smd_hpi

Quantized and binarized operations in mxnet

The structure should be fairly self explanatory.

• src quantized operators that get compiled into the mxnet library
• tools tools developed for working with binarized networks
  • model-converter: a tool to pack a trained binary model so that each weight just uses 1 bit of storage.
  • docker: a simple Dockerfile to setup a container for mxnet with all dependencies, build it with docker build -t mxnet, then run it with docker run -t -i mxnet
  • amalgamate_mxnet_mac.sh: this script will amalgamate the mxnet library into a single file and perform some modifications needed to compile on macOS and iOS
• test unit tests for those operators // @todo
• examples several projects demonstrating the binarized/quantized operators
  • binary_mnist train and predict with a LeNet on the MNIST dataset
  • binary-imagenet1k train and predict with a ResNet18 on the imagenet or cifar10 dataset
• binary_models a collection of pre-trained binarized models over MNIST, CIFAR-10 and ImageNet dataset. The model accuracy has been described in our paper.