FANN

NAME

AI::FANN

SYNOPSIS

# See below for details on export tags
    use AI::FANN :enum;
#                               Hidden
    #                         Input    |    Output
    #                              \   |   /
    given AI::FANN.new: layers => [ 2, 3, 1 ] {
        LEAVE .?destroy; # Make sure to clean up after yourself
# A sample data set for solving the XOR problem
        my $data = AI::FANN::TrainData.new: pairs => [
            [ -1, -1 ] => [ -1 ],
            [ -1,  1 ] => [  1 ],
            [  1, -1 ] => [  1 ],
            [  1,  1 ] => [ -1 ],
        ];
LEAVE $data.?destroy;
.activation-function: FANN_SIGMOID_SYMMETRIC;
# Train for up to 500,000 epochs
        # or until the MSE is less than 0.001
        # with no reports to STDOUT
        .train: $data,
            desired-error          => 0.001,
            max-epochs             => 500_000,
            epochs-between-reports => 0;
say .run: [ 1, -1 ];
    }
    # OUTPUT:
    # (0.9508717060089111)

DESCRIPTION

This distribution provides native bindings for the Fast Artificial Neural Network library (FANN). The aim of the library is to be easy to use, which makes it a good entrypoint and suitable for working on machine learning prototypes.

Creating networks, training them, and running them on input data can be done without much knowledge of the internals of ANNs, although the ANNs created will still be powerful and effective. Users with more experience and desiring more control will also find methods to parameterize most of the aspects of the ANNs, allowing for the creation of specialized and highly optimal ANNs.

Installation

The bindings for Raku make use of the system version of FANN. Please refer to your platform's instructions on how to install the library, or follow the instructions for compiling from source.

Error handling

The default behaviour for libfann is to print errors to standard error. In order to give the user more control over how to handle these errors, AI::FANN will raise exceptions whenever an error is encountered. When possible, these will be raised before an actual call to libfann is ever made.

When this is not possible, errors raised by libfann will be wrapped into exceptions of type X::AI::FANN. When capturing these, a string version of the error will be available in its message method, while its code method will return the error as a member of the AI::FANN::Error enum.

METHODS

The methods described below include readers, mutators, and methods that operate on the internal state of the network in more complex ways.

Some methods, like num-input are only for reading the internal state of the network, and will always return the value that was requested.

Other methods, like activation-function will act as both readers and mutators depending on the arguments that are passed.

When acting as readers, named parameters may be used to specify the scope of the reading. Some of these may be mandatory.

When acting as mutators, the new value should be passed as one or more positional arguments, with any named parameters specifying the possible scope of the mutation. All mutators always return the calling object, to allow for chaining. These will be marked in the signatures as returns self.

Most other methods, like reset-error or train, will also return the calling object, and may take named parameters. Some methods have different return values, like test or save that reflect the result of the operation. In all cases, the signature should specify the return value.

The sections below follow roughly the same structure as that used in the documentation of libfann.

Whenever possible, the underlying method that is being called will be indicated next to the method signatures.

Please refer to the libfann documentation for additional details.

Creation and Execution

new

# fann_create_shortcut
    # fann_create_sparse
    # fann_create_standard
    multi method new (
               :@layers,
        Num()  :$connection-rate,
        Bool() :$shortcut,
    ) returns AI::FANN
# fann_create_from_file
    multi method new (
        IO()   :$path,
    ) returns AI::FANN

Creates a new AI::FANN neural network. The constructor can be called in one of two ways.

If the path parameter is set, it will be coerced to a IO::Path and the network will be created based on the contents of that file (see save, below).

Alternatively, a list of integers can be passed as the layers parameter to specify the number of neurons in each layer, with the input layer being the first in the list, the output layer being the last in the list, and any remaining ones describing hidden layers.

By default, this will create a fully connected backpropagation neural network. There will be a bias neuron in each layer (except the output layer), and this bias neuron will be connected to all neurons in the next layer. When running the network, the bias nodes always emits 1.

To create a neural network that is not fully connected, a connection-rate parameter can be set to a number between 0 and 1, where 0 is a network with no connections, and 1 is a fully connected network.

If the shortcut flag is set, the resulting network will be fully connected, and it will have connections between neurons in non-contiguous layers. A fully connected network with shortcut connections is a network where all neurons are connected to all neurons in later layers, including direct connections from the input layer to the output layer.

The connection-rate and shortcut parameters are not compatible, and using both is an error.

run

# fann_run
    multi method run (
        CArray[num32] $input,
    ) returns CArray[num32]
multi method run (
        *@input,
    ) returns List

Run the input through the neural network, returning an array of outputs. The output array will have one value per neuron in the output layer.

The type of the return value depends on the type of the input.

If the input is provided as a CArray[num32] object, it will be used as-is and the return value will be of the same type. This is the fastest way to call this method.

If the input is passed as a List or Array, it will be internally converted to its C representation, and the return value will be a List object.

bit-fail

# fann_get_bit_fail
    method bit-fail returns Int

Returns the number of fail bits, or the number of output neurons which differ more than the bit fail limit (see bit-fail-limit). The bits are counted in all of the training data, so this number can be higher than the number of training data.

This value is reset by reset-error and updated by all the same functions which also update the mean square error (eg. test).

connection-rate

# fann_get_connection_rate
    method connection-rate returns Num

Get the connection rate used when the network was created.

num-input

# fann_get_num_input
    method num-input returns Int

Get the number of input neurons.

num-layers

# fann_get_num_layers
    method num-layers returns Int

Get the number of layers in the network.

num-output

# fann_get_num_output
    method num-output returns Int

Get the number of output neurons.

total-connections

# fann_get_total_connection
    method total-connections returns Int

Get the total number of connections in the entire network.

total-neurons

# fann_get_total_neurons
    method total-neurons returns Int

Get the total number of neurons in the entire network. This number includes the bias neurons, so a 2-4-2 network has 2+4+2 neurons, plus 2 bias neurons (one for each layer except the output one) for a total of 10.

network-type

# fann_get_network_type
    method network-type returns AI::FANN::NetType

Get the type of neural network it was created as.

layer-array

# fann_get_layer_array
    method layer-array returns List

Get the number of neurons in each layer in the network.

