neural-network-fpga

Description

This project propose an alternative implementation of Perceptron Multilayer Neural Networks to solve binary classification problem in FPGAs devices.

Advices

There are some important observations you may know before using this work:

• This Network is based on fixed point Q4.12 instead float point (you can adapt)
• The layers are on 7-6-1 configuration (you may adapt as your need in network_pkg.vhd file)
• The input is configured to receive 2 values (you may adapt as your need in network_pkg.vhd file)
• The first and second layers are using the relu as activation function (you can adapt)
• The last layer is using and adaptation of logistic sigmoid function (you can adapt)
• The weights and biases are not trained here, you must to add the values in weights.txt and biases.txt
• Values in weights.txt and biases.txt must be following the order from first neuron of first layer to last neuron of last layer
• Is recommended some framework as TensorFlow to calculate the weights and biases
• Is recommended some script to convert the calculated weights and biases to fixed point and write the files

Example

I already added a trained network as propose to solve the classification problem below: