louisstow / fann.js Goto Github PK

FANN compiled through Emscripten

License: MIT License

Shell 14.07% JavaScript 47.51% HTML 38.42%

fann.js's Introduction

FANN.js

The FANN (Fast Artificial Neural Network) library compiled through Emscripten. This library contains some higher level bindings.

Much of the original documentation is still relevant. The noticable changes will be documented below. Note not all functions have a binding. You may see mention of Fixed vs Float mode, FANN.js uses Float mode.

How to use

FANN.js can be used almost like for like with the original library. These bindings provide an object oriented approach. For example take fann_print_connections(ann). In FANN.js this function is available on a Network instance as network.print_connections(). Notice the fann_ prefix isn't necessary nor is passing the neural network reference.

Before using the FANN library you should set a callback to window.FANN_ready.

FANN_ready = function () {
	// FANN.js is ready to use	
	var network = FANN.create([3, 3, 1]);
};

Demos

FANN.

Network create(neurons Array<Number>)

neurons the number of neurons in an array. The length of the array is the number of layers

Create a network with the provided structure. Returns an instance of the Network class. The neurons array argument specifies how many neurons are in that layer. The first index will be the input layer, the last index will be the output layer and anything between will be hidden layer(s).

var network = FANN.create([
	2, // Input layer: 2 neurons
	2, // Hidden layers (1): 2 neurons
	1  // Output layer: 1 neuron
]);

Network create(exported_network String)

exported_network String of a network exported through .export()

Create a network from a previously exported network as a string. This is useful when restoring a network.

TrainingData createTraining(Array data)

data a multidimensional array containing a set of inputs and desired outputs.

Create a TrainingData instance based on the provided data.

var tdata = FANN.createTraining([
    [[-1, -1], [-1]],
    [[ 1,  1], [-1]],
    [[-1,  1], [ 1]],
    [[ 1, -1], [ 1]]
]);

TrainingData createTraining(String data)

data a string of the training data. The required format is documented here.

Create a TrainingData instance based on the provided data. The string format will be the same as when exported via .export().

Network.

Array<Number> run(inputs Array<Number>)

inputs array of number inputs. Length should correspond to the amount of input neurons

Run the network with a set of inputs. Returns an array of the output neurons value.

var network = FANN.create([2, 2, 1]);
console.log(network.run([-1, 1]));

String export()

Returns a large string containing a snapshot of the network. You can store this string and restore a network based on the snapshot data using FANN.create().

Other functions

The following functions are available on a Network instance. The only difference between the original documented function is you do not need to pass the network reference as the first argument or use the fann_ prefix.

TrainingData.

String export()

Returns a large string containing a snapshot of the training data. You can store this string and create a TrainingData instance using FANN.createTraining().

Other functions

Constants

All of the enum and constants from the original library are available under the FANN. namespace. For example, FANN_TRAIN_INCREMENTAL becomes FANN.TRAIN_INCREMENTAL.

Compiling

After cloning, simply run ./build.sh build. The built file will be in the root directory as fann.js.

fann.js's People

Contributors

Stargazers

Watchers

fann.js's Issues

How to save and load ANN data?

I found export(), but no "import". Also export() outputs string, however in JS we expects JSON to play with.

Can you remove the dependency on the window object in fannjs?

Missing Scaling Params Prevents Scaling Data - FANN Error 18

So it seems that the function necessary for setting the the scaling params on a neural network is missing from fann.js.

For example, your dataset might look like this:

var UNSCALED_XOR_DATA = [
    [[0, 0], [0]],
    [[100, 100], [0]],
    [[0, 100], [100]],
    [[100, 0], [100]]
];

Where the range is not between -1 and 1 and FANN can rescale this data for us if we set the parameters based on the range we want and the dataset.

I.E.

FANN can turn the UNSCALED_XOR_DATA into:

var SCALED_XOR_DATA = [
    [[-1, -1], [-1]],
    [[ 1,  1], [-1]],
    [[-1,  1], [ 1]],
    [[ 1, -1], [ 1]]
];

However in order to do so you have to set the params on the ANN object which seems to be missing from FANN.js.

Here is an example of how this would work:

<script async src="../fann.js"></script>
<script>

var UNSCALED_XOR_DATA = [
    [[0, 0], [0]],
    [[ 100, 100], [0]],
    [[0, 100], [100]],
    [[100, 0], [100]]
];

function XOR () {
    
    ////////////////////
    // Configure ANN  //
    ////////////////////
    
    // New ANN
    NN = FANN.create([2, 3, 1]);
    
    // Set activation functions
    NN.set_activation_function_hidden(FANN.SIGMOID_SYMMETRIC);
    NN.set_activation_function_output(FANN.SIGMOID_SYMMETRIC);
    
    // Read raw (un-scaled) training data
    data = FANN.createTraining(UNSCALED_XOR_DATA);
    

    // Given the data range present in the data variable
    // configure the ANN with the desired scaling range
    NN.set_input_scaling_params(data, -1, 1);
    NN.set_output_scaling_params(data, -1, 1);
    
    // Scale the data range to -1 to 1
    // Scale the data - DOES NOT WORK because the ANN has not been configured with scaling params
    data = NN.scale_train(data, -1, 1); // Results in FANN Error 18: Scaling parameters not present.
    
