Skip to content
This repository contains the code for our recent paper `Minimal Random Code Learning: Getting Bits Back from Compressed Model Parameters'
Python
Branch: master
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Permalink
Type Name Latest commit message Commit time
Failed to load latest commit information.
figures
models
.gitignore
LICENSE
README.md
main.py

README.md

MIRACLE

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.

LeNet-5 on MNIST

LENET5 on MNIST

VGG-16 on CIFAR-10

VGG16 on CIFAR10

You can’t perform that action at this time.