Bias is not included so the layers match the ones used in the constructor.

bias-array

# fann_get_bias_array
    method bias-array returns List

Get the number of bias in each layer in the network.

connection-array

# fann_get_connection_array
    method connection-array returns List

Get the connections in the network as a List of AI::FANN::Connection.

These objects encapsulate a connection between two neurons. They hold a number identifying the source and target neurons, which can be read with the from-neuron and to-neuron methods respectively; and the weight of the connection, which can be read with the weight method.

The weight method returns a writable container, which means that a new value can be set by using it on the left side of an assignment. Connection objects thus modified can then be passed to the weights method described below to update the connections of the network.

weights

multi method weights () returns List
# fann_set_weight
    multi method weights (
        Num()  $weight,
        Int() :$from! where * >= 0,
        Int() :$to!   where * >= 0,
    ) returns self
multi method weights (
        *@connections where { .all ~~ AI::FANN::Connection },
    ) returns self

Called with no arguments, returns the list of all connection weights as a List of Num. The weights will be in the same order as the connections returned by connection-array.

This method can also be used as a setter if called with either a weight as a positional argument and the numbers identifying the source and target neurons as the :from and :to named parameters respectively.

Alternatively, one or more AI::FANN::Connection objects (such as those returned by connection-array can be passed as positional arguments, in which case the weight in each connection will be used as the new value. See the documentation of that method for details.

Using this method as a setter returns the calling ANN, to allow for chaining.

connection-array

# fann_get_connection_array
    method connection-array returns List of AI::FANN::Connection

Get the connections in the network.

randomize-weights

# fann_randomize_weights
    method randomize-weights (
        Range:D $range,
    ) returns self

Give each connection a random weight between the endpoints of the specified Range object.

From the beginning the weights are random between -0.1 and 0.1.

This method is an alias for randomise-weights.

randomise-weights

# fann_randomize_weights
    method randomise-weights (
        Range:D $range,
    ) returns self

Give each connection a random weight between the endpoints of the specified Range object.

From the beginning the weights are random between -0.1 and 0.1.

This method is an alias for randomize-weights.

init-weights

# fann_init_weights
    method init-weights (
        AI::FANN::TrainData:D $data,
    ) returns self

Initialize the weights using Widrow + Nguyenโ€™s algorithm.

This function behaves similarly to randomize-weights. It will use the algorithm developed by Derrick Nguyen and Bernard Widrow to set the weights in such a way as to speed up training. This technique is not always successful, and in some cases can be less efficient than a purely random initialization.

The algorithm requires access to the range of the input data (ie, largest and smallest input), and therefore requires an AI::FANN::TrainData as its only positional argument. This should be the same data set used to train the network.

# fann_print_connections
    method print-connections returns self

Will print the connections of the network in a compact matrix, for easy viewing of its internals.

As an example, this is the output from a small (2 2 1) network trained on the xor problem:

Layer / Neuron 012345
    L   1 / N    3 BBa...
    L   1 / N    4 BBA...
    L   1 / N    5 ......
    L   2 / N    6 ...BBA
    L   2 / N    7 ......

This network has five real neurons and two bias neurons. This gives a total of seven neurons named from 0 to 6. The connections between these neurons can be seen in the matrix.

A period (".") indicates there is no connection, while a character tells how strong the connection is on a scale from a-z. The two real neurons in the hidden layer (neuron 3 and 4 in layer 1) have connections from the three neurons in the previous layer as is visible in the first two lines. The output neuron (6) has connections from the three neurons in the hidden layer 3 - 5, as shown in the fourth line.

To simplify the matrix output, neurons are not visible as neurons that connections can come from, and input and bias neurons are not visible as neurons that connections can go to.

# fann_print_parameters
    method print-parameters returns self

Prints all of the parameters and options of the network.

clone

# fann_copy
    method clone returns AI::FANN

Returns an exact copy of the calling AI::FANN object.

destroy

# fann_destroy
    method destroy returns Nil

Destroy the internal representation of this network. It's a good idea to make sure to call this for every object that has been created.

File Input / Output

save

# fann_save
    method save ( IO() $path ) returns Bool

Save the entire network to a configuration file.

The configuration file contains all information about the neural network and can be passed as the path parameter to the constructor to create an exact copy of the network and all of the associated parameters.

The only parameters that are not saved are the callback, error log, and user data, since they cannot safely be ported to a different location. Note that temporary parameters generated during training, like the mean square error, are also not saved.

Training

The methods in this section support fixed topology training.

When using this method of training, the size and topology of the ANN is determined in advance and the training alters the weights in order to minimize the difference between the desired output values and the actual output values.

For evolving topology training, see the Cascade Training section below.

train

multi method train (
        @input,
        @output,
    ) returns self
# fann_train
    multi method train (
        CArray[num32] $input,
        CArray[num32] $output,
    ) returns self
# fann_train_epoch
    # fann_train_on_data
    multi method train (
        AI::FANN::TrainData $data,
        Int() :$max-epochs,
        Int() :$epochs-between-reports,
        Num() :$desired-error,
    ) returns self
# fann_train_epoch
    # fann_train_on_file
    multi method train (
        IO() $path,
        Int() :$max-epochs,
        Int() :$epochs-between-reports,
        Num() :$desired-error,
    ) returns self

This method is used to train the neural network.

The first two candidates train a single iteration using the specified set of inputs and desired outputs in the input and output parameters. Inputs and outputs can be passed as CArray[num32] objects, or as arrays of numeric values, which will be converted internally to their C representation.

Since only one pattern is presented, training done this way is always incremental training (FANN_TRAIN_INCREMENTAL in the AI::FANN::Train enum).

The last two candidates train instead on an entire dataset. The first one takes a mandatory AI::FANN::TrainData object, while the second takes instead a filename that will be used to generate a training dataset internally. Both of these candidates will default to running a single iteration or "epoch". They can instead be used to train for a period of time by specifying the maximum number of iterations, the target error, and the number of iterations between reports. See callback for the code that gets executed to generate this report.

