Apple Silicon M1: conda-forgeのPyTorch (osx-arm64)は遅いぞ。ビルドしようぜ！ #PyTorch

Update: 2021-12-13 macOS monterey, pytorch0.10.0, torchvision0.11.1

Very slowly

なんでこんなに差がでるんだろ？？

https://github.com/pytorch/examples/tree/master/mnist_hogwild

time python main.py --num-processes 2 --epoch 2
（あまりにも遅いのでepoch 2で計測）

	user	system	cpu	total
build	42.40s	7.50s	258%	19.270
conda-forge	53396.66s	1993.15s	929%	1:39:20.09

これは使い物にならないレベル。自分でビルドしようぜ！！

Environment

macOS Monterey 12.0.1 (Apple Silicon M1 Pro MacBook Pro)
Miniforge3

conda install numpy matplotlib cython pandas scipy scikit-learn

PyTorch v0.10.0

conda install typing-extensions pyyaml

git clone https://github.com/pytorch/pytorch
cd pytorch
git checkout v1.10.0
git submodule update --init --recursive --depth 1

python setup.py bdist_wheel
pip install bdist/torch-1.10.0a0+git36449ea-cp39-cp39-macosx_12_0_arm64.whl

torchvison v0.11.1

git clone https://github.com/pytorch/vision.git
cd torchvion
git checkout v0.11.1

python setup.py bdist_wheel
pip install dist/torchvision-0.11.0a0+fa347eb-cp39-cp39-macosx_11_0_arm64.whl

Note

ということでIntelと比較。M1速いぜ！

time python main.py --num-processes 4 --epoch 20

	user	system	cpu	total
MacBook Air 2020 / Apple Silicon M1	465.06s	79.23s	494%	1:50.05
Mac mini 2019 / Intel Core i5	1071.34s	184.12s	560%	3:44.01