Copyright (c) 2019 ETH Zurich, Lukas Cavigelli, Georg Rutishauser, Luca Benini

EBPC: Extended Bit-Plane Compression for Deep Neural Network Inference and Training Accelerators

If you find this work useful in your research, please cite

@article{cavigelli2019epbc,
  title={{EBPC}: {E}xtended {B}it-{P}lane {C}ompression for {D}eep {N}eural {N}etwork {I}nference and {T}raining {A}ccelerators},
  author={Cavigelli, Lukas and Rutishauser, Georg and Benini, Luca},
  year={2019}
}
@inproceedings{cavigelli2018bitPlaneCompr,
  title={{E}xtended {B}it-{P}lane {C}ompression for {C}onvolutional {N}eural {N}etwork {A}ccelerators},
  author={Cavigelli, Lukas and Benini, Luca},
  booktitle={Proc. IEEE AICAS}, year={2018}
}

The paper is available on arXiv at https://arxiv.org/abs/1908.11645 and https://arxiv.org/abs/1810.03979. The code provided here is not intended as a general framework or library of any kind, but is barely-cleaned research code to help reproduce our experimental results.

The code for the hardware implementation is available here: https://github.com/pulp-platform/stream-ebpc

Installation

Clone repo and setup conda environment

Clone this repo including its submodules:

git clone --recursive [URL to Git repo]

Setup the conda environment with the list of provided dependencies:

conda env create -f=./conda-env.yaml -p ./conda-env

Download and prepare dataset

You can get the ILSVRC 2012 data directly from here, or much faster via torrent. Follow the instructions within QuantLab to prepare the dataset (and use the scripts in QuantLab/ImageNet to prepare a standard ILSVRC12 dataset setup). Point the ilsvrc12 link to the right destination or extract the data directly in there.

Install and train models with QuantLab

You can download QuantLab here. For some analyses, e.g. the sparsity analysis by epoch, train the corresponding network (full precision/no quantization; choose to log a snapshot of the network every epoch). This generates several 10s of GBs of data, hence we prefer not to share them unless necessary.

Structure

Algorithm evaluations

Implementation:

filename	content
dataCollect.py	functions to 1) load the models, 2) obtain the feature maps within a model, 3) read data from tensorboard log files
referenceImpl.py	implementations of the baseline compression algorithms (CSC, zeroRLE, ZVC)
bpcUtils.py	contains the implementation of the compressor incl. all its building blocks as well as the quantization function and the value-to-binary conversion
analysisTools.py	provides the Analyzer class which is instantiated for a specific default quantization method and compressor. It provides functions to get properties such as the compression ratio for a given feature map tensor, its sparsity, ...
groupedBarPlot.py	utility function: creates nice grouped bar plots based on PyPlot
reporting.py	contains functions to read, write, and parse CSV files

Analyses and visualizations:

filename	content
totalComprRatio_generate.py	script to generate/collect all the statistics for the total compression ratio figure; writes results to file; long run-timescript to generate/collect all the statistics for the total compression ratio figure; writes results to file; long run-time
totalComprRatio_plot.ipynb	visualizes all the results on the total compression ratio
sparsityByEpoch.ipynb	generates the figure as suggested by the filename
sparsityBoxplot.ipynb	generates the figure as suggested by the filename
histogram.ipynb	generates the figure as suggested by the filename

Hardware implementation

(TBD)

Licensing

Please refer to the LICENSE file for the licensing of the algorithm evaluation code and to the LICENSE_HARDWARE file for the licensing of the files related to the hardware implementation.