Auto-Keras and AutoML

Google introduces Neural Architecture Search, abbreviated as NAS. NAS is essentially an algorithm that searches for the most optimal neural network to perform a specific task for your particular dataset.

For many applications in AI and ML, specialised skills and knowledge were previously required. However, many companies only need deep networks to carry out relatively simple tasks such as image classification.

So, has Google come up with the proverbial silver bullet that allows everyone to wield the powerful force of AI? Not quite!

For instance, using AutoML for computer vision will cost you $20 per hour. And there’s no guarantee that you will achieve better accuracy than a manually designed network.

In light of this, there is an alternative tool called AutoKeras – an open-source package written in Python that is very user-friendly for deep learning in the Keras library. With a straightforward installation (pip install autokeras – I couldn’t find it in Anaconda), it is completely free and has open-source code. If you prefer TensorFlow or Pytorch over Keras, there are links to those alternatives at the end of the article.

So, how can AutoKeras assist me? With hyperparameter tuning such as optimisation algorithms (SGD, Adam, etc.), learning rate, regularisation, and so on.

It’s worth noting that even AutoKeras cannot replace a top-notch professional like Rosebrock, as evidenced by his study on computer vision – medical images with malaria. AutoKeras computed complex models for 24 hours and ultimately achieved a fairly good accuracy of 95.9%. The manually designed variant of ResNet took 54 minutes to compute and achieved an accuracy of 97%. But beware! Adrian Rosebrock is a whiz in computer vision. Despite his young age, he has written three books on the subject, which he continually updates :)

Resources:
Getting started guide: https://www.pyimagesearch.com/…/auto-keras-and-automl-a-ge…/
AutoKeras is better than AutoML: https://towardsdatascience.com/autokeras-the-killer-of-goog…
AutoKeras: https://github.com/jhfjhfj1/autokeras
TensorFlow: https://github.com/melodyguan/enas
PyTorch: https://github.com/carpedm20/ENAS-pytorch
Malaria: https://www.pyimagesearch.com/…/deep-learning-and-medical-…/