Dockerised TensorFlow on an Apple Silicon host (M1 chip)

Introduction

There seems to be a known issue when running TensorFlow inside a Docker image that targets the x86_64 architecture (the one used in Intel 64bit chips):

• Main issue, worth following: tensorflow/tensorflow#52845
• Related, affecting TensorFlow-serving: tensorflow/serving#1948

The root cause seems to be that QEMU, which emulates the x86_64 architecture on behalf of Docker, does not support AVX instructions, which are used in the official releases of TensorFlow in PyPI. More details in the first issue.

One could follow one of the following fixes (the first two mentioned in this comment, the last two mine):

1. Run docker with x86_64 using an unofficial wheel compiled without AVX, which will be slow.
2. Run docker with arm64 (the architecture of M1 chips) using an unofficial wheel for this architecture, which will work fine as long as all other packages and utilities used in your project work with this architecture.
3. Throw the MacBook out of the window, get an old Intel-based MacBook or get a Linux laptop.
4. Wait until someone fixes this mess. Work on something else in the meantime.

Next, we present some simple examples of approaches 1 (tbd) and 2.

Solution 1

TBD

Example with solution 2

See the requirements.txt file, where one of three builds of TensorFlow is installed:

1. the official tensorflow for Linux environments

docker build --platform linux/x86_64 -f Dockerfile -t m1-support-x86_64 .
docker run --platform linux/x86_64 --rm m1-support-x86_64

>> Output when running on a `arm64` host:
>> x86_64
>> The TensorFlow library was compiled to use AVX instructions, but these aren't available on your machine.
>> qemu: uncaught target signal 6 (Aborted) - core dumped
>> Aborted

2. KumaTea's unofficial arm64 tensorflow for Linux environments

docker build --platform linux/arm64 -f Dockerfile -t m1-support-arm64 .
docker run --platform linux/arm64 --rm m1-support-arm64

>> Output in `arm64` host:
>> aarch64
>> tf.Tensor(934.2583, shape=(), dtype=float32)

3. tensorflow-macos for macOS systems (don't forget to use a virtualenv!):

pip install -r requirements.txt
python test.py

>> Output in `x86_64` host:
>> arm64
>> tf.Tensor(-89.86255, shape=(), dtype=float32)