Introducing Keras Recommenders: state-of-the-art recommendation techniques at your fingertips

Key Concepts

• Keras: A high-level deep learning API known for its user-friendliness and intuitive model construction.
• JAX: A library for high-performance numerical computation and machine learning research, featuring JIT (Just-In-Time) compilation.
• Flax NNX: A modular neural network library built on JAX that provides explicit state management and uses standard Python classes for modules.
• Keras-NNX Integration: A bridge allowing Keras variables to be treated as NNX variables, enabling seamless interoperability between the two ecosystems.
• Trace State: A mechanism in NNX that tracks variables to ensure they are compatible with JAX’s JIT compilation.

1. Integration Overview and Setup

The integration of Keras with Flax NNX allows developers to combine the high-level API of Keras with the explicit state management and performance of JAX/NNX. To enable this functionality, users must configure specific environment variables before importing Keras:

• KERAS_BACKEND: Must be set to "jax".
• KERAS_NNX_ENABLED: Must be set to "true".

By enabling this, Keras variables are automatically treated as instances of nnx.Variable, allowing them to be tracked and managed by the NNX ecosystem.

2. Technical Implementation and Verification

The integration allows Keras components to exist within NNX modules. The video demonstrates a model containing both a standard NNX linear layer and a custom Keras variable. To verify the integration, three checks are performed:

1. Trace State Verification: Confirming the Keras variable has a trace_state attribute, which ensures it can be JIT-compiled by JAX.
2. Variable Tracking: Using nnx.variables() to ensure the NNX system recognizes and tracks the custom Keras variable.
3. Direct Access: Confirming that the NNX system can fetch and manipulate the value of the Keras variable directly.

3. Training Workflows

The integration supports two distinct training paradigms:

A. Keras-Managed Workflow (Keras-centric)

• Mechanism: Uses standard Keras APIs like model.compile() and model.fit().
• Benefit: Provides the familiar, high-level developer experience of Keras while running on the high-performance JAX/NNX backend.

B. NNX-Managed Workflow (NNX-centric)

• Mechanism: Treats Keras models as NNX modules. Developers write custom training loops using JAX-native libraries (e.g., optax for optimizers).
• Key Tool: The @nnx.jit decorator is used instead of standard JIT, which optimizes the entire model—including Keras components—for high-performance execution.
• Benefit: Offers maximum flexibility and fine-grained control over gradients, loss computation, and weight updates.

4. Key Arguments and Perspectives

• Unified Framework: The presenter argues that this integration bridges the gap between rapid prototyping (Keras) and high-performance research (JAX/NNX).
• Seamless Interoperability: By making Keras variables compatible with NNX, the system allows developers to "mix and match" components without sacrificing the state management capabilities of either library.
• Performance: The use of nnx.jit ensures that Keras models benefit from JAX’s compilation speed, effectively removing the performance trade-off typically associated with high-level APIs.

5. Conclusion

The Keras-NNX integration represents a significant advancement in the JAX ecosystem. It provides a "best of both worlds" scenario: developers can utilize the intuitive, high-level Keras API for model construction while leveraging the explicit, high-performance control of Flax NNX for training loops and state management. This unified approach is designed to support both simple prototyping and complex, custom research requirements.