Building detectron2 on M1/M2 Macs

[jessecanada]

Hi all,

With the new pytorch support for Apple Silicon, I was eager to try and run my detectron2 projects on my M1 Mac. I successfully used the following recipe to install detectron2.

Important: The grpcio package is required to successfully build detectron2 from source, but it will not install right out of the box on Apple Silicon. To get circumvent this issue, we need to export the following environmental variables. View source.
export GRPC_PYTHON_BUILD_SYSTEM_OPENSSL=1
export GRPC_PYTHON_BUILD_SYSTEM_ZLIB=1

After that, we are good to go.

1. Create a conda virtual environment
2. critical Install pytorch nightly build. However, at the time of writing (6/14/2022), torchaudio is not available yet as a prebuild. Thus, we use this command:
pip install --pre torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/nightly/cpu
3. Run uname -m in terminal to check your archflag (should be "arm64")
4. critical Build detectron2 from source:
```CC=clang CXX=clang++ ARCHFLAGS="-arch arm64” python -m pip install 'git+https://github.com/facebookresearch/detectron2.git'# (add --user if you don't have permission)

[kimile599]

good to know that

[ccerhan]

For step 2, torch now can be installed like this:

conda install pytorch==1.13.1 torchvision==0.14.1 -c pytorch

Before step 4, pycocotools should be installed correctly for arm64:

conda install pycocotools -c conda-forge

Then install detectron2:

CC=clang CXX=clang++ ARCHFLAGS="-arch arm64" python -m pip install 'git+https://github.com/facebookresearch/detectron2.git'

[dhruvsojitra76]

Detectron2 error:

NotImplementedError: The operator 'aten::_foreach_add.List' is not currently implemented for the MPS device.

Is there any way i can utilize my MPS while traning on coustom dataset ?

is MMdetection competable with MPS?

[valentinamisi]

Regarding step 3, I get this: x86_64, does this change step 4? Step 4 is not working for me.

[valentinamisi]

I tried this: CC=clang CXX=clang++ ARCHFLAGS="-arch x86_64” python -m pip install 'git+https://github.com/facebookresearch/detectron2.git'

and I get the following error:

fatal error: too many errors emitted, stopping now [-ferror-limit=]
20 errors generated.
error: command '/usr/bin/clang' failed with exit code 1
[end of output]

note: This error originates from a subprocess, and is likely not a problem with pip.
ERROR: Failed building wheel for detectron2
Running setup.py clean for detectron2
Failed to build detectron2
ERROR: Could not build wheels for detectron2, which is required to install pyproject.toml-based projects

[OBW93]

I get the same error message with arm64 and using Python 3.9.4 with Sonoma 14.3 and a M1. Could you solve this problem?

[EmmaMeeus]

Same here, but with Python 3.10.0 / M1 / Sonoma 14.1.2.

[ezalorsara]

i got the same error

[tuokcalb]

I'm still stuck, did anyone ever get this working? I may just give up at this point.

[peter65374]

it works on my Mac M2 chip with macOS14.2.1 Sonoma. That's cool.