TensorFlow Addons

Warnings

• breaking TensorFlow Addons (TFA) has reached its planned end of life in May 2024. Development of new features has ceased, and the library is in a minimal maintenance mode. Users are advised to migrate to other TensorFlow community repositories (e.g., Keras, Keras-CV, Keras-NLP).
  Fix: Migrate functionality to core TensorFlow, Keras, or other specialized TensorFlow community libraries. Review release notes for specific component migration paths.
• gotcha Compatibility with TensorFlow versions is crucial. TensorFlow Addons is only guaranteed to be compatible with the TensorFlow versions it was tested against. Warnings will be emitted if versions do not match, and custom C++ operations may lead to segmentation faults or crashes with incompatible TensorFlow builds (e.g., `conda` installations).
  Fix: Always install `tensorflow-addons` with the `[tensorflow]` extra (e.g., `pip install tensorflow-addons[tensorflow]`) or verify the compatibility matrix on the official GitHub page. Avoid `conda`-installed TensorFlow for custom ops.
• deprecated Some functionalities, like `tfa.activations.gelu` and `data_format` arguments in `tensorflow_addons/image`, have been deprecated, with `gelu` migrating to core TensorFlow.
  Fix: Consult the `tensorflow-addons` release notes and documentation for specific replacements. For `gelu`, use `tf.keras.activations.gelu` or `tf.nn.gelu`.
• breaking Windows support for `tensorflow-addons` was dropped due to inconsistent TensorFlow 2.15 `.whl` packaging. Users on Windows will fall back to pure TensorFlow Python implementations where possible for custom ops.
  Fix: Use Linux or macOS for full `tensorflow-addons` functionality, especially for GPU custom operations. If on Windows, be aware of potential performance implications or missing custom op functionality.
• gotcha APIs in TensorFlow Addons might evolve more rapidly than those in core TensorFlow. While the SIG strives for stability, frequent updates can introduce changes.
  Fix: Regularly check release notes and changelogs when upgrading `tensorflow-addons` to understand any API modifications.

Install

• pip install tensorflow-addons Stable release
• pip install tensorflow-addons[tensorflow] With compatible TensorFlow version

Imports

• tfa

  import tensorflow_addons as tfa

• tfa.activations

  import tensorflow_addons.activations as tfa_activations

• tfa.layers

  import tensorflow_addons.layers as tfa_layers

• tfa.optimizers

  import tensorflow_addons.optimizers as tfa_optimizers

• tfa.losses

  import tensorflow_addons.losses as tfa_losses

• tfa.metrics

  import tensorflow_addons.metrics as tfa_metrics

Quickstart

This quickstart demonstrates how to integrate a TensorFlow Addons layer (`GroupNormalization`) and an optimizer (`AdamW`) into a standard Keras model. It compiles and runs a single training epoch with dummy data.

import tensorflow as tf
import tensorflow_addons as tfa

# Define a simple Keras model with a TFA layer and optimizer
model = tf.keras.Sequential([
    tf.keras.layers.InputLayer(input_shape=(10,)),
    tf.keras.layers.Dense(128, activation='relu'),
    tfa.layers.GroupNormalization(groups=8, axis=-1), # Example TFA layer
    tf.keras.layers.Dense(10, activation='softmax')
])

# Use a TFA optimizer
optimizer = tfa.optimizers.AdamW(weight_decay=0.001, learning_rate=0.001)

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

# Create dummy data
import numpy as np
x_train = np.random.rand(100, 10).astype(np.float32)
y_train = np.random.randint(0, 10, 100)

print("Model summary:")
model.summary()

print("\nTraining model with TFA components...")
model.fit(x_train, y_train, epochs=1, batch_size=32, verbose=0)
print("Training complete.")