This repository contains the code for our recent paper `Minimal Random Code Learning: Getting Bits Back from Compressed Model Parameters' (https://arxiv.org/abs/1810.00440). This is the implementation of MIRACLE for LeNet-5 on MNIST and VGG-16 on CIFAR-10. It is based on Tensorflow (1.3).
An example model can be trained by running
python main.py
We used Deep Compression, Weightless Encoding and Bayesian Compression as baselines.
The Compression-Error rate exchange is shown below. Lower left is better.