Keras

Keras 3 is a game changer: it is now multi-backend. You can write Keras code and run it on top of JAX, PyTorch, or TensorFlow.

When to Use

• Multi-Backend Deep Learning (Keras 3): Writing models that run interchangeably on JAX, PyTorch, or TensorFlow.
• Rapid Neural Network Prototyping: Developing vision, NLP, and tabular architectures with clean, intuitive APIs.
• Production Model Deployment: Exporting models to ONNX, TensorRT, or TensorFlow Lite.
• Custom Layers & Loss Functions: Subclassing keras.layers.Layer with backend-agnostic keras.ops.

Quick Start

import os
os.environ["KERAS_BACKEND"] = "torch" # Choose: jax, torch, or tensorflow
import keras
from keras import layers

# Build classifier using Functional API
inputs = keras.Input(shape=(28, 28, 1))
x = layers.Conv2D(32, kernel_size=(3, 3), activation="relu")(inputs)
x = layers.MaxPooling2D(pool_size=(2, 2))(x)
x = layers.Flatten()(x)
outputs = layers.Dense(10, activation="softmax")(x)

model = keras.Model(inputs=inputs, outputs=outputs)
model.compile(optimizer="adam", loss="sparse_categorical_crossentropy", metrics=["accuracy"])

Core Concepts

Multi-Backend Functional Model Architecture

Building deep learning architectures compatible with PyTorch, JAX, and TensorFlow:

import os
# Configure backend before importing keras: 'jax', 'torch', or 'tensorflow'
os.environ["KERAS_BACKEND"] = "jax"

import keras
from keras import layers, ops

# Functional API model definition
def create_residual_classifier(input_shape, num_classes):
    inputs = layers.Input(shape=input_shape)

    # Feature extraction block
    x = layers.Dense(128, activation="relu")(inputs)
    x = layers.BatchNormalization()(x)
    x = layers.Dropout(0.2)(x)

    # Residual skip connection
    residual = x
    x = layers.Dense(128, activation="relu")(x)
    x = layers.BatchNormalization()(x)
    x = layers.add([x, residual])

    outputs = layers.Dense(num_classes, activation="softmax")(x)
    return keras.Model(inputs=inputs, outputs=outputs, name="res_classifier")

model = create_residual_classifier(input_shape=(64,), num_classes=10)
model.compile(
    optimizer=keras.optimizers.Adam(learning_rate=1e-3),
    loss="categorical_crossentropy",
    metrics=["accuracy"]
)
model.summary()

Custom Layer with Backend-Agnostic ops

Writing custom layers using universal keras.ops:

import keras
from keras import layers, ops

class ScaleAndShiftLayer(layers.Layer):
    def __init__(self, **kwargs):
        super().__init__(**kwargs)

    def build(self, input_shape):
        self.scale = self.add_weight(
            shape=(input_shape[-1],),
            initializer="ones",
            trainable=True,
            name="scale"
        )
        self.shift = self.add_weight(
            shape=(input_shape[-1],),
            initializer="zeros",
            trainable=True,
            name="shift"
        )

    def call(self, inputs):
        # ops works transparently across JAX, PyTorch, and TensorFlow
        return ops.add(ops.multiply(inputs, self.scale), self.shift)

Robust Training with Modern Callbacks

Automating early stopping, learning rate reduction, and model checkpointing:

callbacks = [
    keras.callbacks.EarlyStopping(
        monitor="val_loss",
        patience=5,
        restore_best_weights=True
    ),
    keras.callbacks.ReduceLROnPlateau(
        monitor="val_loss",
        factor=0.2,
        patience=3,
        min_lr=1e-6
    ),
    keras.callbacks.ModelCheckpoint(
        filepath="best_model.keras",
        monitor="val_accuracy",
        save_best_only=True
    )
]

# model.fit(x_train, y_train, validation_split=0.2, epochs=50, callbacks=callbacks)

Common Patterns

Custom Layer Subclassing with Backend-Agnostic Operations

Problem: Writing custom layers that run seamlessly across JAX, PyTorch, and TensorFlow backends.

Solution: Subclass keras.layers.Layer using keras.ops:

import keras
from keras import ops

class SimpleDense(keras.layers.Layer):
    def __init__(self, units=32):
        super().__init__()
        self.units = units

    def build(self, input_shape):
        self.w = self.add_weight(shape=(input_shape[-1], self.units), initializer="glorot_uniform")
        self.b = self.add_weight(shape=(self.units,), initializer="zeros")

    def call(self, inputs):
        return ops.matmul(inputs, self.w) + self.b

Best Practices

Do:

• Target Keras 3 with multi-backend compatibility (os.environ["KERAS_BACKEND"] = "jax" or "torch").
• Use keras.ops instead of backend-specific tensor libraries (torch.* or tf.*) in custom layers.
• Save models in the native .keras zip-based format (model.save("model.keras")).
• Always include EarlyStopping with restore_best_weights=True to prevent overfitting.

Don't:

• Use legacy keras 2.x patterns tied exclusively to tf.keras.
• Write custom training loops unless necessary; model.compile() and model.fit() provide high optimization.
• Save models using legacy H5 (.h5) format; use modern .keras.

Troubleshooting

| Error | Cause | Solution | | :------------------------------------------------------------- | :---------------------------------------------------------------------------- | :------------------------------------------------------------------------- | | Backend mismatch error in Keras 3 | Changing KERAS_BACKEND after importing Keras. | Set os.environ["KERAS_BACKEND"] before importing keras. | | ValueError: Shapes (None, 1) and (None, 10) are incompatible | Loss function mismatch (e.g. categorical_crossentropy with integer labels). | Use sparse_categorical_crossentropy for integer class labels. | | Layer ... was called on an input with incompatible shape | Input dimension does not match input_shape of the first layer. | Inspect model architecture using model.summary() to verify layer shapes. |

References

• Keras Documentation