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'

[mayank1101]

It doesn't work

[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?

[ianujv4231]

BRO SAME ERROR, HOW DID U RESOLVE IT ???

FOR THE TRAINING SEGMENT :
trainer = DefaultTrainer(cfg)
trainer.resume_or_load(resume=False) # Start from scratch
trainer.train()

[valentinamisi]

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

[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.

[aguevaracaprio]

@valentinamisi you should be using ARCHFLAGS="-arch arm64” as M chips are based on ARM architecture. If you are on an Intel chip the x86 should be good. Was anyone able to make it work using M chips GPU support?

[peter65374]

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

[attentionmech]

wow, it works in my macOS 15.0 Sequoia. Thanks a tonn.

[LatteCoconut]

I failed to install it in pyenv using python3.10.
but I succeeded in install it with python3.9 in conda.
I don't understand

[gbmerrall]

I got it compiling with the following on an M2

CPATH=/Library/Developer/CommandLineTools/SDKs/MacOSX15.2.sdk/usr/include/c++/v1/ CC=clang CXX=clang++ ARCHFLAGS="-arch arm64" python -m pip install 'git+https://github.com/facebookresearch/detectron2.git'

The specific include directory may be different for you based on the version of CommandLineTools you so check that before trying yourself.

[JoostvDoorn]

It seems you don't need to refer to the exact version, since there is a symlink to the latest version. The following worked for me.

CPATH=/Library/Developer/CommandLineTools/SDKs/MacOSX.sdk/usr/include/c++/v1 CC=clang CXX=clang++ ARCHFLAGS="-arch arm64" python -m pip install 'git+https://github.com/facebookresearch/detectron2.git'

[Benji-create]

i got it to install wtih

SDK_PATH=$(xcrun --show-sdk-path)
CPATH="${SDK_PATH}/usr/include/c++/v1/" CC=clang CXX=clang++ ARCHFLAGS="-arch arm64" python -m pip install 'git+https://github.com/facebookresearch/detectron2.git'

[Athleteg]

This is a workable solution for me:
% git clone https://github.com/facebookresearch/detectron2.git
% cd ~/detectron2
% CC=clang CXX=clang++ ARCHFLAGS="-arch arm64" pip3 install -e . --no-build-isolation