TF-Keras (Legacy Keras 2)

Warnings

• breaking TF-Keras (Keras 2) is in maintenance mode. New features are developed in Keras 3. If you install TensorFlow >= 2.16, `tf.keras` will default to Keras 3, which has a different API and multi-backend support. This can cause compatibility issues with code expecting Keras 2 behavior.
  Fix: For new projects, consider migrating to Keras 3 (`pip install keras`). To explicitly use TF-Keras (Keras 2) with TensorFlow >= 2.16, install `tf-keras` and set the environment variable `TF_USE_LEGACY_KERAS=1` before importing TensorFlow or Keras. Alternatively, import directly from `tf_keras` as shown in the imports section.
• breaking Python 3.9 support has been removed as of `tf-keras` version 2.21.0.
  Fix: Upgrade your Python environment to Python 3.10 or newer.
• deprecated The Keras Scikit-learn API wrappers (`KerasClassifier` and `KerasRegressor`) were removed starting with TensorFlow 2.13 and compatible `tf-keras` versions.
  Fix: Migrate to `SciKeras` for Scikit-learn compatible Keras models.
• gotcha The default model saving format (`.keras` extension) in Keras 2 is now the Keras v3 format, not the H5 format. This might break workflows that manually inspected or modified H5 files saved with a `.keras` extension.
  Fix: When saving models, explicitly specify `save_format="h5"` if you need to retain the H5 format for a `.keras` file: `model.save('my_model.keras', save_format='h5')`. For general use, consider adopting the Keras v3 saving format or `model.export()` for inference.

Install

• pip install tf-keras Install TF-Keras

Imports

• keras

  import tf_keras as keras

  Direct `import keras` (without installing tf_keras and setting TF_USE_LEGACY_KERAS=1) will import Keras 3, which is multi-backend and has potential API differences. `tf_keras` is the dedicated legacy Keras 2 package.
• Model, layers

  from tf_keras import models, layers

  While `from tensorflow.keras` used to be the standard for Keras 2, with TensorFlow 2.16+ it now defaults to Keras 3. For explicit Keras 2 usage, import directly from the `tf_keras` package.

Quickstart

This quickstart demonstrates how to build, compile, train, and evaluate a simple neural network for classifying MNIST handwritten digits using the `tf_keras` library. It showcases the Sequential API, common layers like Dense and Dropout, and standard training procedures. It also includes a crucial environment variable setting for compatibility with newer TensorFlow versions.

import os
import numpy as np
import tf_keras as keras
from tf_keras import layers

# Set environment variable to ensure Keras 2 is used if TensorFlow >= 2.16 is also installed
os.environ['TF_USE_LEGACY_KERAS'] = '1'

# Load a dataset (MNIST handwritten digits)
(x_train, y_train), (x_test, y_test) = keras.datasets.mnist.load_data()

# Preprocess the data
x_train = x_train.reshape(-1, 28 * 28).astype('float32') / 255.0
x_test = x_test.reshape(-1, 28 * 28).astype('float32') / 255.0

# Define the model using the Sequential API
model = keras.Sequential([
    layers.Dense(128, activation='relu', input_shape=(784,)),
    layers.Dropout(0.2),
    layers.Dense(10, activation='softmax')
])

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

# Train the model
print("\nTraining the model...")
history = model.fit(x_train, y_train, epochs=5, batch_size=32, validation_split=0.1)

# Evaluate the model
print("\nEvaluating the model...")
loss, accuracy = model.evaluate(x_test, y_test)
print(f"Test Loss: {loss:.4f}, Test Accuracy: {accuracy:.4f}")

# Make predictions
predictions = model.predict(x_test[:5])
predicted_classes = np.argmax(predictions, axis=1)
print(f"\nFirst 5 test samples predictions: {predicted_classes}")
print(f"True labels: {y_test[:5]}")