fann_train_on_file function, without writing the status line.fann_create_array examplefann_create from scratchfann_create loading from a filefann_create from training datafann_runExamplefann - Fast Artificial Neural Network Library is written in ANSI C. The library implements multilayer feedforward ANNs, up to 150 times faster than other libraries. FANN supports execution in fixed point, for fast execution on systems like the iPAQ.
Copies of FANN can be obtained from our SourceForge project page, located at http://www.sourceforge.net/projects/fann/
You can currently get FANN as source code (fann-*.tar.bz2), Debian packages (fann-*.deb), or RPM's (fann-*.rpm).
FANN is available under the terms of the GNU Lesser General Public License.
RPMs are a simple way to manage packages, and is used on many common Linux distributions such as Red Hat, Mandrake, and SuSE.
Two separate packages exist; fann, the runtime library, and fann-devel, the development library and header files.
After downloading FANN, simply run (as root) the following command: rpm -ivh $PATH_TO_RPM
DEBs are packages for the Debian Linux distribution. Two separate packages exists libfann1 and libfann1-dev, where libfann1 is the runtime library and libfann1-dev is the development library.
Fann is included in the testing distribution of Debian, so testing users can simply run (as root) the following command: apt-get install libfann1 libfann1-dev.
After downloading the FANN DEB package, simply run (as root) the following command: dpkg -i $PATH_TO_DEB
FANN >= 1.1.0 includes a Microsoft Visual C++ 6.0 project file, which can be used to compile FANN for Windows. To build the library and examples with MSVC++ 6.0:
First, navigate to the MSVC++ directory in the FANN distribution and open the all.dsw workspace. In the Visual Studio menu bar, choose "Build" -> "Batch build...", select the project configurations that you would like to build (by default, all are selected), and press "rebuild all"
When the build process is complete, the library and examples can be found in the MSVC++\Debug and MSVC++\Release directories and the release versions of the examples are automatically copied into the examples where they are supposed to be run.
Compiling FANN from source code entails the standard GNU autotools technique. First, configure the package as you want it by typing (in the FANN directory), ./configure If you need help choosing the options you would like to use, try ./configure --help
Next, you have to actually compile the library. To do this, simply type make
Finally, to install the library, type make install. Odds are you will have to be root to install, so you may need to su to root before installing. Please remember to log out of the root account immediately after make install finishes.
Some people have experienced problems with compiling the library with some compilers, especially windows compilers which can not use GNU autotools. Please look through the help forum and the mailing list archives for info on how these problems was solved. If you do not find any information here, feel free to ask questions.
An ANN is normally run in two different modes, a training mode and an execution mode. Although it is possible to do this in the same program, using different programs is recommended.
There are several reasons to why it is usually a good idea to write the training and execution in two different programs, but the most obvious is the fact that a typical ANN system is only trained once, while it is executed many times.
The following is a simple program which trains an ANN with a data set and then saves the ANN to a file.
Example 1-1. Simple training example
#include "fann.h"
int main()
{
const float connection_rate = 1;
const float learning_rate = 0.7;
const unsigned int num_input = 2;
const unsigned int num_output = 1;
const unsigned int num_layers = 3;
const unsigned int num_neurons_hidden = 4;
const float desired_error = 0.0001;
const unsigned int max_iterations = 500000;
const unsigned int iterations_between_reports = 1000;
struct fann *ann = fann_create(connection_rate, learning_rate, num_layers,
num_input, num_neurons_hidden, num_output);
fann_train_on_file(ann, "xor.data", max_iterations,
iterations_between_reports, desired_error);
fann_save(ann, "xor_float.net");
fann_destroy(ann);
return 0;
}
The file xor.data, used to train the xor function:
4 2 1 0 0 0 0 1 1 1 0 1 1 1 0The first line consists of three numbers: The first is the number of training pairs in the file, the second is the number of inputs and the third is the number of outputs. The rest of the file is the actual training data, consisting of one line with inputs, one with outputs etc.
This example introduces several fundamental functions, namely fann_create,
fann_train_on_file,
fann_save, and fann_destroy.
The following example shows a simple program which executes a single input on the ANN. The program introduces two new functions
(fann_create_from_file and
fann_run) which were not used in the training procedure, as well as the fann_type
type.
Example 1-2. Simple execution example
#include <stdio.h>
#include "floatfann.h"
int main()
{
fann_type *calc_out;
fann_type input[2];
struct fann *ann = fann_create_from_file("xor_float.net");
input[0] = 0;
input[1] = 1;
calc_out = fann_run(ann, input);
printf("xor test (%f,%f) -> %f\n",
input[0], input[1], *calc_out);
fann_destroy(ann);
return 0;
}
If after reading the documentation you are still having problems, or have a question that is not covered in the documentation, please consult the fann-general mailing list. Archives and subscription information are available here.
This section describes some of the low-level functions and how they can be used to obtain more control of the fann library. For a full list of functions, lease see the API Reference, which has an explanation of all the fann library functions. Also feel free to take a look at the source code.
This section describes different procedures, which can help to get more power out of the fann library: Adjusting Parameters, Network Design, Understanding the Error Value, and Training and Testing.
Several different parameters exists in an ANN, these parameters are given defaults in the fann library, but they can be adjusted at runtime. There is no sense in adjusting most of these parameters after the training, since it would invalidate the training, but it does make sense to adjust some of the parameters during training, as will be described in Training and Testing. Generally speaking, these are parameters that should be adjusted before training.
The learning rate is one of the most important parameters, but unfortunately it is also a parameter which is hard to find a reasonable default for. I
(SN) have several times ended up using 0.7, but it is a good idea to test several different learning rates when training a network. It is also worth
noting that the activation function has a profound effect on the optimal learning rate [Thimm and Fiesler, 1997].
The learning rate can be set when creating the network, but it can also be set by the
fann_set_learning_rate function.
The initial weights are random values between -0.1 and 0.1, if other weights are preferred, the weights can be altered by the
fann_randomize_weights or
fann_init_weights function.
In [Thimm and Fiesler, High-Order and Multilayer Perceptron Initialization, 1997], Thimm and Fiesler state that, "An (sic) fixed weight variance of 0.2, which corresponds to a weight range of [-0.77, 0.77], gave the best mean performance for all the applications tested in this study. This performance is similar or better as compared to those of the other weight initialization methods."
