Awesome

pytorch-macos-bench

This is a fast implementation based on mmcls to test the training speed of PyTorch x Apple Silicon GPU.

Training

For cpu training of Resnet-50 on Imagenet.

PYTHONPATH=. python tools/train.py configs/resnet50/resnet50_b256_in1k.py

For gpu training of Resnet-50 on Imagenet.

PYTHONPATH=. python tools/train.py configs/resnet50/resnet50_b256_in1k_mps.py --use-mps

Speed

For cpu training, M1 macbook air 2020 needs about 332 days (57s per batch).

For gpu training, M1 macbook air 2020 needs about 81 days (14s per batch).

Dependencies

mmcls folder is copy and modified from mmcls@877ea30.

mmcv installation:

python -m pip install opencv-python-headless && MMCV_WITH_OPS=1 pip install git+https://github.com/open-mmlab/mmcv.git@25602c68441d18430d649711a4f3bec576ce327e