AutoKeras

Choosing a neural-network architecture is the part of deep learning that feels least like engineering and most like folklore: how many layers, which widths, where the dropout goes. AutoKeras hands that question to a search algorithm.

How it works

AutoKeras is an open-source neural architecture search library built on Keras, developed by the DATA Lab at Texas A&M University. You declare a task rather than a model; it then runs a bounded number of trials, each training a candidate architecture, and keeps the best. Search is driven by KerasTuner underneath, and the original system's distinctive idea was network morphism — growing a trained network by inserting or widening layers so the child inherits its parent's weights instead of training from scratch — guided by Bayesian optimisation. Preprocessing sits inside the searched pipeline too: text vectorisation, image normalisation and categorical encoding are chosen for you.

Using it

import autokeras as ak
clf = ak.ImageClassifier(max_trials=20)
clf.fit(x_train, y_train, epochs=10)
model = clf.export_model()

Task-level APIs cover the common cases — ImageClassifier, TextClassifier, StructuredDataClassifier and their regression counterparts — while AutoModel lets you pin the parts you are sure about and search only the rest. What comes out is an ordinary Keras model, so it saves, serves and fine-tunes like anything you would have written by hand. That is the appeal: architecture search on a single GPU, with nothing exotic left in your deployment path.

AutoKeras is an open-source AutoML library that applies neural architecture search to Keras models: you specify a task — image, text or tabular classification or regression — plus a trial budget, and it searches architectures and preprocessing together, returning a standard Keras model. It matters because it makes architecture search practical on a single GPU without hand-designing networks, and its output drops straight into normal TensorFlow training and deployment.

Deep learning models are built in layers, a bit like the floors of a building, and deciding how many to use and how wide each one should be is largely a matter of experience and guesswork.

AutoKeras is a free tool that does that guessing for you. You tell it what kind of job you have — sorting photographs, classifying sentences, predicting a number from a spreadsheet — and roughly how many attempts it may make. It then builds, trains and compares different designs, and gives you back the best one it found, ready to use like a model you had designed yourself. It turns an expert's intuition into something a computer can search for.