A set of Python codes and data to benchmark TensorFlow for macOS on a training task of a large CNN model for image segmentation.
I made this set for benchmarking TensorFlow on GPU of M1 SoC in macOS Monterey.
Installation Instructions of TensorFlow for GPU training in macOS Monterey:
Getting Started with tensorflow-metal PluggableDevice (Apple Developer)
Related article of mine in Medium.com:
Training Speed of TensorFlow in macOS Monterey
- Train.py : Main code.
- data : contains 512x512 greyscale images for training (0001.png to 0500.png) and validation (0501.png to 0600.png). Images are simple triangles just for benchmark.
- neural_networks : contains an original TF-Keras model. The model has an U-net-like structure with the total parameters of 20-Mega (20,886,706).
- utils : contains helper codes for Train.py - CSV loader, Keras callbacks, and loss and metrics.
- run_BENCHMARK_5times_5min_interval.sh / .bat : Automation scripts to perform the training task five times with five minutes interval.
- package_list : contains package lists of miniconda (miniforge) in which TensorFlow runs on GPU.
The following packages are required for this benchmark code.
numpy h5py hdf5 joblib opencv graphviz
(The other packages will be installed based on their dependencies.)
In my experience, they should be installed to the miniforge3 BEFORE installing tensorflow-deps, -macos and -metal.
As for versions, please refer to my package list "miniforge3_TF26_macOS_Monterey.txt".
You may be able to install the packages directly into your env by
conda install --name [env_name] --file miniforge3_TF26_macOS_Monterey.txt
But it may be dependent on machine-specific things, which may be UNSAFE.
I performed this benchmark code on my MacBook Air M1 and a WIndows workstation. The red circle in the above figure represents comutation time required for all the training images in an epoch. The yellow circle represents time required for validation.
Please check my Medium article for more details: Training Speed of TensorFlow in macOS Monterey
