Tensorflow < 2.4 芯片 M1

Question

I am working on a project which requirements need python 3.7 and tensorflow 2.3.1. The problem, I have a MacBook Pro with M1 chip. I was able to install and run TF 2.4. However, I ran into more complicated compatibility issues. Does anyone know how could I solve this?

Answer 1

M1 has a compatibility issue with tensorflow. There is a workaround provided by Apple and other blogs. I have recently tried the same and have provided the summary below:

Tensorflow: OS: BigSur(11.2.3)

Install command line tools:

• xcode-select --install

Install miniforge:

• https://github.com/conda-forge/miniforge#miniforge3

After download run

1. chmod +x ~/<dir>/Miniforge3-MacOSX-arm64.sh
2. sh ~/<dir>/Miniforge3-MacOSX-arm64.sh

dir: directory to which miniforge is downloaded.

Or

1. brew install miniforge
2. After download change source to miniforge3
3. source ~/miniforge3/bin/activate

Download the environment.yml file from https://github.com/mwidjaja1/DSOnMacARM

Setup a new conda environment using the yml file:

• conda env create --file=environment.yml --name <env_name>
• Activate new environment

Install tensorflow dependencies:

• conda install -c apple tensorflow-deps

Install base tensorflow:

• python -m pip install tensorflow-macos

Install tensorflow-metal plugin:

• python -m pip install tensorflow-metal

Possible Issues:

• Miniforge uses conda-forge to install packages
• Method to route channel to conda-forge
• conda install -c conda-forge matplotlib
• conda install or pip install might not work

Packages available for libraries in conda-forge:

https://anaconda.org/conda-forge

Only these will work

Sources:

• https://developer.apple.com/metal/tensorflow-plugin/
• https://towardsdatascience.com/installing-tensorflow-on-the-m1-mac-410bb36b776
• https://github.com/mwidjaja1/DSOnMacARM