Keras 3
A major rewrite of the Keras framework, distinguished by its multi-backend support for TensorFlow, PyTorch, and JAX. Through a unified API and backend abstraction layer, the same model code can run interchangeably across different deep learning frameworks. The version was officially released in late 2023.
Timeline (last 90 days)
Keras 3 already supports three major backends: TensorFlow, JAX, and PyTorch.
Keras 3 helps Expedia modernize its ranking technology stack.
Keras 3 is a major rewrite of the Keras framework, with its biggest highlight being multi-backend support, allowing the same model code to switch freely between TensorFlow, PyTorch, and JAX.
Keras 3 addresses the community fragmentation caused by the long-term coexistence of tf.keras and the standalone Keras version by introducing a unified backend abstraction layer.
Switching backends in Keras 3 only requires setting the environment variable KERAS_BACKEND=torch or KERAS_BACKEND=jax, without modifying the model code itself.
2023 年底发布的 Keras 3 进行了架构性重写，回归多后端设计哲学
Keras 3 实现多后端的核心技术机制是统一的张量操作抽象层 keras.ops，提供约 200 个与 NumPy 兼容的操作符
2023 年底发布的 Keras 3 进行了架构性重写，重新回归多后端设计，可在 JAX、PyTorch 与 TensorFlow 之间无缝切换
KerasFormers 是用纯 Keras 3 编写的预训练 Transformer 模型，可同时运行在 JAX、PyTorch 和 TensorFlow 三大后端之上
All Facts (9)
Keras 3 already supports three major backends: TensorFlow, JAX, and PyTorch.
50%UnverifiedKeras 3 addresses the community fragmentation caused by the long-term coexistence of tf.keras and the standalone Keras version by introducing a unified backend abstraction layer.
50%UnverifiedSwitching backends in Keras 3 only requires setting the environment variable KERAS_BACKEND=torch or KERAS_BACKEND=jax, without modifying the model code itself.
50%UnverifiedKeras 3 is a major rewrite of the Keras framework, with its biggest highlight being multi-backend support, allowing the same model code to switch freely between TensorFlow, PyTorch, and JAX.
50%Unverified2023 年底发布的 Keras 3 进行了架构性重写，回归多后端设计哲学
50%UnverifiedKeras 3 实现多后端的核心技术机制是统一的张量操作抽象层 keras.ops，提供约 200 个与 NumPy 兼容的操作符
50%UnverifiedKerasFormers 是用纯 Keras 3 编写的预训练 Transformer 模型，可同时运行在 JAX、PyTorch 和 TensorFlow 三大后端之上
50%Unverified2023 年底发布的 Keras 3 进行了架构性重写，重新回归多后端设计，可在 JAX、PyTorch 与 TensorFlow 之间无缝切换
50%UnverifiedKeras 3 helps Expedia modernize its ranking technology stack.
50%