The standard activation function is the sigmoid activation function, but it is also possible to use the threshold activation function. A list of the
currently available activation functions is available in the Activation Functions
section. The activation functions are chosen using the
fann_set_activation_function_hidden and
fann_set_activation_function_output functions.
These two functions set the activation function for the hidden layers and for the output layer. Likewise the steepness parameter used in the sigmoid
function can be adjusted with the
fann_set_activation_steepness_hidden and
fann_set_activation_steepness_output functions.
FANN distinguishes between the hidden layers and the output layer, to allow more flexibility. This is especially a good idea for users wanting discrete output from the network, since they can set the activation function for the output to threshold. Please note, that it is not possible to train a network when using the threshold activation function, due to the fact, that it is not differentiable.
When creating a network it is necessary to define how many layers, neurons and connections it should have. If the network become too large, the ANN will have difficulties learning and when it does learn it will tend to over-fit resulting in poor generalization. If the network becomes too small, it will not be able to represent the rules needed to learn the problem and it will never gain a sufficiently low error rate.
The number of hidden layers is also important. Generally speaking, if the problem is simple it is often enough to have one or two hidden layers, but as the problems get more complex, so does the need for more layers.
One way of getting a large network which is not too complex, is to adjust the connection_rate parameter given to
fann_create. If this parameter is 0.5, the constructed network will have the same amount of
neurons, but only half as many connections. It is difficult to say which problems this approach is useful for, but if you have a problem which can be
solved by a fully connected network, then it would be a good idea to see if it still works after removing half the connections.
The mean square error value is calculated while the ANN is being trained. Some functions are implemented, to use and manipulate this error value. The
fann_get_MSE function returns the error value and the
fann_reset_MSE resets the error value. The following explains how the mean square error
value is calculated, to give an idea of the value's ability to reveal the quality of the training.
If d is the desired output of an output neuron and y is the actual output of the neuron, the square error is (d - y) squared. If two output neurons exists, then the mean square error for these two neurons is the average of the two square errors.
When training with the fann_train_on_file function, an error value is printed. This
error value is the mean square error for all the training data. Meaning that it is the average of all the square errors in each of the training pairs.
Normally it will be sufficient to use the fann_train_on_file training function, but
sometimes you want to have more control and you will have to write a custom training loop. This could be because you would like another stop criteria,
or because you would like to adjust some of the parameters during training. Another stop criteria than the value of the combined mean square error could
be that each of the training pairs should have a mean square error lower than a given value.
Example 2-1.
The internals of the fann_train_on_file function, without writing the status line.
struct fann_train_data *data = fann_read_train_from_file(filename);
for(i = 1 ; i <= max_epochs ; i++) {
fann_reset_MSE(ann);
for (j = 0 ; j != data->num_data ; j++) {
fann_train(ann, data->input[j], data->output[j]);
}
if ( fann_get_MSE(ann) < desired_error ) {
break;
}
}
fann_destroy_train(data);
This piece of code introduces the fann_train function, which trains the ANN for one iteration
with one pair of inputs and outputs and also updates the mean square error. The
fann_train_data structure is also introduced, this structure is a container for the
training data in the file described in figure 10. The structure can be used to train the ANN, but it can also be used to test the ANN with data which it
has not been trained with.
Example 2-2. Test all of the data in a file and calculates the mean square error.
struct fann_train_data *data = fann_read_train_from_file(filename);
fann_reset_MSE(ann);
for(i = 0 ; i != data->num_data ; i++ ) {
fann_test(ann, data->input[i], data->output[i]);
}
printf("Mean Square Error: %f\n", fann_get_MSE(ann));
fann_destroy_train(data);
This piece of code introduces another useful function: fann_test function, which takes an input
array and a desired output array as the parameters and returns the calculated output. It also updates the mean square error.
With the knowledge of how to train and test an ANN, a new approach to training can be introduced. If too much training is applied to a set of data, the ANN will eventually over-fit, meaning that it will be fitted precisely to this set of training data and thereby loosing generalization. It is often a good idea to test, how good an ANN performs on data that it has not seen before. Testing with data not seen before, can be done while training, to see how much training is required in order to perform well without over-fitting. The testing can either be done by hand, or an automatic test can be applied, which stops the training when the mean square error of the test data is not improving anymore.
If a very low mean square error is required it can sometimes be a good idea to gradually decrease the learning rate during training, in order to make the adjusting of weights more subtle. If more precision is required, it might also be a good idea to use double precision floats instead of standard floats.
The threshold activation function is faster than the sigmoid function, but since it is not possible to train with this function, you may wish to consider an alternate approach:
While training the ANN you could slightly increase the steepness parameter of the sigmoid function. This would make the sigmoid function more steep and make it look more like the threshold function. After this training session you could set the activation function to the threshold function and the ANN would work with this activation function. This approach will not work on all kinds of problems, and has been successfully tested on the XOR function.
It is possible to run the ANN with fixed point numbers (internally represented as integers). This option is only intended for use on computers with no floating point processor, for example, the iPAQ, but a minor performance enhancement can also be seen on most modern computers [IDS, 2000].
The ANN cannot be trained in fixed point, which is why the training part is basically the same as for floating point numbers. The only difference is that
you should save the ANN as fixed point. This is done by the fann_save_to_fixed
function. This function saves a fixed point version of the ANN, but it also does some analysis, in order to find out where the decimal point should be.
The result of this analysis is returned from the function.
The decimal point returned from the function is an indicator of, how many bits is used for the fractional part of the fixed point numbers. If this number is negative, there will most likely be integer overflow when running the library with fixed point numbers and this should be avoided. Furthermore, if the decimal point is too low (e.g. lower than 5), it is probably not a good idea to use the fixed point version.
Please note, that the inputs to networks that should be used in fixed point should be between -1 and 1.
