$ python3 -c 'from numpy.distutils.core import setup'
Traceback (most recent call last):
  File "<string>", line 1, in <module>
ModuleNotFoundError: No module named 'numpy.distutils'

As the message says, it looks like numpy.distutils is now gone and they recommend looking at:

https://numpy.org/devdocs/reference/distutils_status_migration.html

and unfortunately it looks like setuptools does not have Fortran support, so this might mean a move to meson. I see many people having to do this on GitHub (see https://github.com/scipy/scipy/issues/13615). If I had a choice, I'd prefer CMake if only because I don't think meson is installed on any of my machines (but maybe with pip this doesn't matter?). Though looking at things like https://github.com/scikit-fmm/scikit-fmm/issues/78#issuecomment-2004970585, maybe meson is "winning"...

[mrkwjc]

Hi! Thanks for reporting! I pushed a solution with custom setuptools commands. Numpy (with f2py), meson and meson-python must be available at installation time. Does this work for you?

[mathomp4]

Hi! Thanks for reporting! I pushed a solution with custom setuptools commands. Numpy (with f2py), meson and meson-python must be available at installation time. Does this work for you?

Ooh. Thanks! I'll let you know tomorrow...once I figure out how to install meson on the cluster. 😄

[mrkwjc]

meson is pip installable...

[mathomp4]

Good news, it installs!

Bad news, I think more changes are needed underneath. When I try to load it:

$ python3 -c 'import ffnet'
Traceback (most recent call last):
  File "<string>", line 1, in <module>
  File "/discover/nobackup/mathomp4/Miniconda-Test/MINIpyD/24.1.2-0_py3.12/2024-03-22/lib/python3.12/site-packages/ffnet/__init__.py", line 17, in <module>
    import ffnet.ffnet as ffnetmodule
  File "/discover/nobackup/mathomp4/Miniconda-Test/MINIpyD/24.1.2-0_py3.12/2024-03-22/lib/python3.12/site-packages/ffnet/ffnet.py", line 16, in <module>
    from scipy import zeros, ones, optimize, sqrt, ndarray, array
ImportError: cannot import name 'zeros' from 'scipy' (/discover/nobackup/mathomp4/Miniconda-Test/MINIpyD/24.1.2-0_py3.12/2024-03-22/lib/python3.12/site-packages/scipy/__init__.py)

I think these lines:

https://github.com/mrkwjc/ffnet/blob/67279ab68469f07c0212228bc1ed41e68f49dcd6/ffnet/ffnet.py#L16-L17

need to become:

from numpy import zeros, ones, sqrt, ndarray, array
from scipy import optimize

If I do that, the README example works:

$ python3 ffnet_example.py
Feed-forward neural network:
inputs:     2
hiddens:    2
outputs:    1
connections and biases:    9

Testing results for 4 testing cases:
OUTPUT 1 (node nr 5):
Targets vs. outputs:
   1      1.000000      1.000000
   2      0.000000      0.500000
   3      0.000000      0.000001
   4      1.000000      0.499999
Regression line parameters:
slope         =  0.499999
intercept     =  0.250000
r-value       =  0.707106
p-value       =  0.292894
slope stderr  =  0.353553
estim. stderr =  0.353553

(or, I guess, doesn't crash? Not sure what the answer should be).

That said, it's possible my suggested changes might need to be protected with try/except? Not sure when the bits moved from scipy to numpy...

[mrkwjc]

Change must be introduced in >scipy-1.10... Nevertheless, I introduced your changes, without try/except, this always should be written so :)

[mathomp4]

@mrkwjc Things seem to be working well for me now! Thanks! I'll close this.