In both cases, the training uses the algorithm set with training-algorithm, and the parameters set for these training algorithms (see Training Algorithm Parameters below).

test

multi method test (
        @input,
        @output,
    ) returns List
# fann_test
    multi method test (
        CArray[num32] $input,
        CArray[num32] $output,
    ) returns CArray[num32]
multi method test (
        AI::FANN::TrainData $data,
    ) returns Num
multi method train (
        IO() $path,
    ) returns Num

Test the network with a set of inputs and desired outputs. This operation updates the mean square error, but does not change the network in any way.

Inputs and outputs can be passed as CArray[num32] objects, or as arrays of numeric values, which will be converted internally to their C representation.

These candidates return the same as the equivalent invokations of run.

Two more calling patterns are offered as shortcuts.

A AI::FANN::TrainData object can be passed as the data parameter, in which case the network will be tested with all the input and output data it contains.

Alternatively, the path parameter can be set to a value that can be coerced to a IO::Path object. In this case, an AI::FANN::TrainData will be internally read from the contents of this file and used as above.

These candidates return the updated mean square error for the network.

callback

multi method callback (
        :$delete where :so,
    ) returns self
# fann_set_callback
    method callback (
        &callback where {
            .cando: \(
                AI::FANN            $fann,
                AI::FANN::TrainData $data,
                uint32              $max-epochs,
                uint32              $epochs-between-reports,
                num32               $desired-error,
                uint32              $epoch,
            );
        }
    ) returns self

If called with a Callable as the first positional argument, this method will set that as the training callback. If called with a single :delete argument that evaluates to True, any previously set callback will be cleared, and the default callback will be restored.

The default callback function simply prints out some status information.

The callback will be called during training if using a AI::FANN::TrainData object either directly (with the :data argument to train) or indirectly (with the :path argument to the same method). It will be called once during the first epoch, and again every time the epoch is divisible by the value provided in the :epochs-between-reports argument to train.

The callback will be called with the AI::FANN object being trained, the AI::FANN::TrainData object that is being used for training, as well as the maximum number of allowed training epochs, the number of epochs between reports, and the target error for training that were set when training started as positional arguments. Additionally, the current epoch will also be passed as the final argument to the callback.

The callback can interrupt the training by returning False or a value that, when coerced into an Int evaluates to -1.

activation-function

# fann_get_activation_function
    multi method activation-function (
        Int    :$layer!,
        Int    :$neuron!,
    ) returns AI::FANN::ActivationFunc
# fann_set_activation_function
    # fann_set_activation_function_layer
    multi method activation-function (
        AI::FANN::ActivationFunc $function,
        Int    :$layer!,
        Int    :$neuron,
    ) returns self
# fann_set_activation_function_hidden
    # fann_set_activation_function_output
    multi method activation-function (
        AI::FANN::ActivationFunc $function,
        Bool() :$hidden,
        Bool() :$output,
    ) returns self

If called with no positional arguments, this method returns the activation function for neuron number and layer specified in the :neuron and :layer parameters respectively, counting the input layer as layer 0. It is not possible to get activation functions for the neurons in the input layer: doing so is an error.

If called with a member of the AI::FANN::ActivationFunc enum as the first positional argument, then this function will instead set this as the activation function for the specified layer and neuron, and return the calling AI::FANN object.

When used as a setter, specifying the layer is always required. This can be done with the :layer parameter, as described above, or with the :hidden or :output flags. The :hidden flag will set the activation function for all neurons in all hidden layers, while the :output flag will do so only for those in the output layer.

When setting the activation function using the :layer parameter, the :neuron parameter is optional. If none is set, all neurons in the specified layer will be modified.

activation-steepness

# fann_get_activation_steepness
    multi method activation-steepness (
        Int    :$layer!,
        Int    :$neuron!,
    ) returns Num
# fann_set_activation_steepness
    # fann_set_activation_steepness_layer
    multi method activation-steepness (
        Num()   $steepness,
        Int    :$layer!,
        Int    :$neuron,
    ) returns self
# fann_set_activation_steepness_hidden
    # fann_set_activation_steepness_output
    multi method activation-steepness (
        Num()   $steepness,
        Bool() :$hidden,
        Bool() :$output,
    ) returns self

If called with no positional arguments, this method returns the activation steepness for the neuron number and layer specified in the :neuron and :layer parameters respectively, counting the input layer as layer 0. It is not possible to get activation functions for the neurons in the input layer: trying to do so is an error.

If called with a positional argument, it will be coerced to a Num and this function will instead set this as the activation steepness for the specified layer and neuron and return the calling AI::FANN object.

When used as a setter, specifying the layer is always required. This can be done with the :layer parameter, as described above, or with the :hidden or :output flags. The :hidden flag will set the activation function for all neurons in all hidden layers, while the :output flag will do so only for those in the output layer.

When setting the activation steepness using the :layer parameter, the :neuron parameter is optional. If none is set, all neurons in the specified layer will be modified.

training-algorithm

# fann_get_training_algorithm
    multi method training-algorithm returns AI::FANN::Train
# fann_set_training_algorithm
    multi method training-algorithm (
        AI::FANN::Train $algorithm,
    ) returns self

If called with no positional arguments, this method returns the training algorithm as per the AI::FANN::Train enum. The training algorithm is used eg. when running train or cascade-train with a AI::FANN::TrainData object.

If a member of that enum is passed as the first positional argument, this method instead sets that as the new training algorithm and returns it.

Note that only FANN_TRAIN_RPROP and FANN_TRAIN_QUICKPROP are allowed during cascade training.

The default training algorithm is FANN_TRAIN_RPROP.

train-error-function

# fann_get_train_error_function
    multi method train-error-function returns AI::FANN::ErrorFunc
# fann_set_train_error_function
    multi method train-error-function (
        AI::FANN::ErrorFunc $function,
    ) returns self

If called with no positional arguments, this method returns the error function used during training as per the AI::FANN::ErrorFunc enum.

If a member of that enum is passed as the first positional argument, this method instead sets that as the new training error function and returns it.

The default training error function if FANN_ERRORFUNC_TANH.

train-stop-function

# fann_get_train_stop_function
    multi method train-stop-function returns AI::FANN::StopFunc
# fann_set_train_stop_function
    multi method train-stop-function (
        AI::FANN::StopFunc $function,
    ) returns self

If called with no positional arguments, this method returns the stop function used during training as per the AI::FANN::StopFunc enum.

If a member of that enum is passed as the first positional argument, this method instead sets that as the new training stop function and returns it.

