Keras (bundled v2)

The legacy v2 copy of Keras shipped with TensorFlow. Lives in tensorflow/python/keras/.

Heads-up: Keras 3 is developed in keras-team/keras and is the recommended Keras for new projects. The code in this directory is a frozen v2 snapshot kept around to keep tf.keras.* working for existing users. New features go to keras-team/keras.

Purpose

• Provide the historical tf.keras symbols (Sequential, Model, Layer, Optimizer, Loss, Metric, callbacks.*, applications.*).
• Keep older user code (using tf.keras.Model, model.fit, model.compile) functional through deprecation cycles.
• Serve as a reference v2 implementation tightly coupled to the bundled TF runtime.

Directory layout

tensorflow/python/keras/
├── __init__.py
├── backend.py             # ~206 KB — low-level math used by layers
├── callbacks.py           # ~110 KB — TensorBoard, ModelCheckpoint, EarlyStopping, ...
├── metrics.py             # ~129 KB — every built-in metric
├── losses.py              # ~77 KB — losses
├── models.py              # Model, Sequential, save/load
├── activations.py         # ReLU, GELU, ...
├── constraints.py
├── regularizers.py
├── distribute/            # Distribution support for Model.fit
├── engine/                # Layer + Model base classes (the heart)
├── initializers/
├── layers/                # The big collection: Dense, Conv2D, LSTM, ...
├── legacy_tf_layers/      # Even older `tf.layers` (pre-Keras-as-default)
├── mixed_precision/
├── optimizer_v1.py        # Pre-2.0 optimizer base class
├── optimizer_v2/          # The widely-used Keras V2 optimizers
├── protobuf/              # Keras-specific protos (saved model)
├── saving/                # HDF5, SavedModel, keras_saved_model formats
└── utils/

Key abstractions

Class	File	Role
Layer	tensorflow/python/keras/engine/base_layer.py	Base class for all layers.
Model	tensorflow/python/keras/engine/training.py	Trainable graph; provides fit/evaluate/predict.
Sequential	tensorflow/python/keras/engine/sequential.py	Linear stack of layers.
Optimizer	tensorflow/python/keras/optimizer_v2/optimizer_v2.py	Base optimizer.
Callback	tensorflow/python/keras/callbacks.py	Training-time hook.
Metric	tensorflow/python/keras/metrics.py	Stateful metric.
Loss	tensorflow/python/keras/losses.py	Loss base class.
keras.backend namespace	tensorflow/python/keras/backend.py	Kitchen-sink helpers (deprecated, but still imported widely).

How model.fit works

sequenceDiagram
    participant User
    participant Model
    participant Distribute as DistributionStrategy
    participant Function as tf.function
    participant Runtime as TF runtime

    User->>Model: model.fit(x, y, epochs=...)
    Model->>Distribute: scope, replicate dataset
    Distribute->>Function: train_step traced as tf.function
    loop epochs * steps
        Function->>Runtime: forward pass + backward pass + apply_gradients
        Runtime-->>Function: losses, metrics
    end
    Function-->>Model: history
    Model-->>User: History

Internally model.fit builds a per-step tf.function that runs forward, computes loss, computes gradients (via GradientTape), and applies the optimizer update. Distribution is layered on by entering a tf.distribute.Strategy.scope() before building the model.

Bundled-vs-standalone Keras

import keras (the standalone) reaches Keras 3, which can target multiple backends (TF, JAX, PyTorch). import tensorflow as tf; tf.keras.* resolves into this directory and only targets the bundled TF runtime. Some new APIs land in standalone Keras and are not mirrored here.

Saving and loading

• model.save("path") writes a SavedModel by default. Implementation in tensorflow/python/keras/saving/.
• HDF5 format (.h5) is supported via tensorflow/python/keras/saving/hdf5_format.py.
• keras.models.load_model re-constructs the model from disk and rebuilds the trackable graph.

Integration points

• Builds on: tf.Tensor, tf.Variable, tf.function, tf.GradientTape, tf.distribute.Strategy.
• Saves through: tensorflow/python/saved_model/ (see features/saved-model).
• Logs to: tensorflow/python/summary/ for TensorBoard; the TensorBoard callback is in tensorflow/python/keras/callbacks.py.

Entry points for modification

• Bug fixes that affect bundled tf.keras users should land in this directory.
• New features should land in keras-team/keras (Keras 3) and only get backported here when they fix a critical regression.
• The Keras team treats this directory as nearly read-only — touch with care, and ping the Keras maintainers (comp:keras label).

Related

• features/tf-function-and-autograph — the trace mechanism model.fit uses.
• features/saved-model — what model.save writes.
• features/distribution-strategy — how multi-GPU model.fit works.