TensorFlow

Warnings

• breaking TF 2.16+ ships Keras 3 as default. tf.keras now points to Keras 3, which has breaking API differences from Keras 2. Code written for Keras 2 (tf.keras with TF <2.16) may fail silently or with cryptic errors.
  Fix: To keep Keras 2: pip install tf-keras, then set environment variable TF_USE_LEGACY_KERAS=1 before any tensorflow import. For new code, migrate to Keras 3 API. Key changes: tf.Variable attributes → keras.Variable, TF SavedModel save/load API changed, jit_compile=True by default.
• breaking tf.estimator API fully removed in TF 2.16. Any code using tf.estimator.Estimator, tf.estimator.DNNClassifier, etc. raises AttributeError.
  Fix: Migrate to Keras model API. The Keras training API (model.fit, model.evaluate, model.predict) covers all use cases from tf.estimator.
• breaking Keras 3: model.save() to TF SavedModel format no longer supported. model.save('path') now saves in .keras format by default.
  Fix: Use model.save('model.keras') for Keras format. To export as TF SavedModel: tf.saved_model.save(model, 'saved_model_dir'). To load a SavedModel as a Keras layer: keras.layers.TFSMLayer('saved_model_dir', call_endpoint='serving_default').
• breaking Keras 3: tf.Variable assigned as Layer attributes is NOT tracked as a weight. This silently breaks custom layers that assign tf.Variable in __init__.
  Fix: Use self.add_weight() or assign keras.Variable instead of tf.Variable for tracked layer weights.
• breaking Windows: TF GPU support above 2.10 dropped for Windows Native. tensorflow>=2.11 on Windows only runs on CPU. GPU on Windows requires WSL2.
  Fix: Use tensorflow<2.11 for native Windows GPU, or use WSL2 for GPU support with newer versions.
• gotcha Mixing standalone keras package and tf.keras objects causes isinstance failures. Libraries like tensorflow_hub use tf.keras internally — adding hub.KerasLayer to a standalone keras.Sequential raises ValueError: not an instance of keras.Layer.
  Fix: Use consistent imports: either always use tf.keras (from tensorflow import keras) or always use standalone import keras. Do not mix objects from both in the same model.
• gotcha Keras 3 sets jit_compile=True by default (XLA compilation). Custom layers using TensorFlow-specific ops not supported by XLA will silently fail or error. Was False by default in Keras 2.
  Fix: Pass jit_compile=False to model.compile() if you encounter XLA errors with custom layers or TF ops.

Install

• pip install tensorflow Standard (CPU + GPU via nvidia packages)
• pip install tensorflow[and-cuda] With CUDA dependencies bundled
• pip install tensorflow-cpu CPU-only (smaller, no GPU)
• pip install tf-keras Keras 2 compatibility shim (maintenance mode)

Imports

• keras

  # Option 1: Use standalone Keras 3 (recommended for new code)
  import keras
  model = keras.Sequential([keras.layers.Dense(64, activation='relu')])
  
  # Option 2: Access via tf.keras (same Keras 3 in TF 2.16+)
  import tensorflow as tf
  model = tf.keras.Sequential([tf.keras.layers.Dense(64)])

  Since TF 2.16, tf.keras and import keras both point to Keras 3 — but third-party libraries like tensorflow_hub, tensorflow_probability may still use the old bundled keras, causing isinstance failures.
• tf.function

  @tf.function
  def train_step(x, y):
      with tf.GradientTape() as tape:
          pred = model(x, training=True)
          loss = loss_fn(y, pred)
      grads = tape.gradient(loss, model.trainable_variables)
      optimizer.apply_gradients(zip(grads, model.trainable_variables))
      return loss

  TF 1.x session-based API (sess = tf.Session(), sess.run()) fully removed in TF 2.0. Use eager execution or @tf.function for graph compilation.

Quickstart

Keras 3 model with TensorFlow backend. Use .keras format for saving.

import tensorflow as tf
import keras

# Build model (Keras 3)
model = keras.Sequential([
    keras.layers.Dense(64, activation='relu', input_shape=(10,)),
    keras.layers.Dense(1)
])

model.compile(
    optimizer='adam',
    loss='mse',
    metrics=['mae']
)

# Train
model.fit(X_train, y_train, epochs=10, batch_size=32, validation_split=0.2)

# Save / load (.keras format recommended)
model.save('model.keras')
loaded = keras.models.load_model('model.keras')