The default training stop function if FANN_STOPFUNC_MSE.

bit-fail-limit

# fann_get_bit_fail_limit
    multi method bit-fail-limit returns Num
# fann_set_bit_fail_limit
    multi method bit-fail-limit (
        Num() $limit,
    ) returns self

If called with no positional arguments, this method returns the bit fail limit used during training. If called with a positional argument, it will be coerced to a Num and set as the new limit. In that case, this method returns the value that has been set.

The bit fail limit is used during training when the stop function is set to FANN_STOPFUNC_BIT (see train-stop-function).

The limit is the maximum accepted difference between the desired output and the actual output during training. Each output that diverges more than this limit is counted as an error bit. This difference is divided by two when dealing with symmetric activation functions, so that symmetric and asymmetric activation functions can use the same limit.

The default bit fail limit is 0.35.

learning-rate

multi method learning-rate returns Num
multi method learning-rate (
        Num() $rate,
    ) returns self

If called with no positional arguments, this method returns the learning rate used during training. If called with a positional argument, it will be coerced to a Num and set as the new learning rate.

The learning rate is used to determine how aggressive training should be for some of the training algorithms (FANN_TRAIN_INCREMENTAL, FANN_TRAIN_BATCH, FANN_TRAIN_QUICKPROP). Do however note that it is not used in FANN_TRAIN_RPROP.

The default learning rate is 0.7.

learning-momentum

multi method learning-momentum returns Num
multi method learning-momentum (
        Num() $momentum,
    ) returns self

If called with no positional arguments, this method returns the learning momentum used during training. If called with a positional argument, it will be coerced to a Num and set as the new learning momentum.

The learning momentum can be used to speed up FANN_TRAIN_INCREMENTAL training. Too high a momentum will however not benefit training. Setting the momentum to 0 will be the same as not using the momentum parameter. The recommended value of this parameter is between 0 and 1.

The default momentum is 0.

scale

# fann_scale_train
    multi method scale (
        AI::FANN::TrainData:D $data,
    ) returns self
# fann_scale_input
    # fann_scale_output
    multi method scale (
        CArray[num32] :$input,
        CArray[num32] :$output,
    ) returns self
# fann_scale_input
    # fann_scale_output
    multi method scale (
        :@input,
        :@output,
    ) returns self

This method will scale a set of inputs and outputs according to the scaling parameters set in this network (see scaling for how these are calculated and set).

If called with an AI::FANN::TrainData object, the scaling will apply to its input and output data. Alternatively, the :input and :output named parameters can be set to either CArray[num32] or to Array objects with the data to scale, which will be modified in-place according to the scaling parameters calculated for inputs and outputs respectively.

Calling this method before setting scaling parameters (with scaling) is an error. Calling this method after clearing the scaling parameters is not.

descale

# fann_descale_train
    multi method descale (
        AI::FANN::TrainData:D $data,
    ) returns self
# fann_descale_input
    # fann_descale_output
    multi method descale (
        CArray[num32] :$input,
        CArray[num32] :$output,
    ) returns self
# fann_descale_input
    # fann_descale_output
    multi method descale (
        :@input,
        :@output,
    ) returns self

This method will reverse the scaling performed by scale.

If called with an AI::FANN::TrainData object, the descaling will apply to its input and output data. Alternatively, the :input and :output named parameters can be set to either CArray[num32] or to Array objects with the data to descale, which will be modified in-place according to the scaling parameters calculated for inputs and outputs respectively.

Calling this method before setting scaling parameters (with scaling) is an error. Calling this method after clearing the scaling parameters is not.

scaling

# fann_set_scaling_params
    # fann_set_input_scaling_params
    # fann_set_output_scaling_params
    multi method scaling (
        AI::FANN::TrainData:D $data,
        Range :$output,
        Range :$input,
    ) returns self
# fann_clear_scaling_params
    multi method scaling (
        :$delete! where :so,
    ) returns self

Takes an AI::FANN::TrainData object that will be used to calculate the scaling parameters as a positional parameter, and Range objects representing the desired range for input and output values in the :input and :output named parameters respectively. At least one of these must be specified.

The scaling parameters set by this method can be cleared with the :delete flag. This will reset them a default value of -1..1.

reset-error

# fann_reset_MSE
    method reset-error returns self

Resets the mean square error from the network, and the number of bits that fail.

mean-square-error

# fann_get_MSE
    method mean-square-error returns Num

Reads the mean square error from the network. This value is calculated during training or testing (see train and test above), and can therefore sometimes be a bit off if the weights have been changed since the last calculation of the value.

Training Algorithm Parameters

These methods control the parameters used for specific training algorithms.

quickprop-decay

multi method quickprop-decay returns Num
multi method quickprop-decay (
        Num() $value where * <= 0,
    ) returns self

The decay is a small negative valued number which is the factor that the weights should become smaller in each iteration during quickprop training. This is used to make sure that the weights do not become too high during training.

If called with no positional arguments, this method returns the current decay value. If called with a positional argument, it will be coerced to a Num and set as the new decay.

The default decay is -0.0001.

quickprop-mu

multi method quickprop-mu returns Num
multi method quickprop-mu (
        Num() $value,
    ) returns self

The mu factor is used to increase and decrease the step-size during quickprop training. The mu factor should always be above 1, since it would otherwise decrease the step-size when it was supposed to increase it.

If called with no positional arguments, this method returns the current mu factor. If called with a positional argument, it will be coerced to a Num and set as the new mu factor.

The default mu factor is 1.75.

rprop-increase

multi method rprop-increase returns Num
multi method rprop-increase (
        Num() $value where * > 1,
    ) returns self

The increase factor is a value larger than 1, which is used to increase the step-size during RPROP training.

If called with no positional arguments, this method returns the current increase factor. If called with a positional argument, it will be coerced to a Num and set as the new increase factor.

The default increase factor is 1.2.

rprop-decrease

multi method rprop-decrease returns Num
multi method rprop-decrease (
        Num() $value where * < 1,
    ) returns self

The increase factor is a value larger than 1, which is used to decrease the step-size during RPROP training.

If called with no positional arguments, this method returns the current decrease factor. If called with a positional argument, it will be coerced to a Num and set as the new decrease factor.

