About

This repository contains the code used in the research paper "MARLIN: a Co-design Methodology for Approximate ReconfigurabLe Inference of Neural Networks at the Edge", published in "IEEE Transactions on Circuits and Systems I: Regular Papers". The code can be used as it is to reproduce the results presented in the paper.

MARLIN is a methodology that enables the implementation of approximate neural networks (NNs) on a RISC-V core. It is built from 4 components: - A runtime reconfigurable multiplier with 256 approximation levels, each one with its own error distribution and average power. - A modified RISC-V core, with the multiplier inserted in the execution stage. - A modified PULP toolchain, with approximation control automatically embedded in the compiled code. - A search algorithm based on NSGA-II that evaluates thousands of approxximate NNs and selects those with the best energy/accuracy trade-off.

Install instructions

Execute the following commands to install Clone the repository with bash git clone https://github.com/vlsi-lab/MARLIN.git or bash git clone git@github.com:vlsi-lab/MARLIN.git Then open a terminal and intall ninja-build bash sudo apt update sudo apt install ninja-build Install marlin environment (requires miniconda or anaconda). ```bash cd MARLIN

conda env create -f marlin.yml ``` Inside each folder there are README files with additional information useful to install additional components, execute the code and replicate the data.

In order to use the multiplier LUTs and pre-trained neural networks, it is necessary to install 7-zip and extract the files according to the README contained in each folder. You can download 7-zip from here. When launching any python file, remember to keep the folder hierarchy as defined in this repository. You can add the folders to the PYTHONPATH, after entering the MARLIN repo, with the following command: bash export PYTHONPATH="${PYTHONPATH}:./:./neural_networks:./benchmark_CIFAR10:./approximate_multiplier:./riscv_characterization"

Run simulation

Support to verify the experimental results obtained on the modified version of PULP Platform is provided. vendor is used to keep track of dependencies across multiple github repositories. Execute the command: bash make build This command will setup the required scripts and repository. Further information on the configuration can be found in the folder ./riscv_characterization.\ Note: to avoid possible path errors, we highly suggest to launch python scripts from inside MARLIN folder directly.

Citation

If you used this code in your work, consider adding a reference to our paper @ARTICLE{marlin, author={Guella, Flavia and Valpreda, Emanuele and Caon, Michele and Masera, Guido and Martina, Maurizio}, journal={IEEE Transactions on Circuits and Systems I: Regular Papers}, title={MARLIN: A Co-Design Methodology for Approximate ReconfigurabLe Inference of Neural Networks at the Edge}, year={2024}, volume={71}, number={5}, pages={2105-2118}, doi={10.1109/TCSI.2024.3365952}}