Example 3-1. An example of a program written to support training in both fixed point and floating point numbers
#include "fann.h"
#include <stdio.h>
int main()
{
fann_type *calc_out;
const float connection_rate = 1;
const float learning_rate = 0.7;
const unsigned int num_input = 2;
const unsigned int num_output = 1;
const unsigned int num_layers = 3;
const unsigned int num_neurons_hidden = 4;
const float desired_error = 0.001;
const unsigned int max_iterations = 20000;
const unsigned int iterations_between_reports = 100;
struct fann *ann;
struct fann_train_data *data;
unsigned int i = 0;
unsigned int decimal_point;
printf("Creating network.\n");
ann = fann_create(connection_rate, learning_rate, num_layers,
num_input,
num_neurons_hidden,
num_output);
printf("Training network.\n");
data = fann_read_train_from_file("xor.data");
fann_train_on_data(ann, data, max_iterations, iterations_between_reports, desired_error);
printf("Testing network.\n");
for(i = 0; i < data->num_data; i++){
calc_out = fann_run(ann, data->input[i]);
printf("XOR test (%f,%f) -> %f, should be %f, difference=%f\n",
data->input[i][0], data->input[i][1], *calc_out, data->output[i][0], fann_abs(*calc_out - data->output[i][0]));
}
printf("Saving network.\n");
fann_save(ann, "xor_float.net");
decimal_point = fann_save_to_fixed(ann, "xor_fixed.net");
fann_save_train_to_fixed(data, "xor_fixed.data", decimal_point);
printf("Cleaning up.\n");
fann_destroy_train(data);
fann_destroy(ann);
return 0;
}
Running a fixed point ANN is done much like running an ordinary ANN. The difference is that the inputs and outputs should be in fixed point
representation. Furthermore the inputs should be restricted to be between -multiplier and multiplier to
avoid integer overflow, where the multiplier is the value returned from
fann_get_multiplier. This multiplier is the value that a floating point number should
be multiplied with, in order to be a fixed point number, likewise the output of the ANN should be divided by this multiplier in order to be between zero
and one.
To help using fixed point numbers, another function is provided.
fann_get_decimal_point which returns the decimal point. The decimal point is the
position dividing the integer and fractional part of the fixed point number and is useful for doing operations on the fixed point inputs and outputs.
Example 3-2. An example of a program written to support both fixed point and floating point numbers
#include <time.h>
#include <sys/time.h>
#include <stdio.h>
#include "fann.h"
int main()
{
fann_type *calc_out;
unsigned int i;
int ret = 0;
struct fann *ann;
struct fann_train_data *data;
printf("Creating network.\n");
#ifdef FIXEDFANN
ann = fann_create_from_file("xor_fixed.net");
#else
ann = fann_create_from_file("xor_float.net");
#endif
if(!ann){
printf("Error creating ann --- ABORTING.\n");
return 0;
}
printf("Testing network.\n");
#ifdef FIXEDFANN
data = fann_read_train_from_file("xor_fixed.data");
#else
data = fann_read_train_from_file("xor.data");
#endif
for(i = 0; i < data->num_data; i++){
fann_reset_MSE(ann);
calc_out = fann_test(ann, data->input[i], data->output[i]);
#ifdef FIXEDFANN
printf("XOR test (%d, %d) -> %d, should be %d, difference=%f\n",
data->input[i][0], data->input[i][1], *calc_out, data->output[i][0], (float)fann_abs(*calc_out - data->output[i][0])/fann_get_multiplier(ann));
if((float)fann_abs(*calc_out - data->output[i][0])/fann_get_multiplier(ann) > 0.1){
printf("Test failed\n");
ret = -1;
}
#else
printf("XOR test (%f, %f) -> %f, should be %f, difference=%f\n",
data->input[i][0], data->input[i][1], *calc_out, data->output[i][0], (float)fann_abs(*calc_out - data->output[i][0]));
#endif
}
printf("Cleaning up.\n");
fann_destroy_train(data);
fann_destroy(ann);
return ret;
}
The fixed point ANN is not as precise as a floating point ANN, furthermore it approximates the sigmoid function by a stepwise linear function. Therefore,
it is always a good idea to test the fixed point ANN after loading it from a file. This can be done by calculating the mean square error as described
earlier. There is, however, one problem with this approach: The training data stored in the file is in floating
point format. Therefore, it is possible to save this data in a fixed point format from within the floating point program. This is done by the function
fann_save_train_to_fixed. Please note that this function takes the decimal point
as an argument, meaning that the decimal point should be calculated first by using the
fann_save_to_fixed function.
This section will briefly explain the theory of neural networks (hereafter known as NN) and artificial neural networks (hereafter known as ANN). For a more in depth explanation of these concepts please consult the literature; [Hassoun, 1995] has good coverage of most concepts of ANN and [Hertz et al., 1991] describes the mathematics of ANN very thoroughly, while [Anderson, 1995] has a more psychological and physiological approach to NN and ANN. For the pragmatic I (SN) could recommend [Tettamanzi and Tomassini, 2001], which has a short and easily understandable introduction to NN and ANN.
The human brain is a highly complicated machine capable of solving very complex problems. Although we have a good understanding of some of the basic operations that drive the brain, we are still far from understanding everything there is to know about the brain.
In order to understand ANN, you will need to have a basic knowledge of how the internals of the brain work. The brain is part of the central nervous system and consists of a very large NN. The NN is actually quite complicated, so the following discussion shall be relegated to the details needed to understand ANN, in order to simplify the explanation.
The NN is a network consisting of connected neurons. The center of the neuron is called the nucleus. The nucleus is connected to other nucleuses by means of the dendrites and the axon. This connection is called a synaptic connection.
The neuron can fire electric pulses through its synaptic connections, which is received at the dendrites of other neurons.
When a neuron receives enough electric pulses through its dendrites, it activates and fires a pulse through its axon, which is then received by other neurons. In this way information can propagate through the NN. The synaptic connections change throughout the lifetime of a neuron and the amount of incoming pulses needed to activate a neuron (the threshold) also change. This behavior allows the NN to learn.
The human brain consists of around 10^11 neurons which are highly interconnected with around 10^15 connections [Tettamanzi and Tomassini, 2001]. These neurons activates in parallel as an effect to internal and external sources. The brain is connected to the rest of the nervous system, which allows it to receive information by means of the five senses and also allows it to control the muscles.
It is not possible (at the moment) to make an artificial brain, but it is possible to make simplified artificial neurons and artificial neural networks. These ANNs can be made in many different ways and can try to mimic the brain in many different ways.