The default increase factor is 0.5.

rprop-delta-range

multi method rprop-delta-range returns Range
multi method rprop-delta-range (
        Range $value where { not .infinite },
    ) returns self

The delta range determines the minimum and maximum allowed values for the step-size used during RPROP training.

If called with no positional arguments, this method returns the current delta range. If called with a Range as a positional argument, it will be set as the new delta range.

The default delta range is 0..50.

rprop-delta-zero

multi method rprop-delta-zero returns Num
multi method rprop-delta-zero (
        Num() $value where * > 0,
    ) returns self

The delta zero is a positive number determining the initial step size used during RPROP training.

If called with no positional arguments, this method returns the current initial step size. If called with a positional argument, it will be coerced to a Num and set as the new step size.

The default delta zero is 0.1.

sarprop-weight-decay-shift

multi method sarprop-weight-decay-shift returns Num
multi method sarprop-weight-decay-shift (
        Num() $value,
    ) returns self

If called with no positional arguments, this method returns the current weight decay shift used during SARPROP training. If called with a positional argument, it will be coerced to a Num and set as the new weight decay shift.

The default value is -6.644.

sarprop-error-threshold

multi method sarprop-error-threshold returns Num
multi method sarprop-error-threshold (
        Num() $value,
    ) returns self

If called with no positional arguments, this method returns the current error threshold factor used during SARPROP training. If called with a positional argument, it will be coerced to a Num and set as the new error threshold.

The default value is 0.1.

sarprop-step-error-shift

multi method sarprop-step-error-shift returns Num
multi method sarprop-step-error-shift (
        Num() $value,
    ) returns self

If called with no positional arguments, this method returns the current step error shift used during SARPROP training. If called with a positional argument, it will be coerced to a Num and set as the new step error shift.

The default value is 1.385.

sarprop-temperature

multi method sarprop-temperature returns Num
multi method sarprop-temperature (
        Num() $value,
    ) returns self

If called with no positional arguments, this method returns the current decay shift used during SARPROP training. If called with a positional argument, it will be coerced to a Num and set as the new decay shift.

The default value is 0.015.

Cascade Training

Cascade training differs from ordinary training in that it starts with an empty neural network and then adds neurons one by one, while it trains the neural network. The main benefit of this approach is that you do not have to guess the number of hidden layers and neurons prior to training, but cascade training has also proved better at solving some problems.

The basic idea of cascade training is that a number of candidate neurons are trained separate from the real network, then the most promising of these candidate neurons is inserted into the neural network. Then the output connections are trained and new candidate neurons are prepared. The candidate neurons are created as shortcut connected neurons in a new hidden layer, which means that the final neural network will consist of a number of hidden layers with one shortcut connected neuron in each.

For methods supporting ordinary, or fixed topology training, see the Training section above.

cascade-train

# fann_cascadetrain_on_data
    multi method cascade-train (
        AI::FANN::TrainData:D $data,
        Int() :$max-neurons!,
        Int() :$neurons-between-reports!,
        Num() :$desired-error!,
    ) returns self
# fann_cascadetrain_on_file
    multi method cascade-train (
        IO() $path,
        Int() :$max-neurons!,
        Int() :$neurons-between-reports!,
        Num() :$desired-error!,
    ) returns self

Trains the network on an entire dataset for a period of time using the Cascade2 training algorithm. The dataset can be passed as an AI::FANN::TrainData object in the data parameter. Alternatively, if the path is set, it will be coerced to an IO::Path object and the training data will be read from there instead.

This algorithm adds neurons to the neural network while training, which means that it needs to start with an ANN without any hidden layers. The neural network should also use shortcut connections, so the shortcut flag should be used when invoking new, like this

my $ann = AI::FANN.new: :shortcut,
        layers => [ $data.num-input, $data.num-output ];

cascade-num-candidates

# fann_get_cascade_num_candidates
    multi method cascade-num-candidates returns Int

This method returns the number of candidates used during training.

The number of candidates is calculated by multiplying the value returned by cascade-activation-functions-count, cascade-activation-steepnesses-count, and cascade-num-candidate-groups.

The actual candidates is defined by the cascade-activation-functions and cascade-activation-steepnesses arrays. These arrays define the activation functions and activation steepnesses used for the candidate neurons. If there are 2 activation functions in the activation function array and 3 steepnesses in the steepness array, then there will be 2x3=6 different candidates which will be trained. These 6 different candidates can be copied into several candidate groups, where the only difference between these groups is the initial weights. If the number of groups is set to 2, then the number of candidate neurons will be 2x3x2=12. The number of candidate groups can be set with cascade-num-candidate-groups.

The default number of candidates is 6x4x2 = 48

cascade-num-candidate-groups

# fann_get_cascade_num_candidate_groups
    multi method cascade-num-candidate-groups returns Int
# fann_set_cascade_num_candidate_groups
    multi method cascade-num-candidate-groups ( Int $groups ) returns self

If called with no positional arguments, this method returns the number of candidate groups used during training. If called with an Int as a positional argument, it will be set as the new value.

The number of candidate groups is the number of groups of identical candidates which will be used during training.

This number can be used to have more candidates without having to define new parameters for the candidates.

See cascade-num-candidates for a description of which candidate neurons will be generated by this parameter.

The default number of candidate groups is 2

cascade-candidate-limit

# fann_get_cascade_candidate_limit
    multi method cascade-candidate-limit returns Num
# fann_set_cascade_candidate_limit
    multi method cascade-candidate-limit ( Num() $value ) returns self

The candidate limit is a limit for how much the candidate neuron may be trained. It limits the proportion between the MSE and candidate score. Set this to a lower value to avoid overfitting and to a higher if overfitting is not a problem.

If called with no positional arguments, this method returns the current candidate limit. If called with a positional argument, it will be coerced to a Num and set as the new candidate limit.

The default candidate limit is 1000.

cascade-weight-multiplier

# fann_get_cascade_weight_multiplier
    multi method cascade-weight-multiplier returns Num
# fann_set_cascade_weight_multiplier
    multi method cascade-weight-multiplier ( Num() $value ) returns self

The weight multiplier is a parameter which is used to multiply the weights from the candidate neuron before adding the neuron to the neural network. This parameter is usually between 0 and 1, and is used to make the training a bit less aggressive.

If called with no positional arguments, this method returns the current weight multiplier. If called with a positional argument, it will be coerced to a Num and set as the new weight multiplier.

The default weight multiplier is 0.4

cascade-output-change-fraction

# fann_get_cascade_output_change_fraction
    multi method cascade-output-change-fraction returns Num
# fann_set_cascade_output_change_fraction
    multi method cascade-output-change-fraction ( Num() $value ) returns self

The cascade output change fraction is a number between 0 and 1 determining how large a fraction the mean-square-error should change within cascade-output-stagnation-epochs during training of the output connections, in order for the training not to stagnate. If the training stagnates, the training of the output connections will be ended and new candidates will be prepared.

If the MSE does not change by a fraction of the value returned by this method during a period of cascade-output-stagnation-epochs, the training of the output connections is stopped because the training has stagnated.

If the cascade output change fraction is low, the output connections will be trained more and if the fraction is high they will be trained less.

If called with no positional arguments, this method returns the current output change fraction. If called with a positional argument, it will be coerced to a Num and set as the new fraction.

The default cascade output change fraction is 0.01, which is equivalent to a 1% change in MSE.

cascade-candidate-change-fraction

# fann_get_cascade_candidate_change_fraction
    multi method cascade-candidate-change-fraction returns Num
# fann_set_cascade_candidate_change_fraction
    multi method cascade-candidate-change-fraction ( Num() $value ) returns self

The cascade candidate change fraction is a number between 0 and 1 determining how large a fraction the mean-square-error should change within cascade-output-stagnation-epochs during training of the candidate neurons, in order for the training not to stagnate. If the training stagnates, the training of candidate neurons will be ended and the best candidate will be selected.

If the MSE does not change by a fraction of the value returned by this method during a period of cascade-candidate-stagnation-epochs, the training of the candidate neurons is stopped because the training has stagnated.

If the cascade candidate change fraction is low, the candidate neurons will be trained more and if the fraction is high they will be trained less.

If called with no positional arguments, this method returns the current candidate change fraction. If called with a positional argument, it will be coerced to a Num and set as the new fraction.

The default cascade candidate change fraction is 0.01, which is equivalent to a 1% change in MSE.

cascade-candidate-stagnation-epochs

# fann_get_cascade_candidate_stagnation_epochs
    multi method cascade-candidate-stagnation-epochs returns Num
# fann_set_cascade_candidate_stagnation_epochs
    multi method cascade-candidate-stagnation-epochs ( Num() $value ) returns self

The number of cascade candidate stagnation epochs determines the number of epochs training is allowed to continue without changing the MSE by a fraction of cascade-candidate-change-fraction.

If called with no positional arguments, this method returns the current candidate stagnation epochs. If called with a positional argument, it will be coerced to a Num and set as the new candidate stagnation epochs.

The default number of cascade candidate stagnation epochs is 12.

cascade-output-stagnation-epochs

# fann_get_cascade_output_stagnation_epochs
    multi method cascade-output-stagnation-epochs returns Num
# fann_set_cascade_output_stagnation_epochs
    multi method cascade-output-stagnation-epochs ( Num() $value ) returns self

The number of cascade output stagnation epochs determines the number of epochs training is allowed to continue without changing the MSE by a fraction of cascade-output-change-fraction.

If called with no positional arguments, this method returns the current output stagnation epochs. If called with a positional argument, it will be coerced to a Num and set as the new output stagnation epochs.

The default number of cascade output stagnation epochs is 12.

cascade-activation-steepnesses-count

# fann_get_cascade_activation_steepnesses_count
    multi method cascade-activation_steepnesses_count returns Int

Returns the number of activation steepnesses in the list returned by cascade-activation-functions.

The default number of activation steepnesses is 4.

cascade-candidate-epochs

# fann_get_cascade_min_cand_epochs
    # fann_get_cascade_max_cand_epochs
    multi method cascade-candidate-epochs returns Range
# fann_set_cascade_min_cand_epochs
    # fann_set_cascade_max_cand_epochs
    multi method cascade-candidate-epochs (
        Range $value where { not .infinite },
    ) returns self
multi method cascade-candidate-epochs (
        Int :$min,
        Int :$max,
    ) returns self

The candidate epochs determines the minimum and maximum number of epochs the input connections to the candidates may be trained before adding a new candidate neuron.

If called with no positional arguments, this method returns the current candidate epoch range. If called with a Range as a positional argument, it will be set as the new candidate epoch range. This method can also be called with a value for the minimum or maximum end of the range as the :min and :max named parameters respectively.

The default candidate epoch range is 50..150

cascade-output-epochs

# fann_get_cascade_min_out_epochs
    # fann_get_cascade_max_out_epochs
    multi method cascade-output-epochs returns Range
# fann_set_cascade_min_out_epochs
    # fann_set_cascade_max_out_epochs
    multi method cascade-output-epochs (
        Range $value where { not .infinite },
    ) returns self
multi method cascade-output-epochs (
        Int :$min,
        Int :$max,
    ) returns self

The output epochs determines the minimum and maximum number of epochs the output connections may be trained after adding a new candidate neuron.

If called with no positional arguments, this method returns the current output epoch range. If called with a Range as a positional argument, it will be set as the new output epoch range. This method can also be called with a value for the minimum or maximum end of the range as the :min and :max named parameters respectively.

The default output epoch range is 50..150

cascade-activation-steepnesses

# fann_get_cascade_activation_steepnesses
    multi method cascade-activation-steepnesses returns List
# fann_set_cascade_activation_steepnesses
    multi method cascade-activation-steepnesses (
        CArray[num32] $steepnesses,
    ) returns self
multi method cascade-activation-steepnesses (
        *@steepnesses,
    ) returns self

If called with no positional arguments, this method returns the array of activation steepnesses used by the candidates. See cascade-num-candidates for a description of which candidate neurons will be generated by this array.

If called with a CArray[num32] object as the first positional argument, this method will instead use that as the new value. Alternatively, the values that would be in that array can be passed as positional arguments and they'll be internally converted to a C representation to use instead.

In either case, the new array must be just as long as defined by the count (see cascade-activation-steepnesses-count).

The default activation steepnesses are [ 0.25, 0.50, 0.75, 1.00 ].

cascade-activation-functions

# fann_get_cascade_activation_functions
    multi method cascade-activation-functions returns List
# fann_set_cascade_activation_functions
    multi method cascade-activation-functions (
        CArray[num32] $functions,
    ) returns self
multi method cascade-activation-functions (
        *@functions,
    ) returns self

If called with no positional arguments, this method returns the array of activation functions used by the candidates. See cascade-num-candidates for a description of which candidate neurons will be generated by this array.

If called with a CArray[num32] object as the first positional argument, this method will instead use that as the new value. Alternatively, the values that would be in that array can be passed as positional arguments and they'll be internally converted to a C representation to use instead.

In either case, the new array must be just as long as defined by the count (see cascade-activation-functions-count).

The default activation functions are [ FANN_SIGMOID, FANN_SIGMOID_SYMMETRIC, FANN_GAUSSIAN, FANN_GAUSSIAN_SYMMETRIC, FANN_ELLIOT, FANN_ELLIOT_SYMMETRIC, FANN_SIN_SYMMETRIC, FANN_COS_SYMMETRIC, FANN_SIN, FANN_COS ].

EXPORT TAGS

AI::FANN exports nothing by default. However, the following enums are available and can be exported using the :enum tag to export all enums, or the :error tag to export only the AI::FANN::Error enum.

AI::FANN::NetType

  • FANN_NETTYPE_LAYER

  • FANN_NETTYPE_SHORTCUT

AI::FANN::ActivationFunc

The activation functions used for the neurons during training. The activation functions can either be defined for a group of neurons by calling activation-function with the :hidden or :output parameters or it can be defined for a single neuron or layer with the :layer and :neuron parameters.

The functions are described with functions where

  • x is the input to the activation function,

  • y is the output,

  • s is the steepness and

  • d is the derivation.

The steepness of an activation function is defined in the same way by calling activation-steepness.

See the documentation for those functions for details.

  • FANN_LINEAR

    Linear activation function.

    -โˆž < y < โˆž
        y = xโ‹…s
        d = s
    
  • FANN_THRESHOLD

    Threshold activation function. Cannot be used during training.

    y = 0 if x < 0
        y = 1 if x โ‰ฅ 0
    
  • FANN_THRESHOLD_SYMMETRIC

    Symmetric threshold activation function. Cannot be used during training.

    y = -1 if x < 0
        y =  1 if x โ‰ฅ 0
    
  • FANN_SIGMOID

    Sigmoid activation function. This function is very commonly used.

    0 < y < 1
        y = 1 / ( 1 + exp( -2โ‹…sโ‹…x ) ) - 1
        d = 2โ‹…sโ‹…yโ‹…( 1 - yยฒ )
    
  • FANN_SIGMOID_STEPWISE

    Stepwise linear approximation to sigmoid. Faster than sigmoid, but a little less precise.

  • FANN_SIGMOID_SYMMETRIC

    Symmetric sigmoid activation function, also known as "tanh". This function is very commonly used.

    -1 < y < 1
        y = tanh(sโ‹…x) = 2 / ( 1 + exp( -2โ‹…sโ‹…x ) ) - 1
        d = sโ‹…( 1 - yยฒ )
    
  • FANN_SIGMOID_SYMMETRIC_STEPWISE

    Stepwise linear approximation to symmetric sigmoid. Faster than symmetric sigmoid, but a little less precise.

  • FANN_GAUSSIAN

    Gaussian activation function.

    0 < y < 1
        y = 0 when x = -โˆž
        y = 1 when x = 0
        y = 0 when x = โˆž
        y = exp( -xโ‹…sโ‹…xโ‹…s )
        d = -2โ‹…xโ‹…sโ‹…yโ‹…s
    
  • FANN_GAUSSIAN_SYMMETRIC

    Symmetric Gaussian activation function.

    -1 < y < 1
        y = -1 when x = -โˆž
        y =  1 when x = 0
        y = -1 when x = โˆž
        y = exp( -xโ‹…sโ‹…xโ‹…s )โ‹…2 - 1
        d = -2โ‹…xโ‹…sโ‹…yโ‹…s
    
  • FANN_GAUSSIAN_STEPWISE

    Not yet implemented.

  • FANN_ELLIOT

    Fast (sigmoid like) activation function defined by David Elliott.

    0 < y < 1
        y = xโ‹…s / 2 / ( 1 + |xโ‹…s| ) + 0.5
        d = s / ( 2 โ‹… ( 1 + |xโ‹…s| )ยฒ )
    
  • FANN_ELLIOT_SYMMETRIC

    Fast (symmetric sigmoid like) activation function defined by David Elliott.

    -1 < y < 1
        y = xโ‹…s  / ( 1 + |xโ‹…s| )
        d = s / ( 1 + |xโ‹…s| )ยฒ
    
  • FANN_LINEAR_PIECE

    Bounded linear activation function.

    0 โ‰ค y โ‰ค 1
        y = xโ‹…s
        d = s
    
  • FANN_LINEAR_PIECE_SYMMETRIC

    Bounded linear activation function.

    -1 โ‰ค y โ‰ค 1
        y = xโ‹…s
        d = s
    
  • FANN_SIN_SYMMETRIC

    Periodical sinus activation function.

    -1 โ‰ค y โ‰ค 1
        y = sin( xโ‹…s )
        d = sโ‹…cos( xโ‹…s )
    
  • FANN_COS_SYMMETRIC

    Periodical cosinus activation function.

    -1 โ‰ค y โ‰ค 1
        y = cos( xโ‹…s )
        d = sโ‹…-sin( xโ‹…s )
    
  • FANN_SIN

    Periodical sinus activation function.

    0 โ‰ค y โ‰ค 1
        y = sin( xโ‹…s ) / 2 + 0.5
        d = sโ‹…cos( xโ‹…s ) / 2
    
  • FANN_COS

    Periodical cosinus activation function.

    0 โ‰ค y โ‰ค 1
        y = cos( xโ‹…s ) / 2 + 0.5
        d = sโ‹…-sin( xโ‹…s ) / 2
    

AI::FANN::Train

The training algorithms used when training on AI::FANN::TrainData with functions like train with the :path or :data arguments. The incremental training alters the weights after each time it is presented an input pattern, while batch only alters the weights once after it has been presented to all the patterns.

