keras-team/keras

Keras is a high-level deep learning library that makes it much easier to build and train neural networks without needing to write low-level mathematical code. The problem it solves is that deep learning frameworks like TensorFlow, PyTorch, and JAX each have their own APIs that are powerful but complex, Keras sits on top of all of them, providing a simpler, more consistent interface so developers can focus on their model design rather than framework details. Keras 3 works by letting you write model code once using Keras's intuitive API, then choose which underlying framework (called a backend) actually runs the computations. You can switch between JAX, TensorFlow, PyTorch, or OpenVINO by setting a single environment variable. This is significant because different backends have different performance characteristics, JAX is often fastest for training on hardware accelerators, while TensorFlow has a mature production ecosystem. The library supports building models for computer vision, natural language processing, audio, time series forecasting, and recommendation systems. You can scale from a laptop to a cluster of GPUs or TPUs using the same code. Keras is also designed to be backward-compatible with the older TensorFlow-only version (tf.keras), so existing code can often be migrated with minimal changes. You would use Keras when you want to build and experiment with neural network models quickly, without getting tangled in framework-specific details. The tech stack is Python, requiring Python 3.10 or higher, installed via pip, with your choice of TensorFlow, JAX, PyTorch, or OpenVINO as the backend.