Chapter 3: TensorFlow vs Keras¶

“Abstraction isn’t just convenience—it’s control disguised as simplicity.”

3.1 The Relationship: Is Keras Part of TensorFlow?¶

Short answer: Yes. But not always. Keras started as an independent deep learning API in 2015, aiming to make neural networks user-friendly. Originally, it supported multiple backends like TensorFlow, Theano, and CNTK.

But in 2017, Google integrated Keras as TensorFlow’s official high-level API, now accessed as:

import tensorflow as tf
from tensorflow import keras

You’ll often see it written as tf.keras—that’s the TensorFlow-native version. And it’s not just a wrapper anymore—it’s part of TensorFlow’s core ecosystem.

3.2 Key Differences and Integration Points¶

Let’s compare them side by side:

Feature	tf.keras	Standalone Keras (pip install keras)
Backend Engine	TensorFlow only	Can run on TensorFlow, Theano, etc.
Performance Optimizations	Supports @tf.function, XLA, mixed-precision	Limited (no deep TF graph integration)
Distributed Training	Integrated with tf.distribute.Strategy	Manual or not available
Mobile Deployment	TFLite-compatible	Not directly
Versioning	Follows TensorFlow versioning	Versioned independently
Recommended?	✅ Yes, always use tf.keras today	❌ Deprecated for most TF users

3.3 Why TensorFlow Adopted Keras¶

Because building models directly with low-level TensorFlow APIs was... painful.

# Old-school TF 1.x style (yikes)
W = tf.Variable(tf.random.normal((784, 64)))
b = tf.Variable(tf.zeros(64))
y = tf.nn.relu(tf.matmul(x, W) + b)

Now with tf.keras:

from tensorflow.keras import layers

model = tf.keras.Sequential([
    layers.Dense(64, activation='relu', input_shape=(784,))
])

Cleaner. Safer. Reusable. And it connects directly to:

Loss functions (tf.keras.losses)
Optimizers (tf.keras.optimizers)
Metrics (tf.keras.metrics)
Model training (model.fit, model.evaluate, etc.)

3.4 The Layer Cake of TensorFlow¶

     [ tf.keras (user API) ]  
                ↓  
  [ Computation Graph (tf.function) ]  
                ↓  
    [ Core TensorFlow Engine ]  
                ↓  
  [ Hardware Execution (CPU/GPU/TPU) ]

The beauty of tf.keras is this: you stay at the top layer, and TensorFlow handles the heavy machinery underneath.

You can still drop down to lower levels when needed—but tf.keras is designed for 90% of use cases.

3.5 When to Use tf.keras vs Raw TensorFlow¶

Situation	Use tf.keras?	Use raw tf?
Building standard neural nets	✅ Yes	❌ Too verbose
Custom ops / experimental graph behavior	❌ Maybe	✅ Yes
Writing production models	✅ Yes	✅ Sometimes
Debugging gradients manually	✅ + GradientTape	✅

Use tf.keras unless you're experimenting with internals or need full control.

“Keras makes TensorFlow human—so you can focus on ideas, not boilerplate.”