  • FANN_TRAIN_INCREMENTAL

    Standard backpropagation algorithm, where the weights are updated after each training pattern. This means that the weights are updated many times during a single epoch. For this reason some problems will train very fast with this algorithm, while other more advanced problems will not train very well.

  • FANN_TRAIN_BATCH

    Standard backpropagation algorithm, where the weights are updated after calculating the mean square error for the whole training set. This means that the weights are only updated once during an epoch. For this reason some problems will train slower with this algorithm. But since the mean square error is calculated more correctly than in incremental training, some problems will reach better solutions with this algorithm.

  • FANN_TRAIN_RPROP

    A more advanced batch training algorithm which achieves good results for many problems. The RPROP training algorithm is adaptive, and does therefore not use the value set with learning-rate. Some other parameters can however be set to change the way the RPROP algorithm works, but it is only recommended for users with insight in how the RPROP training algorithm works. The RPROP training algorithm is described by Riedmiller and Braun (1993), but the actual learning algorithm used here is the iRPROP- training algorithm which is described by Igel and Hรผsken (2000) which is a variant of the standard RPROP training algorithm.

  • FANN_TRAIN_QUICKPROP

    A more advanced batch training algorithm which achieves good results for many problems. The quickprop training algorithm uses the learning-rate parameter along with other more advanced parameters, but it is only recommended to change these advanced parameters, for users with insight in how the quickprop training algorithm works. The quickprop training algorithm is described by Fahlman (1988).

  • FANN_TRAIN_SARPROP

    This is the same algorithm described in Nicholas and Tamas (1998).

AI::FANN::ErrorFunc

Error function used during training.

  • FANN_ERRORFUNC_LINEAR

    Standard linear error function.

  • FANN_ERRORFUNC_TANH

    Tanh error function, usually better but can require a lower learning rate. This error function aggressively targets outputs that differ much from the desired, while not targeting outputs that only differ a little that much. This activation function is not recommended for cascade training and incremental training.

AI::FANN::StopFunc

Stop criteria used during training.

  • FANN_STOPFUNC_MSE

    Stop criterion is Mean Square Error (MSE) value.

  • FANN_STOPFUNC_BIT

    Stop criterion is number of bits that fail. The number of bits means the number of output neurons which differ more than the bit fail limit (see bit-fail-limit). The bits are counted in all of the training data, so this number can be higher than the number of training data.

AI::FANN::Error

Used to define error events on AI::FANN and AI::FANN::TrainData objects.

  • FANN_E_NO_ERROR

    No error.

  • FANN_E_CANT_OPEN_CONFIG_R

    Unable to open configuration file for reading.

  • FANN_E_CANT_OPEN_CONFIG_W

    Unable to open configuration file for writing.

  • FANN_E_WRONG_CONFIG_VERSION

    Wrong version of configuration file.

  • FANN_E_CANT_READ_CONFIG

    Error reading info from configuration file.

  • FANN_E_CANT_READ_NEURON

    Error reading neuron info from configuration file.

  • FANN_E_CANT_READ_CONNECTIONS

    Error reading connections from configuration file.

  • FANN_E_WRONG_NUM_CONNECTIONS

    Number of connections not equal to the number expected.

  • FANN_E_CANT_OPEN_TD_W

    Unable to open train data file for writing.

  • FANN_E_CANT_OPEN_TD_R

    Unable to open train data file for reading.

  • FANN_E_CANT_READ_TD

    Error reading training data from file.

  • FANN_E_CANT_ALLOCATE_MEM

    Unable to allocate memory.

  • FANN_E_CANT_TRAIN_ACTIVATION

    Unable to train with the selected activation function.

  • FANN_E_CANT_USE_ACTIVATION

    Unable to use the selected activation function.

  • FANN_E_TRAIN_DATA_MISMATCH

    Irreconcilable differences between two AI::FANN::TrainData objects.

  • FANN_E_CANT_USE_TRAIN_ALG

    Unable to use the selected training algorithm.

  • FANN_E_TRAIN_DATA_SUBSET

    Trying to take subset which is not within the training set.

  • FANN_E_INDEX_OUT_OF_BOUND

    Index is out of bound.

  • FANN_E_SCALE_NOT_PRESENT

    Scaling parameters not present.

  • FANN_E_INPUT_NO_MATCH

    The number of input neurons in the ANN and data donโ€™t match.

  • FANN_E_OUTPUT_NO_MATCH

    The number of output neurons in the ANN and data donโ€™t match.

REFERENCES

  • Fahlman, S.E. (1988). "Faster-Learning Variations on Back-Propagation: An Empirical Study" in Proceedings of the 1988 Connectionist Models Summer School, Morgan Kaufmann.

  • Igel, C., Hรผsken, M. (2000) "Improving the Rprop Learning Algorithm" in Proceedings of the Second International ICSC Symposium on Neural Computation (NC 2000), pp. 115โ€”121. ICSC Academic Press.

  • Nicholas, K.T., Tamas, D.G. (1998) "Simulated Annealing and Weight Decay in Adaptive Learning: The SARPROP Algorithm". IEEE Transactions on Neural Networks 9(4), pp. 662โ€”668

  • Riedmiller, M., Braun, H. (1993). "A Direct Adaptive Method for Faster Backpropagation Leaning: the RPROP Algorithm" in IEEE International Conference on Neural Networks, pp. 586โ€”591, IEEE.

COPYRIGHT AND LICENSE

Copyright 2021 Josรฉ Joaquรญn Atria

This library is free software; you can redistribute it and/or modify it under the Artistic License 2.0.

AI::FANN v0.2.0

Raku binding to the Fast Artificial Neural Network library

Authors

  • Jonathan Scott Duff
  • Josรฉ Joaquรญn Atria

License

Artistic-2.0

Dependencies

Test Dependencies

Provides

  • AI::FANN
  • AI::FANN::Raw
  • AI::FANN::Raw::Base
  • AI::FANN::Raw::Cascade
  • AI::FANN::Raw::Creation
  • AI::FANN::Raw::Error
  • AI::FANN::Raw::IO
  • AI::FANN::Raw::Training

The Camelia image is copyright 2009 by Larry Wall. "Raku" is trademark of the Yet Another Society. All rights reserved.