Introduction to TensorFlow

Discover Google's open-source machine learning framework — its origins, architecture, and why TensorFlow 2 with Keras has become the go-to platform for building and deploying AI models at scale.

What is TensorFlow?

TensorFlow is an open-source machine learning framework developed by the Google Brain team. Originally released in 2015, it has grown into one of the most widely used platforms for building, training, and deploying machine learning models. TensorFlow powers everything from Google Search and Google Photos to YouTube recommendations and Gmail Smart Reply.

TensorFlow supports a wide range of tasks: image recognition, natural language processing, speech recognition, time-series forecasting, reinforcement learning, and more. It runs on CPUs, GPUs, and TPUs (Google's custom AI accelerators), and can deploy to mobile devices, web browsers, edge devices, and cloud servers.

A Brief History

Understanding TensorFlow's evolution helps explain its current design:

TensorFlow 2 vs TensorFlow 1

TensorFlow 2 was a major rewrite that addressed many pain points from TF1. If you're starting today, you'll use TF2 exclusively, but understanding the differences is useful context:

Eager Execution

Eager execution is one of the most important changes in TF2. It means operations execute immediately, returning concrete values instead of building a computational graph for later execution:

import tensorflow as tf

# In TF2, operations run eagerly by default
a = tf.constant([[1, 2], [3, 4]])
b = tf.constant([[5, 6], [7, 8]])

# This immediately computes the result - no session needed!
c = tf.matmul(a, b)
print(c)
# tf.Tensor(
#   [[19 22]
#    [43 50]], shape=(2, 2), dtype=int32)

# You can even convert to NumPy seamlessly
print(c.numpy())
# [[19 22]
#  [43 50]]

The Keras API

Keras is the high-level API for building and training models in TensorFlow. Originally a standalone library created by François Chollet, Keras was integrated into TensorFlow as tf.keras and is now the recommended way to build models:

import tensorflow as tf
from tensorflow import keras

# Build a simple neural network in just a few lines
model = keras.Sequential([
    keras.layers.Dense(128, activation='relu', input_shape=(784,)),
    keras.layers.Dropout(0.2),
    keras.layers.Dense(10, activation='softmax')
])

# Compile with optimizer, loss, and metrics
model.compile(
    optimizer='adam',
    loss='sparse_categorical_crossentropy',
    metrics=['accuracy']
)

# Train on data
model.fit(x_train, y_train, epochs=5, validation_split=0.2)

Keras provides three levels of API complexity to match your needs:

• Sequential API — Simple stack of layers, perfect for most models
• Functional API — Build complex architectures with multiple inputs/outputs and shared layers
• Model subclassing — Full flexibility for custom training loops and research models

The TensorFlow Ecosystem

TensorFlow is more than just a training library. It is an entire ecosystem: