In TensorFlow 2.21, what does tf.keras resolve to, and how do you get Keras 2 back?

## The one-sentence version In TensorFlow 2.21, `tf.keras` is **Keras 3** wearing a TensorFlow-shaped namespace, and Keras 2 survives only as a separate, opt-in package called `tf_keras`. ## What actually changed For most of the TF 2.x era, Keras shipped *inside* TensorFlow. `tf.keras` was a Keras 2 implementation whose tensors were `tf.Tensor`, whose graph mode was TensorFlow's, and whose only backend was TensorFlow. A separate `keras` PyPI package also existed, and in practice it was a thin re-export of the same code. Starting with TensorFlow 2.16, that inversion happened: Keras became an independent, **multi-backend** library again (Keras 3), able to run on TensorFlow, JAX or PyTorch, and TensorFlow started *depending on it* rather than containing it. `tf.keras` is now an alias namespace onto that installed Keras 3 package. TensorFlow 2.21 is on the far side of that change. So in a fresh TF 2.21 environment: - `import keras` and `tf.keras` reach the same library. - `tf.keras.__version__` reports 3.x, not 2.x. - Models are `keras.Model`, layers are `keras.layers.Layer`, and their internals are backend-agnostic rather than TensorFlow-specific. ## The legacy escape hatch Keras 2 was not deleted; it was extracted into the PyPI package `tf_keras`. Two steps switch to it: 1. `pip install tf_keras` 2. Set the environment variable `TF_USE_LEGACY_KERAS=1` With that variable set, TensorFlow points `tf.keras` at `tf_keras` instead of Keras 3. Two details matter and both are asked: - **Ordering.** The variable is read while TensorFlow is being imported. Setting it after `import tensorflow as tf` (or from a notebook cell after the kernel already imported TF) does nothing. In code you set it with `os.environ[...]` at the very top of the entry point, before any TF import; in production you usually set it in the container environment instead. - **It is a process-wide switch, not a per-model one.** Everything that goes through `tf.keras` in that process becomes Keras 2. You cannot build a Keras 3 `Model` out of `tf_keras` layers or vice versa — they are separate class hierarchies with separate base classes, and mixing them fails with type errors or silently unregistered weights. ## Why anyone still sets it The usual reasons are downstream: a pinned training script that relies on Keras 2 saving behaviour, a custom layer that reaches into Keras 2 internals, an in-house library that subclasses Keras 2 classes, or another ecosystem package whose released version still expects the Keras 2 object model. `TF_USE_LEGACY_KERAS=1` buys time on all of those without pinning TensorFlow itself to an old release — which is the real appeal, since you may want the newer TF runtime, ops and CUDA support. It is a **pin, not a plan**: `tf_keras` is a maintenance-only package, and new Keras features (backend portability, `keras.ops`, the newer saving format) only exist on the Keras 3 side. ## How to detect which side you are on At runtime, print the version rather than guessing: ``` import tensorflow as tf print(tf.__version__, tf.keras.__version__) ``` A `3.x` Keras version means Keras 3; a `2.x` one means the legacy package is active. This is the first thing to check when an error message mentions a Keras class you thought you knew, or when the same script behaves differently on two machines — a stray `TF_USE_LEGACY_KERAS` in one environment produces exactly that symptom. ## What changes for you when tf.keras is Keras 3 The headline items a TensorFlow user notices: - Symbolic tensors from `keras.Input` are `KerasTensor` objects, not `tf.Tensor`, so raw TensorFlow ops cannot be applied to them while wiring a functional model. - Saving defaults changed: `model.save()` writes the `.keras` archive (legacy `.h5` is still accepted), and producing a TensorFlow SavedModel directory is now `model.export()`. - `save_weights` expects a `.weights.h5` filename. - Layer code that only calls `tf.*` ops still works fine **under the TensorFlow backend**, but is no longer portable to other backends. ## How to answer in an interview Say the seam out loud: "In TF 2.16 and later — so 2.21 — `tf.keras` *is* Keras 3; Keras 2 lives in the `tf_keras` package behind `TF_USE_LEGACY_KERAS=1`, which must be set before importing TensorFlow, and it is a whole-process switch." That single sentence tells the interviewer you have upgraded a real codebase rather than read a tutorial written for TF 2.10.