Keras 3 backend selection before import: tf.keras aliasing and legacy API support

Direct Answers

Q1: With KERAS_BACKEND unset and TensorFlow backend active (TF 2.16+), tf.keras aliases Keras 3. Only optimizers that were ported to the Keras 3 multi-backend API remain resolvable—typically Adam, AdamW, SGD, RMSprop, Adadelta, Adagrad, Adamax, Nadam, Ftrl, and Lion. Legacy Keras 2 symbols such as tf.keras.optimizers.legacy.Adam or tf.keras.optimizers.experimental.* are not exposed through the aliasing layer.

Q2: If KERAS_BACKEND=jax is set before any import, the Keras 3 core uses JAX. Importing tf.keras afterward still loads TensorFlow and registers the aliasing shim, but the underlying Keras operations route to JAX. The same optimizer set as Q1 is available via tf.keras.optimizers; no additional Keras 2 legacy symbols appear.

Q3: Formal, version-pinned documentation of the aliasing contract appears in the TensorFlow 2.16 release notes and the Keras 3.0 migration guide. Deterministic behavior for removed legacy symbols is guaranteed from Keras 3.0.x + TensorFlow 2.16.x onward, with the symbol surface stabilizing in Keras 3.1 + TF 2.16.1.

Confirmed Facts

• Backend must be fixed before the first import keras or import tf.keras.
• keras.config.set_backend() is equivalent to the environment variable but must also run before first import.
• Keras 3 core does not contain the legacy or experimental optimizer namespaces from Keras 2.
• Switching backends in-process is unsupported and leads to undefined state.

Likely Explanation

The aliasing shim in TensorFlow redirects attribute access to the active Keras 3 backend. Because the wrapper only exposes the multi-backend API surface, any symbol that was TF-specific in Keras 2 is absent unless TensorFlow separately re-exports it, which it does not do for removed symbols.

Verification Steps

1. unset KERAS_BACKEND && python -c "import tf.keras; print(tf.keras.__version__)"
2. export KERAS_BACKEND=jax && python -c "import keras; print(keras.backend.backend())"
3. Attempt python -c "import tf.keras; tf.keras.optimizers.legacy.Adam()" to confirm AttributeError.

One Diagnostic Detail Needed

Which exact TensorFlow version and Keras 3 version are you targeting? The optimizer symbol surface was finalized in Keras 3.1 / TF 2.16.1; earlier patch releases may differ.