Install TensorFlow on M1 Macs: Mission (It’s)possible
To install TensorFlow on a new Mac M1 is no simple task, unless you have priviledged access to the magic receipe.
Well, I just found a copy on the bright side of the Web that I want to share with you to make the installation a breeze.
0. Requirements
According to TensorFlow installation requirement notes we will need Python 3.6 to 3.9 version so the best way to install it is taking advantage of pyenv if you need other versions on other projects:
- Python 3.9 -> TensorFlow 2.5 or later.
- Python 3.8 -> TensorFlow 2.2 or later.
- pip 19.0 or later (requieres compatibility with
manylinux2010) - macOS 10.12.6 (Sierra) or later versions(64-bit) _(no compatibles with GPU)
- macOS requieres pip 20.3 or later
Installing tensorflow on M1 with pyenv
- https://www.fredlich.com/works/installing-tensorflow-on-m1-with-pyenv
Tensorflow is very tricky to install on an M1 mac. Here’s how I got it working:
Install Python version for TensorFlow
TensorFlow needs Python 3.6 a 3.9 so install it with pyenv
pyenv install 3.8.13
pyenv global 3.8.13 to set it as the main global version
Install Miniforge (for a Conda environment) with pyenv
pyenv install --list to get latest version of miniforge
pyenv install miniforge3-4.10.3-10 to install conda
Activate miniforge env
pyenv activate miniforge3-4.10.3-10
Create a Conda env
conda create --name env_tensorflow python=3.9
Activate env: conda activate env_tensorflow
… you are now in a conda env in a pyenv env
You can also activate
Conda Baseenvironment.
Install tensorflow dependencies
conda install -c apple tensorflow-deps
Install tensorflow-macos version
You’ll have to use pip for this.
pip install tensorflow-macos
ignore any errors.
To use GPU install tensorflow-metal
pip install tensorflow-metal
Try and import tensorflow in the python shell
Enter python shell to check if you can import TensorFlow
python >>> import tensorflow as tf print(tf.__version__)
If it works, it works! … if it doesn’t work, well, it doesn’t work !!!
It didn’t work for me. But I didn’t quit. Continue reading for the official workaround.
Solving the error !!!
In my case I had the following error:
If this call came from a _pb2.py file, your generated code is out of date and must be regenerated with protoc >= 3.19.0. If you cannot immediately regenerate your protos, some other possible workarounds are:
- Downgrade the protobuf package to 3.20.x or lower.
- Set PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION=python (but this will use pure-Python parsing and will be much slower).
More information: https://developers.google.com/protocol-buffers/docs/news/2022-05-06#python-updates
“Python upb requires generated code that has been generated from protoc 3.19.0 or newer.”
So… just do what it says: downgrade protobuf package to the recommended version:
$ pip uninstall protobuf
...
$ pip install protobuf~=3.20.0
#or ... pip install protobuf~=3.19.0 (as the error message says)
Example: Working with Tensorflow in a Conda and Pyenv dual environment
- Enter in mini-forge pyenv environment ```bash pyenv virtualenvs (list all env created in pyenv)
pyenv activate miniforge3-4.10.3-10
(you are now inside the minforge env within pyenv)
2. **Enter in conda environment**
```bash
conda env list (list all environment created within conda)
conda activate base
(more info on how to work with conda environments in conda docs)
- Install all needed packages in conda base env ```bash conda list (list all packages available in conda) pip list pip install package
**Open Jupyter Notebook and work**
```bash
jupyter notebook
...
work with tensorflow
...
Deactivate Conda Env, then Pyenvenv
bash
conda deactivate
pyenv deactivate
`
You can now use TensorFlow in any working directory, since it is installed in a separate Conda environment !!!