Install TensorFlow on Apple M1 (M1, Pro, Max) with GPU (Metal)

Note: Uninstall Anaconda/Anaconda Navigator and other related previously installed version of conda-based installations. Anaconda and Miniforge cannot co-exist together.

Installing Miniforge

Install miniforge from brew: brew install miniforge
Create an anaconda environment: conda create -n tf
Activate the environment: conda activate tf
Install Python: conda install python

Installing TensorFlow

Run: conda install -c apple tensorflow-deps to install Apple's TensorFlow dependencies
Run: pip install tensorflow-metal to install Apple's Metal GPU APIs for TensorFlow
Execute: pip install tensorflow-macos to install MacOS arm64 version of TensorFlow
Execute: pip install tensorflow-datasets pandas jupyterlab to install relevant dependencies to run sample code.

Verifying Installation

Execute: jupyter-lab to open a Jupyter Notebook and run the following code:

import tensorflow as tf
import tensorflow_datasets as tfds
print("TensorFlow version:", tf.__version__)
print("Num GPUs Available: ", len(tf.config.experimental.list_physical_devices('GPU')))
print("Num CPUs Available: ", len(tf.config.experimental.list_physical_devices('CPU')))

Checking GPU and CPU

GPU

Running A Sample Code (MNIST)

(ds_train, ds_test), ds_info = tfds.load(
    'mnist',
    split=['train', 'test'],
    shuffle_files=True,
    as_supervised=True,
    with_info=True,
)

def normalize_img(image, label):
  """Normalizes images: `uint8` -> `float32`."""
  return tf.cast(image, tf.float32) / 255., label

batch_size = 128
ds_train = ds_train.map(
    normalize_img, num_parallel_calls=tf.data.experimental.AUTOTUNE)
ds_train = ds_train.cache()
ds_train = ds_train.shuffle(ds_info.splits['train'].num_examples)
ds_train = ds_train.batch(batch_size)
ds_train = ds_train.prefetch(tf.data.experimental.AUTOTUNE)
ds_test = ds_test.map(
    normalize_img, num_parallel_calls=tf.data.experimental.AUTOTUNE)

ds_test = ds_test.batch(batch_size)
ds_test = ds_test.cache()
ds_test = ds_test.prefetch(tf.data.experimental.AUTOTUNE)

model = tf.keras.models.Sequential([
  tf.keras.layers.Conv2D(32, kernel_size=(3, 3),
                 activation='relu'),
  tf.keras.layers.Conv2D(64, kernel_size=(3, 3),
                 activation='relu'),
  tf.keras.layers.MaxPooling2D(pool_size=(2, 2)),
  tf.keras.layers.Flatten(),
  tf.keras.layers.Dense(128, activation='relu'),
  tf.keras.layers.Dense(10, activation='softmax')
])

model.compile(
    loss='sparse_categorical_crossentropy',
    optimizer=tf.keras.optimizers.Adam(0.001),
    metrics=['accuracy'],
)

model.fit(
    ds_train,
    epochs=12,
    validation_data=ds_test,
)

MNSIT

Credits

Nachiket Sirsikar