ANNs are not intelligent, but they are good for recognizing patterns and making simple rules for complex problems. They also have excellent training capabilities which is why they are often used in artificial intelligence research.
ANNs are good at generalizing from a set of training data. E.g. this means an ANN given data about a set of animals connected to a fact telling if they are mammals or not, is able to predict whether an animal outside the original set is a mammal from its data. This is a very desirable feature of ANNs, because you do not need to know the characteristics defining a mammal, the ANN will find out by itself.
When training an ANN with a set of input and output data, we wish to adjust the weights in the ANN, to make the ANN give the same outputs as seen in the training data. On the other hand, we do not want to make the ANN too specific, making it give precise results for the training data, but incorrect results for all other data. When this happens, we say that the ANN has been over-fitted.
The training process can be seen as an optimization problem, where we wish to minimize the mean square error of the entire set of training data. This problem can be solved in many different ways, ranging from standard optimization heuristics like simulated annealing, through more special optimization techniques like genetic algorithms to specialized gradient descent algorithms like backpropagation.
The most used algorithm is the backpropagation algorithm, but this algorithm has some limitations concerning, the extent of adjustment to the weights in each iteration. This problem has been solved in more advanced algorithms like RPROP [Riedmiller and Braun, 1993] and quickprop [Fahlman, 1988].
This is a list of all functions and structures in FANN.
struct fann * fann_create(float connection_rate, float learning_rate, unsigned int num_layers, unsigned int ...);
fann_create will create a new artificial neural network, and return
a pointer to it. The connection_rate controls how many
connections there will be in the network. If the connection rate is set to 1, the
network will be fully connected, but if it is set to 0.5 only half of the connections
will be set.
The num_layers is the number of layers including the input and output layer. This parameter is followed by one parameter for each layer telling how many neurons there should be in the layer.
This function appears in FANN >= 1.0.0.
struct fann * fann_create_array(float connection_rate, float learning_rate, unsigned int num_layers, unsigned int * neurons_per_layer);
fann_create_array will create a new artificial neural network, and return a pointer to
it. It is the same as fann_create, only it accepts an array as its final parameter
instead of variable arguments.
Example 5-1. fann_create_array example
unsigned int neurons_per_layer[3] = {2, 3, 1};
// The following two calls have identical results
struct fann * ann = fann_create_array(1.0f, 0.7f, 3, neurons_per_layer);
struct fann * ann2 = fann_create(1.0f, 0.7f, 3, 2, 3, 1);
fann_destroy(ann);
fann_destroy(ann2);
This function appears in FANN >= 1.0.5.
struct fann * fann_create_shortcut(float learning_rate, unsigned int num_layers, unsigned int ...);
fann_create_shortcut will create a new artificial neural network, and return
a pointer to it. The network will be fully connected, and will furthermore have all shortcut
connections connected.
Shortcut connections are connections that skip 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 num_layers is the number of layers including the input and output layer. This parameter is followed by one parameter for each layer telling how many neurons there should be in the layer.
This function appears in FANN >= 1.2.0.
struct fann * fann_create_shortcut_array(float learning_rate, unsigned int num_layers, unsigned int * neurons_per_layer);
fann_create_shortcut_array will create a new artificial neural network, and return a pointer to
it. It is the same as fann_create_shortcut, only it accepts an array as its final parameter
instead of variable arguments.
This function appears in FANN >= 1.2.0.
void fann_destroy(struct fann * ann);
fann_destroy will destroy an artificial neural network, properly freeing all associate
memory.
This function appears in FANN >= 1.0.0.
fann_type * fann_run(struct fann * ann, fann_type * input);
fann_run will run input through ann,
returning an array of outputs, the number of which being equal to the number of neurons in the output
layer.
This function appears in FANN >= 1.0.0.
void fann_randomize_weights(struct fann * ann, fann_type min_weight, fann_type max_height);
Randomizes the weight of each connection in ann, effectively resetting the network.
See also: Adjusting Parameters,
fann_init_weights
This function appears in FANN >= 1.0.0.
void fann_init_weights(struct fann * ann, struct fann_train_data * train_data);
This function behaves similarly to fann_randomize_weights.
It will use the algorithm developed by Derrick Nguyen and Bernard Widrow
[Nguyen and Widrow, 1990] 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 accepts a second argument, data, which is the training data that will be used to train the network.
See also: Adjusting Parameters,
fann_randomize_weights
This function appears in FANN >= 1.1.0.
void fann_print_connections(struct fann * ann);
fann_print_connections will print the connections of the ann in a compact matrix, for easy viewing of the internals of the ann.
The output from fann_print_connections on a small (2 2 1) network trained on the xor problem:
Layer / Neuron 012345 L 1 / N 3 ddb... L 1 / N 4 bbb... L 2 / N 6 ...cdaThis network have 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.
"." is a place where 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) has connection from the three neurons in the previous layer as is visible in the first two lines. The output neuron (6) has connections form the three neurons in the hidden layer 3 - 5 as is visible in the last line.
To simplify the matrix output neurons is not visible as neurons that connections can come from, and input and bias neurons are not visible as neurons that connections can go to.
This function appears in FANN >= 1.2.0.
void fann_save(struct fann * ann, const char * configuration_file);
fann_save will attempt to save ann to the file located at
configuration_file
This function appears in FANN >= 1.0.0.
void fann_save_to_fixed(struct fann * ann, const char * configuration_file);
fann_save_to_fixed will attempt to save ann to the file located at
configuration_file as a fixed-point network.
This is useful for training a network in floating points, and then later executing it in fixed point.
The function returns the bit position of the fix point, which can be used to find out how accurate the fixed point network will be. A high value indicates high precision, and a low value indicates low precision.
A negative value indicates very low precision, and a very strong possibility for overflow. (the actual fix point will be set to 0, since a negative fix point does not make sense).
Generally, a fix point lower than 6 is bad, and should be avoided. The best way to avoid this, is to have less connections to each neuron, or just less neurons in each layer.
The fixed point use of this network is only intended for use on machines that have no floating point processor, like an iPAQ. On normal computers the floating point version is actually faster.
