Overview / Usage

The Google Brain team recently released a platform-aware neural architecture search network called MnasNet that helps to classify images quickly and efficiently. The method of searching for an architecture is different as it uses a factorized hierarchial search space. It beats out previous architectures such as MobileNetV2 and ResNet on ImageNet. It reaches a 76.13% Top-1 accuracy with 19× fewer parameters and 10× fewer multiply-add operations.

Methodology / Approach

The implementation can be easily modified, trained on new datasets, and be deployed to mobile devices. I used Keras to construct the convolution blocks, separable convolution blocks, and the inverted residual block. The toy dataset to train on is the CIFAR-10 dataset.

Technologies Used

Keras, Intel® Distribution for Python*

Repository

https://github.com/abhoi/Keras-MnasNet