mac m1 tensorflow dependency

[hmckay]

Currently using a mac M1, I have tensorflow-deps, tensorflow-macos(2.9.0) and tensorflow-metal installed (all working fine), but when I try to "pip install adapt" I have the following conflict:
adapt 0.4.1 depends on tensorflow>=2.0

Checking my current tensorflow version:
python3 -c "import tensorflow as tf; print(tf.version)"
2.9.2

Is there a way to resolve this/ can you change the requirements so this package will work with tensorflow-macos?

[antoinedemathelin]

Hi @hmckay,
adapt is tested on macos by github at every commit whithout issue, but I think it install tensorflow instead of tensorflow-macos. Does tensorflow-macos work the same as tensorflow?
Have you tried pip install git+https://github.com/adapt-python/adapt.git ? Or to manually install adapt by cloning the repository, removing the requirement and run pip install . in the adapt repo?

[hmckay]

I think you can install tensorflow on a mac, but miss out on a lot of performance gains (I'm not even sure tensorflow hooks into the GPU on the m1s properly). I'm not an expert on using tensorflow, but from what I can tell tensorflow and tensorflow-macos have the same functionality, but tensorflow-macos has been optimised for macs. Once it's been installed you import everything the same way as you would from tensorflow (so as far as I can tell it's just a pip issue)

When I manually try to install using pip install . I get the following error:
ERROR: Could not find a version that satisfies the requirement tensorflow>=2.0 (from adapt) (from versions: none)
ERROR: No matching distribution found for tensorflow>=2.0

I modified the install_requires in setup.py to the following:

install_requires=["numpy>=1.16", "scipy>=1.0", "scikit-learn>=0.2", "cvxopt>=1.2", "tensorflow>=2.0; platform_system !='Darwin' and platform_machine!='arm64'", "tensorflow-macos>=2.0; platform_system =='Darwin' and platform_machine=='arm64'", "tensorflow-metal; platform_system =='Darwin' and platform_machine=='arm64'"],

I was then able to install with pip install .

Edit: I think tensorflow-macos has also been optimised for the intel macs too, so the requirement of platform_machine=='arm64' may not be necessary

[antoinedemathelin]

Hi @hmckay,
Yes it seems to be a good idea to allow tensorflow-macos instead of tensorflow for mac users. However I wonder if we can make a condition like:

if OS == "macos", then requirements = ["tensorflow>=2.0 OR tensorflow-macos>=2.0"]
if OS != "macos", then requirements = ["tensorflow>=2.0"]

Do you think, we can do something like that in setup.py ?

[hmckay]

This seems to work for me (haven't tested it on anything other than a mac with tensorflow-macos installed though):

from setuptools import setup, find_packages
from pathlib import Path
import platform
import pkg_resources

this_directory = Path(__file__).parent
long_description = (this_directory / "README.md").read_text()

INSTALL_REQUIRES = ["numpy>=1.16", "scipy>=1.0", "scikit-learn>=0.2", "cvxopt>=1.2"]

def get_install_requires():
    if platform.system() != 'Darwin':
        INSTALL_REQUIRES.append("tensorflow>=2.0")
    else:
        installed_pkgs = [pkg_name.project_name for pkg_name in pkg_resources.working_set]
        if "tensorflow-macos" in ','.join(installed_pkgs):
            INSTALL_REQUIRES.extend(["tensorflow-macos>=2.0","tensorflow-metal"])
        else:
            INSTALL_REQUIRES.append("tensorflow>=2.0")
    return INSTALL_REQUIRES

setup(
    name='adapt',
    version='0.4.1',
    description='Awesome Domain Adaptation Python Toolbox for Tensorflow and Scikit-learn',
    url='https://github.com/adapt-python/adapt.git',
    author='Antoine de Mathelin',
    author_email='[email protected]',
    license='BSD-2',
    packages=find_packages(exclude=["tests"]),
    install_requires=get_install_requires(),
    zip_safe=False,
    long_description=long_description,
    long_description_content_type='text/markdown'
)

[antoinedemathelin]

Hi @hmckay,
It seems nice! Can you please open a pull-request with your modification of the setup.py? It will launch the tests on other OS, we will see if it works.
Best,