This function appears in FANN >= 1.0.0.
void fann_train(struct fann * ann, fann_type * input, fann_type * output);
fann_train will train one iteration with a set of inputs, and a set of desired
outputs. The training will be done by the standard backpropagation algorithm.
This function appears in FANN >= 1.0.0.
fann_type * fann_test(struct fann * ann, fann_type * input, fann_type * desired_output);
Test with a set of inputs, and a set of desired outputs. This operation updates the mean square error, but does not change the network in any way.
This function appears in FANN >= 1.0.0.
float fann_get_MSE(struct fann * ann);
Reads the mean square error from the network. This value is calculated during training or testing, and can therefore sometimes be a bit off if the weights have been changed since the last calculation of the value.
This function appears in FANN >= 1.1.0. (before this
fann_get_error is used)
void fann_reset_MSE(struct fann * ann);
Resets the mean square error from the network.
This function appears in FANN >= 1.1.0. (before this
fann_reset_error is used)
struct fann_train_data * fann_read_train_from_file(char * filename);
fann_read_train_from_filewill load training data from a file.
The file should be formatted in the following way:
num_train_data num_input num_output
inputdata seperated by space
outputdata seperated by space
.
.
.
inputdata seperated by space
outputdata seperated by space
An example of a properly formatted file is provided in the Introduction.
This function appears in FANN >= 1.0.0.
void fann_save_train(struct data * train_data, FILE * filename);
Save train_data to filename.
This function appears in FANN >= 1.0.0.
void fann_save_to_fixed(struct data * train_data, FILE * filename, unsigned int decimal_point);
Save train_data as fixed point to filename.
This function appears in FANN >= 1.0.0.
void fann_destroy_train_data(struct fann_train_data * train_data);
Destroy the training data stored in train_data, freeing the associated memory.
This function appears in FANN >= 1.0.0.
float fann_train_epoch(struct fann * ann, struct fann_train_data * data);
Train one epoch with the training data stored in data. One epoch is where all of the training data is considered exactly once.
This function returns the MSE error as it is calculated either before or during the actual training. This is not the actual MSE after the training epoch, but since calculating this will require to go through the entire training set once more, it is more than adequate to use this value during training.
The training algorithm used by this function is chosen by the
fann_set_training_algorithm
function. The default training algorithm is FANN_TRAIN_RPROP.
This function appears in FANN >= 1.2.0.
float fann_test_data(struct fann * ann, struct fann_train_data * data);
Calculates the mean square error for a set of data.
This function appears in FANN >= 1.2.0.
void fann_train_on_data(struct fann * ann, struct fann_train_data * data, unsigned int max_epochs, unsigned int epochs_between_reports, float desired_error);
Trains annusing datauntil desired_erroris reached, or until max_epochsis surpassed.
The training algorithm used by this function is chosen by the
fann_set_training_algorithm
function. The default training algorithm is FANN_TRAIN_RPROP.
This function appears in FANN >= 1.0.0.
void fann_train_on_data_callback(struct fann * ann, struct fann_train_data * data, unsigned int max_epochs, unsigned int epochs_between_reports, float desired_error, int (*callback)(unsigned int epochs, float error));
Trains ann using data until desired_error is reached, or until max_epochs is surpassed.
This function behaves identically to
fann_train_on_data, except that
fann_train_on_data_callbackallows you to specify a function to be called every
epochs_between_reportsinstead of using the default reporting mechanism.
If the callback function returns -1 the training will terminate.
The callback function is very useful in GUI applications or in other applications which do not wish to report the progress on standard output. Furthermore the callback function can be used to stop the training at non standard stop criteria (see Training and Testing.)
This function appears in FANN >= 1.0.5.
The training algorithm used by this function is chosen by the
fann_set_training_algorithm
function. The default training algorithm is FANN_TRAIN_RPROP.
void fann_train_on_file(struct fann * ann, char * filename, unsigned int max_epochs, unsigned int epochs_between_reports, float desired_error);
Trains ann using the data in filename until desired_error is reached, or until max_epochs is surpassed.
The training algorithm used by this function is chosen by the
fann_set_training_algorithm
function. The default training algorithm is FANN_TRAIN_RPROP.
This function appears in FANN >= 1.0.0.
void fann_train_on_file_callback(struct fann * ann, char * filename, unsigned int max_epochs, unsigned int epochs_between_reports, float desired_error, int (*callback)(unsigned int epochs, float error));
Trains ann using the data in filename until desired_error is reached, or until max_epochs is surpassed.
This function behaves identically to
fann_train_on_file, except that
fann_train_on_file_callback allows you to specify a function to be called every
epochs_between_reports instead of using the default reporting mechanism.
The callback function works as described in
fann_train_on_data_callback
The training algorithm used by this function is chosen by the
fann_set_training_algorithm
function. The default training algorithm is FANN_TRAIN_RPROP.
This function appears in FANN >= 1.0.5.
void fann_shuffle_train_data(struct fann_train_data * data);
fann_shuffle_train_datawill randomize the order of the training data contained in
data.
This function appears in FANN >= 1.1.0.
struct fann_train_data * fann_merge_train_data(struct fann_train_data * data1, struct fann_train_data * data2);
fann_merge_train_datawill return a single set of training data which contains all data
from data1 and data2.
This function appears in FANN >= 1.1.0.
void fann_print_parameters(struct fann * ann);
Prints all the parameters of the network, for easy viewing of all the values.
This function appears in FANN >= 1.2.0.
unsigned int fann_get_training_algorithm(struct fann * ann);
Return the training algorithm (as described in Training algorithms) for a given network.
The default training algorithm is FANN_TRAIN_RPROP.
This function appears in FANN >= 1.2.0.
void fann_set_training_algorithm(struct fann * ann, unsigned int training_algorithm);
Set the training algorithm (as described in Training algorithms) of a network.
The default training algorithm is FANN_TRAIN_RPROP.
This function appears in FANN >= 1.2.0.
float fann_get_learning_rate(struct fann * ann);
Return the learning rate for a given network.
This function appears in FANN >= 1.0.0.
void fann_set_learning_rate(struct fann * ann, float learning_rate);
Set the learning rate of a network.
This function appears in FANN >= 1.0.0.
unsigned int fann_get_activation_function_hidden(struct fann * ann);
Return the activation function used in the hidden layers.
