Overview

TensorFlow 2.12, released on March 24, 2023, introduces DTensor for distributed computing and prepares the transition to Keras 3.

Main Features

DTensor

DTensor is a distributed tensor API that automatically shards data and computation across multiple devices (GPU/TPU) transparently.

python 
import tensorflow as tf
from tensorflow.experimental import dtensor

# Create a compute mesh across 2 devices
mesh = dtensor.create_mesh([("batch", 2)])
layout = dtensor.Layout(["batch", dtensor.UNSHARDED], mesh)

# Automatically distributed tensor
x = dtensor.call_with_layout(
    tf.ones, layout, shape=(8, 4)
)
print(f'Shape: {x.shape}')  # (8, 4)

Keras 3 preparation

TensorFlow 2.12 introduces API changes that prepare the migration to Keras 3, with better separation between the backend and the high-level API.

python 
import tensorflow as tf

# Keras API integrated in TensorFlow
model = tf.keras.Sequential([
    tf.keras.layers.Dense(128, activation='relu'),
    tf.keras.layers.Dropout(0.2),
    tf.keras.layers.Dense(10, activation='softmax'),
])

model.compile(
    optimizer='adam',
    loss='sparse_categorical_crossentropy',
    metrics=['accuracy'],
)

Sources