Keras

3.13.2 · active · verified Sun Mar 29

Keras 3 is a multi-backend deep learning framework providing a high-level API for building and training neural networks. It supports JAX, TensorFlow, PyTorch, and OpenVINO (for inference-only) as computational backends, allowing users to leverage the same codebase across different frameworks. Focused on fast experimentation and user experience, Keras 3 enables efficient development and deployment of deep learning models across various domains. The current version is 3.13.2, and the library maintains an active release cadence with frequent updates.

Warnings

breaking Keras 3.13.0 introduced a breaking change by requiring Python 3.11 or higher. Earlier Python versions are not supported.
Fix: Upgrade your Python environment to version 3.11 or newer. `pip install --upgrade python` (if using pyenv/conda, manage environment accordingly).
gotcha The Keras backend (TensorFlow, JAX, or PyTorch) must be configured *before* importing Keras. Attempting to change it after import will not work.
Fix: Set the `KERAS_BACKEND` environment variable (e.g., `os.environ["KERAS_BACKEND"] = "jax"`) or configure `~/.keras/keras.json` before any `import keras` statement in your code.
breaking Model saving in Keras 3.x has changed. The `model.save()` method now expects the native Keras `.keras` format. Saving to the TensorFlow SavedModel format directly via `model.save()` is no longer supported and will raise a ValueError.
Fix: Use `model.save('my_model.keras')` for the native Keras format. For SavedModel/TFLite export, use `model.export(filepath)`.
gotcha When using TensorFlow versions 2.0 through 2.15, `pip install tensorflow` would install Keras 2.x and make it available via `import keras` and `tf.keras`. If you install TensorFlow 2.15, it will overwrite a Keras 3 installation with Keras 2.15.
Fix: For TensorFlow versions <=2.15, if you intend to use Keras 3, you must reinstall Keras 3 *after* installing TensorFlow 2.15. TensorFlow 2.16+ installs Keras 3 by default, but direct `import keras` is still recommended for clarity.
deprecated Setting a `tf.Variable` directly as an attribute of a Keras 3 layer or model will no longer automatically track that variable as a trainable weight, unlike in Keras 2.
Fix: To ensure variables are tracked, use `self.add_weight()` within custom layers/models, or use `keras.Variable` instead of `tf.Variable`.
gotcha Security hardening was introduced to disallow `TFSMLayer` deserialization in `safe_mode`, preventing potential execution of attacker-controlled graphs during model loading from external TensorFlow SavedModels.
Fix: Upgrade to Keras 3.12.1, 3.13.2, or newer to benefit from this security fix. Avoid loading untrusted models.

Install

pip install keras --upgrade Install Keras Core
pip install tensorflow # for TensorFlow backend pip install jax jaxlib # for JAX backend pip install torch torchvision torchaudio # for PyTorch backend Install a Backend (Optional)

Imports

Model
```
from keras import Model
```
Keras 3.x is imported directly as `keras`. While `tf.keras` can point to Keras 3 in TensorFlow 2.16+, direct `import keras` is recommended for explicit Keras 3 usage and backend-agnostic development.
layers
```
from keras import layers
```
Similar to `Model`, layers and other Keras components should be imported directly from the `keras` namespace for Keras 3.x.

Quickstart

This quickstart demonstrates building, compiling, and training a simple convolutional neural network using Keras 3.x for image classification on the MNIST dataset. It includes setting the backend via an environment variable before importing Keras, which is a critical step for Keras 3.x.

import os
os.environ["KERAS_BACKEND"] = os.environ.get("KERAS_BACKEND", "tensorflow") # Set backend before importing keras
import keras
from keras import layers
import numpy as np

# Load example data (e.g., MNIST for a simple classification task)
(x_train, y_train), (x_test, y_test) = keras.datasets.mnist.load_data()
x_train = x_train.reshape(-1, 28, 28, 1).astype("float32") / 255.0
x_test = x_test.reshape(-1, 28, 28, 1).astype("float32") / 255.0

# Define a simple Sequential model
model = keras.Sequential([
    keras.Input(shape=(28, 28, 1)),
    layers.Conv2D(32, kernel_size=(3, 3), activation="relu"),
    layers.MaxPooling2D(pool_size=(2, 2)),
    layers.Conv2D(64, kernel_size=(3, 3), activation="relu"),
    layers.MaxPooling2D(pool_size=(2, 2)),
    layers.Flatten(),
    layers.Dropout(0.5),
    layers.Dense(10, activation="softmax"),
])

# Compile the model
model.compile(
    loss=keras.losses.SparseCategoricalCrossentropy(),
    optimizer=keras.optimizers.Adam(learning_rate=1e-3),
    metrics=["accuracy"],
)

# Train the model
print("\nTraining model...")
model.fit(x_train, y_train, batch_size=128, epochs=3, validation_split=0.1)

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

view raw JSON →