See Activation Functions for details on the activation functions.
This function appears in FANN >= 1.0.0.
fann_set_activation_function_hidden(struct fann * ann, unsigned int activation_function);
Set the activation function used in the hidden layers to activation_function.
See Activation Functions for details on the activation functions.
This function appears in FANN >= 1.0.0.
unsigned int fann_get_activation_function_output(struct fann * ann);
Return the activation function of the output layer.
See Activation Functions for details on the activation functions.
This function appears in FANN >= 1.0.0.
void fann_set_activation_function_output(struct fann * ann, unsigned int activation_function);
Set the activation function of the output layer to activation_function.
See Activation Functions for details on the activation functions.
This function appears in FANN >= 1.0.0.
fann_type fann_get_activation_steepness_hidden(struct fann * ann);
Return the steepness of the activation function of the hidden layers.
The steepness defaults to 0.5 and a larger steepness will make the slope of the activation function more steep, while a smaller steepness will make the slope less steep. A large steepness is well suited for classification problems while a small steepness is well suited for function approximation.
This function appears in FANN >= 1.2.0. and replaces the fann_get_activation_hidden_steepness function from FANN >= 1.0.0.
void fann_set_activation_steepness_hidden(struct fann * ann, fann_type steepness);
Set the steepness of the activation function of the hidden layers of ann to steepness.
The steepness defaults to 0.5 and a larger steepness will make the slope of the activation function more steep, while a smaller steepness will make the slope less steep. A large steepness is well suited for classification problems while a small steepness is well suited for function approximation.
This function appears in FANN >= 1.2.0. and replaces the fann_set_activation_hidden_steepness function from FANN >= 1.0.0.
fann_type fann_get_activation_steepness_output(struct fann * ann);
Return the steepness of the activation function of the hidden layers.
The steepness defaults to 0.5 and a larger steepness will make the slope of the activation function more steep, while a smaller steepness will make the slope less steep. A large steepness is well suited for classification problems while a small steepness is well suited for function approximation.
This function appears in FANN >= 1.2.0. and replaces the fann_get_activation_output_steepness function from FANN >= 1.0.0.
void fann_set_activation_steepness_output(struct fann * ann, fann_type steepness);
Set the steepness of the activation function of the hidden layers of ann to steepness.
The steepness defaults to 0.5 and a larger steepness will make the slope of the activation function more steep, while a smaller steepness will make the slope less steep. A large steepness is well suited for classification problems while a small steepness is well suited for function approximation.
This function appears in FANN >= 1.2.0. and replaces the fann_set_activation_output_steepness function from FANN >= 1.0.0.
void fann_set_train_error_function(struct fann * ann, unsigned int train_error_function);
Set the training error function (as described in Training Error Functions) of a network.
The default training error function is FANN_ERRORFUNC_TANH.
This function appears in FANN >= 1.2.0.
unsigned int fann_get_train_error_function(struct fann * ann);
Get the training error function (as described in Training Error Functions) of a network.
The default training error function is FANN_ERRORFUNC_TANH.
This function appears in FANN >= 1.2.0.
float fann_get_quickprop_decay(struct fann * ann);
The decay is a small negative valued number which is the factor that the weights should become smaller in each iteration. This is used to make sure that the weights do not become too high during training.
The default value for this parameter is -0.0001.
This function appears in FANN >= 1.2.0.
void fann_set_quickprop_decay(struct fann * ann, float quickprop_decay);
The decay is a small negative valued number which is the factor that the weights should become smaller in each iteration. This is used to make sure that the weights do not become too high during training.
The default value for this parameter is -0.0001.
This function appears in FANN >= 1.2.0.
float fann_get_quickprop_mu(struct fann * ann);
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 suppose to increase it.
The default value for this parameter is 1.75.
This function appears in FANN >= 1.2.0.
void fann_set_quickprop_mu(struct fann * ann, float quickprop_mu);
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 suppose to increase it.
The default value for this parameter is 1.75.
This function appears in FANN >= 1.2.0.
float fann_get_rprop_increase_factor(struct fann * ann);
The increase factor is a value larger than 1, which is used to increase the step-size during RPROP training.
The default value for this parameter is 1.2.
This function appears in FANN >= 1.2.0.
void fann_set_rprop_increase_factor(struct fann * ann, float rprop_increase_factor);
The increase factor is a value larger than 1, which is used to increase the step-size during RPROP training.
The default value for this parameter is 1.2.
This function appears in FANN >= 1.2.0.
float fann_get_rprop_decrease_factor(struct fann * ann);
The increase factor is a value smaller than 1, which is used to decrease the step-size during RPROP training.
The default value for this parameter is 0.5.
This function appears in FANN >= 1.2.0.
void fann_set_rprop_decrease_factor(struct fann * ann, float rprop_decrease_factor);
The increase factor is a value smaller than 1, which is used to decrease the step-size during RPROP training.
The default value for this parameter is 0.5.
This function appears in FANN >= 1.2.0.
float fann_get_rprop_delta_min(struct fann * ann);
The minimum step-size is a small positive number determining how small the minimum step may be.
The default value for this parameter is 0.0.
This function appears in FANN >= 1.2.0.
void fann_set_rprop_delta_min(struct fann * ann, float rprop_delta_min);
The minimum step-size is a small positive number determining how small the minimum step may be.
The default value for this parameter is 0.0.
This function appears in FANN >= 1.2.0.
float fann_get_rprop_delta_max(struct fann * ann);
The maximum step-size is a small positive number determining how small the minimum step may be.
The default value for this parameter is 50.0.
This function appears in FANN >= 1.2.0.
void fann_set_rprop_delta_max(struct fann * ann, float rprop_delta_max);
The maximum step-size is a small positive number determining how small the minimum step may be.
The default value for this parameter is 50.0.
This function appears in FANN >= 1.2.0.
unsigned int fann_get_num_input(struct fann * ann);
Return the number of neurons in the input layer of ann.
This function appears in FANN >= 1.0.0.
unsigned int fann_get_num_output(struct fann * ann);
Return the number of neurons in the output layer of ann.
This function appears in FANN >= 1.0.0.
unsigned int fann_get_total_neurons(struct fann * ann);
Return the total number of neurons in ann. This number includes the bias neurons.
