mlp

A multilayer perceptron implementation in C++ and Coffeescript

C++

Build

The `g` executable must be available in your system path, and must support the c11 standard.

Run cd c++ and use make to compile the test and cli programs.

Test with the XOR function

Run ./test, it will output the predictions for the XOR truth table:

2.1271e-15
1
1
3.44844e-15

CLI

Run ./cli

Example:

# of inputs> 2

# of hidden neurons> 2

# of hidden layers> 1

# of outputs> 1

# of samples> 4

# of epochs> 10000

Learning rate> 0.1

Training sample #1:
Input 2 numbers for the input features, space-separated:
0 0
Input 1 numbers for the expected output, space-separated:
0


Training sample #2:
Input 2 numbers for the input features, space-separated:
0 1
Input 1 numbers for the expected output, space-separated:
1


Training sample #3:
Input 2 numbers for the input features, space-separated:
1 0
Input 1 numbers for the expected output, space-separated:
1


Training sample #4:
Input 2 numbers for the input features, space-separated:
1 1
Input 1 numbers for the expected output, space-separated:
0


Training...
MSE: 1.36784e-29
Enter the testing samples: 2-dimensional vectors separated by spaces.
0 0
2.71603e-15
0 1
1
1 0
1
1 1
3.22927e-15

API

You can also use c++/mlp.hh as a library. It has no dependencies!

MLP(uint n_inputs, uint n_hidden_layers, uint n_hidden_neurons, uint n_outputs, double learning_rate,
    double min_weight_init, double max_weight_init); // Last two arguments are optional

double MLP::train(double* input_values, double* expected); // Returns Mean Squared Error

double* MLP::recall(double* input_values);

void set_learning_rate(double v);

void pretty_print(std::ostream& o);

It does also implement the << and >> operators for streams to dump and import trained neural networks 🔥

TODOs

• Parallelize the implementation

• Custom activation functions (with lambdas)

• Support for specifying the number of neurons for each layer

Coffeescript

Test with the XOR function

Run cd coffeescript and coffee test.coffee

Example output:

Testing the XOR function:
MSE: 1.814996379590531e-30917
1,1: 1.9984014443252818e-15
0,1: 0.9999999999999989
1,0: 0.9999999999999989
0,0: 0

API

    mlp = new MLP(nInputs, nOutputs, nHiddenLayers, nHiddenNeurons, learningRate)
    mse = mlp.train([0,0], [0])
    out = mlp.recall([0,0])