Skip to content

Latest commit

 

History

History
146 lines (116 loc) · 3.73 KB

README.md

File metadata and controls

146 lines (116 loc) · 3.73 KB

ffnn-c

This project is too old. TensorflowLite/Micro is better in most cases. You can use this project as an example if you want to custom build your own inference engine for small and simple ML models.

Feed-forward neural network in C programing language

ffnn-c is a minimal high-performance feed-forward neural networks implementation in C. Suitable for embeded systesms to run pre-trained neural networks. This library is useful when you have a pre-trained neural network structure and you want to use it in an embeded system.

The library supports parsing network specified in JSON.

For any questions please contact [email protected].

Running examples and tests with Makefile

make clean && make test && ./test
make clean && make example && ./example

Getting started

  • Prepare a json network string content
    {
        "layerSizes":[3,2,2],
        "activations":["sigmoid","sigmoid"],
        "weights":[
            {"col": 3, "row": 2, "grid":[1,1, 1.0, 1.0, 1.0, 1.0]},
            {"col\": 2, "row": 2, "grid":[1,1,1,1]}
        ],
        "biases":[
            {"vector" :[ 0.25, 0.25]},
            {"vector" :[ 0.5, 0.5]}]}
        ]
    }
  • NOTE: The first layerSizes value is the signal input length, weights, biases and activations should be specified for each layer.
  • Create a ffnn_network struct from json network content
Network* network = create_network(JSON_NETWORK_ACTIVATION_BY_LAYERS);
  • Run your input through the network
double * input = (double *) alloca(3 * sizeof(double));
input[0] = 1.0; input[1] = 1.0; input[2] = 1.0;
double * output = run_network(network, input);
  • Note: You are in charge of manage the memory of the input
  • Retrieve previous output
printf("Network saved response: %lf,%lf \n", network -> output[0], network -> output[1]);
  • Release memory of the network free_network(your_network);

Use protobuf to transfer network

Build proto-buf code from source on a Mac

  1. Prepare environments
brew install pkg-config
brew install protobuf
brew install automake
brew install libtool

cd into https://github.com/protobuf-c/protobuf-c
./autogen.sh && ./configure && make && make install
  1. Compile for C
protoc --c_out=. network.proto

Use Protobuf2 to serialize a neural network

from utility.network_pb2 import *

layerSizes = [5, 2, 2]
network = Network()
network.layerSizes.extend(layerSizes)
weights = [Weight(), Weight()]

weights[0].col = layerSizes[0]
weights[0].row = layerSizes[1]
weights[0].grid.extend([0,0,0,0,0,0,0,0,0,0])

weights[1].col = layerSizes[1]
weights[1].row = layerSizes[2]
weights[1].grid.extend([0,0,0,0])

biases = [Bias(), Bias()]

biases[0].vector.extend([0,0,0,0])
biases[1].vector.extend([0,0,0,0])

network.activations.extend([Network.RELU, Network.SOFTMAX])
network.weights.extend(weights)
network.biases.extend(biases)

print(network.SerializeToString())

Use Protobuf3 to serialize a neural network

from utility.network_pb3 import *

network = Network()
network.layerSizes.append(3)
network.layerSizes.append(2)
network.layerSizes.append(2)

network.activations.append(Network.ActivationType.RELU)
network.activations.append(Network.ActivationType.LINEAR)

weight1 = Weight()
weight1.col = 3
weight1.row = 2
weight1.grid.append(0.1)
weight1.grid.append(0.2)
weight1.grid.append(0.3)
weight1.grid.append(0.4)
weight1.grid.append(0.5)
weight1.grid.append(0.6)
network.weights.append(weight1)

weight2 = Weight()
weight2.col = 2
weight2.row = 2
weight2.grid.append(0.1)
weight2.grid.append(0.2)
weight2.grid.append(0.3)
weight2.grid.append(0.4)
network.weights.append(weight2)

bias1 = Bias()
bias1.vector.append(0.6)
bias1.vector.append(0.7)
bias2 = bias1

network.biases.append(bias1)
network.biases.append(bias2)

print(network.encode_to_bytes())