This function appears in FANN >= 1.0.0.
unsigned int fann_get_total_connections(struct fann * ann);
Return the total number of connections in ann.
This function appears in FANN >= 1.0.0.
unsigned int fann_get_decimal_point(struct fann * ann);
Return the position of the decimal point in ann.
This function is only available when the ANN is in fixed point mode.
This function appears in FANN >= 1.0.0.
unsigned int fann_get_errno(struct fann_error * errdat);
Returns the numerical representation of the last error. The error codes are defined in fann_errno.h.
This function appears in FANN >= 1.1.0.
char * fann_get_errstr(struct fann_error * errdat);
Returns the last error.
Note: This will reset the network's error- any subsequent calls to fann_get_errno or
fann_get_errstr will yield 0 and NULL, respectively.
This function appears in FANN >= 1.1.0.
void fann_reset_errno(struct fann_error * errdat);
Reset the last error number.
This function appears in FANN >= 1.1.0.
void fann_reset_errstr(struct fann_error * errdat);
Reset the last error string.
This function appears in FANN >= 1.1.0.
void fann_set_error_log(struct fann_error * errdat, FILE * log);
Set the error log to log.
The error log defaults to stderr.
This function appears in FANN >= 1.1.0.
This structure is subject to change at any time. If you need to use the values contained herein, please see the Options functions. If these functions do not fulfill your needs, please open a feature request on our SourceForge project page.
Properties
The type of error that last occurred.
Where to log error messages.
A string representation of the last error.
The learning rate of the network.
The connection rate of the network. Between 0 and 1, 1 meaning fully connected.
Is 1 if shortcut connections are used in the ann otherwise 0 Shortcut connections are connections that skip layers. A fully connected ann with shortcut connections is an ann where neurons have connections to all neurons in all later layers.
ANNs with shortcut connections are created by fann_create_shortcut.
Pointer to the first layer (input layer) in an array of all the layers, including the input and output layer.
Pointer to the layer past the last layer in an array of all the layers, including the input and output layer.
Total number of neurons. Very useful, because the actual neurons are allocated in one long array.
Number of input neurons (not calculating bias)
Number of output neurons (not calculating bias)
Used to contain the error deltas used during training Is allocated during first training session, which means that if we do not train, it is never allocated.
Used to choose which activation function to use in the output layer.
Used to choose which activation function to use in the hidden layers.
Parameters for the activation function in the hidden layers.
Parameters for the activation function in the output layer.
Training algorithm used when calling fann_train_on_... and fann_train_epoch.
Fixed point only. The decimal point, used for shifting the fix point in fixed point integer operations.
Fixed point only. The multiplier, used for multiplying the fix point in fixed point integer operations. Only used in special cases, since the decimal_point is much faster.
An array of six members used by some activation functions to hold results for the hidden layer(s).
An array of six members used by some activation functions to hold values for the hidden layer(s).
An array of six members used by some activation functions to hold results for the output layer.
An array of six members used by some activation functions to hold values for the output layer.
Total number of connections. Very useful, because the actual connections are allocated in one long array.
Used to store outputs in.
The number of data used to calculate the mean square error.
The total error value. The real mean square error is MSE_value/num_MSE.
When using this, training is usually faster. Makes the error used for calculating the slopes higher when the difference is higher.
Decay is used to make the weights not go so high.
Mu is a factor used to increase and decrease the step-size.
Tells how much the step-size should increase during learning.
Tells how much the step-size should decrease during learning.
The minimum step-size.
The maximum step-size.
Used to contain the slope errors used during batch training Is allocated during first training session, which means that if we do not train, it is never allocated.
The previous step taken by the quickprop/rprop procedures. Not allocated if not used.
The slope values used by the quickprop/rprop procedures. Not allocated if not used.
This structure is subject to change at any time. If you need to use the values contained herein, please see the Training Data functions. If these functions do not fulfill your needs, please open a feature request on our SourceForge project page.
Properties
The type of error that last occurred.
Where to log error messages.
A string representation of the last error.
The number of sets of data in the array.
The number of inputs per set of data.
The number of outputs per set of data.
An array of num_data elements, each of which contain an array of num_input elements, which represent every item of input data.
An array of num_data elements, each of which contain an array of num_output elements, which represent every item of output data.
This structure is subject to change at any time. If you need to use the values contained herein, please see the Error Handling functions. If these functions do not fulfill your needs, please open a feature request on our SourceForge project page.
You may notice that this structure is identical to the first three properties of the fann and fann_train_data structures. This is so you can cast each of those structures to struct fann_error * when calling the Error Handling functions.
Properties
The type of error that last occurred.
Where to log error messages.
A string representation of the last error.
This structure is subject to change at any time. If you require direct access to the contents of this structure, you may want to consider contacting the FANN development team.
Properties
This property is not yet documented.
This property is not yet documented.
This property is not yet documented.
This property is not yet documented.
This structure is subject to change at any time. If you require direct access to the contents of this structure, you may want to consider contacting the FANN development team.
Properties
A pointer to the first neuron in the layer. When allocated, all the neurons in all the layers are actually in one long array, this is because we want to easily clear all the neurons at once.
A pointer to the neuron past the last neuron in the layer the number of neurons is last_neuron - first_neuron
These constants represent the training algorithms available within the fann library. The list will grow over time, but probably not shrink.
The training algorithm used by this function is chosen by the
fann_set_training_algorithm
function. The default training algorithm is FANN_TRAIN_RPROP.
Constants
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.
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 a 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 a better solutions with this algorithm.
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 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 in [Riedmiller and Braun, 1993], but the actual learning algorithm used here is the iRPROP- training algorithm [Igel and Hüsken, 2000] which is an variety of the standard RPROP training algorithm.
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 in [Fahlman, 1988].
These constants represent the activation functions available within the fann library. The list will grow over time, but probably not shrink.
Constants
Execution only - Threshold activation function.
This activation function gives output that is either 0 or 1.
Execution only - Threshold activation function.
This activation function gives output that is either -1 or 1.
Can not be used in fixed point - Linear activation function.
This activation function gives output that is unbounded.
Sigmoid activation function. One of the most used activation functions.
