Skip to content
If FPGAs are universal function approximators, can they be used like neural networks?
Python C++ Makefile Other
Branch: master
Clone or download

README.md

Project Varro

If FPGAs are universal function approximators, can they be used like neural networks?

This project expands upon Adrian Thompson's famous paper using modern FPGAs and advancements in evolutionary algorithms with the goal of universal function approximation (just like a neural network!)

Installation

Training (Fitting)

To train the individual to solve a problem, for example to evolve the neural network over 500 generations using the multiobjective novelty search - reward (nsr-es) just run:

  • python -m varro.algo.experiment --purpose 'fit' --cxpb 0 --ngen 500 --strategy 'nsr-es' --problem_type 'sinx'

Prediction

To predict from a checkpoint using a numpy file of inputs:

  • python -m varro.algo.experiment --purpose 'predict' --ckptfolder ./checkpoint/varro/algo/sinx_2019-Nov-16-19\:00\:41 --strategy 'nsr-es' --X ./varro/algo/X_test.npy

Results

  • python -m varro.algo.experiment --purpose 'fit' --cxpb 0.0 --mutpb 1.0 --imutsigma 0.1 --ngen 100 --popsize 500 --strategy 'sga' --problem_type 'sinx' sinx_evolve_sga

  • python -m varro.algo.experiment --purpose 'fit' --cxpb 0.0 --mutpb 1.0 --imutsigma 0.1 --ngen 100 --popsize 500 --strategy 'ns-es' --problem_type 'sinx' sinx_evolve_ns-es

  • python -m varro.algo.experiment --purpose 'fit' --cxpb 0.0 --mutpb 1.0 --imutsigma 0.1 --ngen 100 --popsize 500 --strategy 'nsr-es' --problem_type 'sinx' sinx_evolve_nsr-es

You can’t perform that action at this time.