tensorflow-metal for Python3.12 and tensorflow 2.17.x

I have figured out a way to use it with Python 3.12 (and MacOS 15). This method relies on the fact that the tensorflow-metal library is not a python library per se, but a whl package containing a MacOS library. As long as there are no API chances since it was compiled (MacOS 12), the library itself should work with any version of Python. With this in mind:

1. Get the whl package of the tensorflow-metal library from Pypi:

https://files.pythonhosted.org/packages/52/56/8373f5751011304a346f07e5423e69f809b626989d2541ae9e816ae7ced2/tensorflow_metal-1.1.0-cp311-cp311-macosx_12_0_arm64.whl

1. Decompress the whl using Keka or 7zip.

2. Identify where your base installation folder for your packages (i.e. where the packages installed with Pip are located). The <base> installation folder for the stock version of Python in MacOS is:

~/Library/Python/3.9/lib/python/site-packages

if you installed Python with MacPorts the <base> location is:

/opt/local/Library/Frameworks/Python.framework/Versions/3.12/lib/python3.12/site-packages/

1. Now, create a new folder inside the site-package folder above, called: tensorflow-metal (you may need sudo rights).

2. From the decompressed folder from the whl package, copy the library into the site-folder/tensorflow-plugins folder:

cp tensorflow_metal-1.1.0-cp311-cp311-macosx_12_0_arm64/tensorflow-plugins/libmetal_plugin.dylib <base>/tensorflow-plugins

1. This assumes you have tensorflow already installed. Run the script below, which I tested with TF 2.17.0 and tf_keras installed as well. You can try to run the script with the plugin installed and without (just remove it from <base>) to see a massive difference in performance, as expected.

import time
import tensorflow as tf
import tf_keras as keras
#from tensorflow import keras

start_time = time.perf_counter()

cifar = keras.datasets.cifar100
(x_train, y_train), (x_test, y_test) = cifar.load_data()
model = keras.applications.ResNet50(
    include_top=True,
    weights=None,
    input_shape=(32, 32, 3),
    classes=100,)

loss_fn = keras.losses.SparseCategoricalCrossentropy(from_logits=False)
model.compile(optimizer="adam", loss=loss_fn, metrics=["accuracy"])
model.fit(x_train, y_train, epochs=5, batch_size=64)

total_time = time.perf_counter() - start_time
print(" Total time: {0:.1f}s or {1:.1f}m or {2:.1f}h".format(total_time,
                            total_time/60, total_time/3600),"\n")

Of course, it would take Apple 10 seconds to create a Python 3.12 wheel. And yet, here we are.