ANN

thx your code , can you provide testbench

VHDL test bench is provided upon request as it is explained in the documentation. The reason is that the VHDL test bench requires a lot of RAM and time to finish.

Example application along with C codes to use the IP form Microblaze or Zynq PS will be provided soon.

I have added a test bench which uses Octave/Matlab to initialize the weights and biases of the network.

Your code is great?

I wonder how to run your testbench, could you explain more detail of testbench, what is the input_data file's content?

Hi Dongyang, The testbench requires octave and ghdl to run. You will also need the nnet package for Octave. The package is here: https://octave.sourceforge.io/nnet/. Install it by following https://www.gnu.org/software/octave/doc/v4.2.0/Installing-and-Removing-Packages.html

Run the testbench script in linux by executing run.sh. The script first runs a basic neural net learning example in Octave, then exports the weights and biases to a vhdl file. This file is next used by the VHDL testbench (which runs in the GHDL simulator) to initialize the weight and bias RAM and perform output computation based on the input file. The input is a signal that goes from -1 to 1 a couple of times. We are testing how well the neural network approximates a sine (with the input normalized to -1 to 1 instead of -pi to pi). That's just some random example. Maybe a classification example would be more clear, you could make one if you would like to :). regards, Jerzy Stefanowicz