    
    ///////////
    // Train //
    ///////////
    NN.train_on_data(data, 1000, 10, 0.01);
    

    //////////
    // Test //
    //////////
    var results = "";
    results += " -1 , -1 => " + NN.run([-1, -1])[0] + '\n<br>';
    results += "-1 ,  1 => " + NN.run([-1, 1])[0] + '\n<br>';
    results += " 1 , -1 => " + NN.run([1, -1])[0] + '\n<br>';
    results += " 1 ,  1 => " + NN.run([1, 1])[0] + '\n<br>';
    
    ////////////////////
    // Output Results //
    ////////////////////
    document.getElementById('results').innerHTML = results;
}

FANN_ready = function () {
    XOR();
};
</script>
<div id='results'></div>

What is the license which applies to this repository?

Hi,

What is the license that you're releasing this work?

Thanks.

Missing FANN Merge Training Data Binding

The binding for fann_merge_train_data is not available.

In an ideal world that binding would exist but here is a workaround if/until it ever gets added to FANN.js:

<meta charset="utf-8"/>
<script async src="../fann.js"></script>
<script>

function MergeData(data1, data2, data_has_header){
    
    // Split by new-lines
    data1 = data1.split(/\n/);
    data2 = data2.split(/\n/);
	
    // Remove header if present
    data1 = data1.slice(data_has_header ? 1 : 0);
    data2 = data2.slice(data_has_header ? 1 : 0);
  
    // Merge arrays
    var data = data1.concat(data2);
 
    // Create new FANN header
    var input_count = data1[0].split(' ').length;
    var output_count = data1[1].split(' ').length;
    var header = ((data1.length / 2)+(data2.length / 2)) + " " + input_count + " " + output_count;

    // Add the header to the beginning of the data
    data.unshift(header); 
    
    // Convert the data back into a new-line delimited string and return
    return data.join('\n', data);
}

function XOR () {

    // New ANN
    NN = FANN.create([2, 3, 1]);
	
    ////////////////////
    // Configure ANN  //
    ////////////////////
    NN.set_activation_steepness_hidden(1);
    NN.set_activation_steepness_output(1);
    NN.set_activation_function_hidden(FANN.SIGMOID_SYMMETRIC);
    NN.set_activation_function_output(FANN.SIGMOID_SYMMETRIC);
    NN.set_train_stop_function(FANN.STOPFUNC_BIT);
    NN.set_bit_fail_limit(0.01);
    NN.set_training_algorithm(FANN.TRAIN_RPROP);

    ////////////////////
    // Training Data  //
    ////////////////////
    var data_has_header = true;
    var data1 = "2 2 1\n-1 -1\n-1\n-1 1\n1";
    var data2 = "2 2 1\n1 -1\n1\n1 1\n-1";

    // Merge data1 & data2 
    // NOTE: The inputs and outputs MUST be the same length in both datasets
    var merged_xor = MergeData(data1, data2, data_has_header);
    //console.log(merged_xor);
    /*
	Result:

	"4 2 1
	-1 -1
	-1
	-1 1
	1
	1 -1
	1
	1 1
	-1"

    */
	
    data = FANN.createTraining(merged_xor);
	
    ///////////
    // Train //
    ///////////
    NN.init_weights(data);
    NN.train_on_data(data, 1000, 10, 0.01);
	
    //////////
    // Test //
    //////////
    var results = "";
    results += " -1 , -1 => " + NN.run([-1, -1])[0] + '\n<br>';
    results += "-1 ,  1 => " + NN.run([-1, 1])[0] + '\n<br>';
    results += " 1 , -1 => " + NN.run([1, -1])[0] + '\n<br>';
    results += " 1 ,  1 => " + NN.run([1, 1])[0] + '\n<br>';
    
    ////////////////////
    // Output Results //
    ////////////////////
    document.getElementById('results').innerHTML = results;
}

FANN_ready = function () {
    XOR();
};
</script>

<div id='results'></div>

louisstow / fann.js Goto Github PK

fann.js's Introduction

FANN.js

How to use

Demos

FANN.

Network.

Other functions

TrainingData.

Other functions

Constants

Compiling

fann.js's People

Contributors

Stargazers

Watchers

Forkers

fann.js's Issues

Recommend Projects

Recommend Topics

Recommend Org