TensorFlow

Google's deep learning framework. Still widely deployed in production, especially via TF Serving, TFLite, and TF.js.

At a glance

Field	Value
Category	Deep learning framework
Difficulty	Intermediate
When to use	Legacy TF code, mobile/edge deployment, TF Serving
When not to use	Greenfield research in 2026
Alternatives	PyTorch, JAX, ONNX Runtime

What it is

TensorFlow 2 with Keras is a high-level framework for building and training neural networks. tf.function compiles Python into a static graph that runs on CPU, GPU, or TPU. The big wins over PyTorch are deployment — TFLite for mobile, TF.js for browser, and TF Serving for high-throughput inference.

When we reach for it at Ephizen

On-device inference where TFLite’s quantization and hardware delegates matter.
Browser-side models via TF.js for interactive demos.
Interop with an existing TF pipeline we’re not rewriting from scratch.

Getting started

import tensorflow as tf
from tensorflow.keras import layers

model = tf.keras.Sequential([
    layers.Dense(64, activation="relu", input_shape=(10,)),
    layers.Dense(1),
])
model.compile(optimizer="adam", loss="mse")
model.fit(X, y, epochs=10, validation_split=0.2)

Gotchas

The TF 1.x → TF 2.x migration scarred a lot of codebases; expect dusty corners in older projects.
Errors inside tf.function can be hard to trace because the graph swallows Python stack frames.
TPU training still requires TF or JAX — PyTorch/XLA is improving but not seamless.
Keras 3 now runs on JAX, PyTorch, or TF backends — confirm which one your team is actually using.