This activation function gives output that is between 0 and 1.
Stepwise linear approximation to sigmoid. Faster than sigmoid but a bit less precise.
This activation function gives output that is between 0 and 1.
Symmetric sigmoid activation function, AKA tanh. One of the most used activation functions.
This activation function gives output that is between -1 and 1.
Stepwise linear approximation to symmetric sigmoid. Faster than symmetric sigmoid but a bit less precise.
This activation function gives output that is between -1 and 1.
These constants represent the error functions used when calculating the error during training.
The training error function used is chosen by the
fann_set_train_error_function
function. The default training error function is FANN_ERRORFUNC_TANH.
Constants
The basic linear error function which simply calculates the error as the difference between the real output and the desired output.
The tanh error function is an error function that makes large deviations stand out, by altering the error value used when training the network. The idea behind this is that it is worse to have 1 output that misses the target by 100%, than having 10 outputs that misses the target by 10%.
This is the default error function and it is usually better. It can however give poor results with high learning rates.
These constants represent the various errors possible in fann, as defined by fann_errno.h.
Constants
No error.
Unable to open configuration file for reading
Unable to open configuration file for writing
Wrong version of configuration file
Error reading info from configuration file
Error reading neuron info from configuration file
Error reading connections from configuration file
Number of connections not equal to the number expected
Unable to open train data file for writing
Unable to open train data file for reading
Error reading training data from file
Unable to allocate memory
Unable to train with the selected activation function
Unable to use the selected activation function
Irreconcilable differences between two fann_train_data structures
int fann_save_internal(struct fann * ann, const char * configuration_file, unsigned int save_as_fixed);
fann_save_internal_fd is used internally to save an ANN to a file.
This function appears in FANN >= 1.0.0.
int fann_save_internal_fd(struct fann * ann, FILE * conf, const char * configuration_file, unsigned int save_as_fixed);
fann_save_internal_fd is used internally to save an ANN to a location pointed to by
conf. configuration_file is the name of the file, used only
for debugging purposes.
This function appears in FANN >= 1.1.0.
void fann_save_train_internal(struct fann_train_data * data, char * filename, unsigned int save_as_fixed, unsigned int decimal_point);
Saves the data in data to filename. save_as_fixed is either TRUE or FALSE. decimal_point tells FANN where the decimal point may be if using fixed point math.
This function appears in FANN >= 1.0.0.
void fann_save_train_internal_fd(struct fann_train_data * data, FILE * file, char * filename, unsigned int save_as_fixed, unsigned int decimal_point);
Saves the data in data to file. save_as_fixed is either TRUE or FALSE. decimal_point tells FANN where the decimal point may be if using fixed point math.
filename is used for debugging output only.
This function appears in FANN >= 1.1.0.
void fann_update_stepwise_hidden(struct fann * ann);
Update the stepwise function in the hidden layers of ann.
This function appears in FANN >= 1.0.0.
float fann_get_error(struct fann * ann);
This function is deprecated and will be removed in a future version. Use
fann_get_MSE instead.
This function appears in FANN >= 1.0.0, but is deprecated in FANN >= 1.1.0.
void fann_reset_error(struct fann * ann);
This function is deprecated and will be removed in a future version. Use
fann_reset_MSE instead.
This function appears in FANN >= 1.0.0, but is deprecated in FANN >= 1.1.0.
fann_type fann_get_activation_hidden_steepness(struct fann * ann);
Return the steepness of the activation function of the hidden layers.
The steepness defaults to 0.5 and a larger steepness will make the slope of the activation function more steep, while a smaller steepness will make the slope less steep. A large steepness is well suited for classification problems while a small steepness is well suited for function approximation.
This function is deprecated and will be removed in a future version. Use
fann_get_activation_steepness_hidden instead.
This function appears in FANN >= 1.0.0. and is deprecated in FANN >= 1.2.0.
void fann_set_activation_hidden_steepness(struct fann * ann, fann_type steepness);
Set the steepness of the activation function of the hidden layers of ann to steepness.
The steepness defaults to 0.5 and a larger steepness will make the slope of the activation function more steep, while a smaller steepness will make the slope less steep. A large steepness is well suited for classification problems while a small steepness is well suited for function approximation.
This function is deprecated and will be removed in a future version. Use
fann_set_activation_steepness_hidden instead.
This function appears in FANN >= 1.0.0. and is deprecated in FANN >= 1.2.0.
fann_type fann_get_activation_output_steepness(struct fann * ann);
Return the steepness of the activation function of the output layer.
The steepness defaults to 0.5 and a larger steepness will make the slope of the activation function more steep, while a smaller steepness will make the slope less steep. A large steepness is well suited for classification problems while a small steepness is well suited for function approximation.
This function is deprecated and will be removed in a future version. Use
fann_get_activation_steepness_output instead.
This function appears in FANN >= 1.0.0. and is deprecated in FANN >= 1.2.0.
void fann_set_activation_output_steepness(struct fann * ann, fann_type steepness);
Set the steepness of the activation function of the hidden layers of ann to steepness.
The steepness defaults to 0.5 and a larger steepness will make the slope of the activation function more steep, while a smaller steepness will make the slope less steep. A large steepness is well suited for classification problems while a small steepness is well suited for function approximation.
This function is deprecated and will be removed in a future version. Use
fann_set_activation_steepness_output instead.
This function appears in FANN >= 1.0.0. and is deprecated in FANN >= 1.2.0.
These functions allow you to interact with the FANN library from PHP.
This extension requires the FANN library, version 1.1.0 or later.
This extension supports the same activation functions as the library, a list of which can be found in the Activation Functions section.
The easiest way to install FANN-PHP is to use PEAR- if you have a fairly recent version of PHP installed, simply run pear install fann. Note that if there are no stable releases of FANN-PHP, you may have to specify the URI for the package, which can be obtained from http://pecl.php.net/fann.
If you cannot install FANN-PHP using PEAR, you can try following the (obsolete) instructions at http://www.cs.utexas.edu/users/UTCS/online-docs/php/pear/faq.install-pecl.html.
If you use one of these methods, you'll need to either dl('fann.so') or add it to your php.ini
If you use either of the above methods, you will probably need to be root.
Please only use this method if using the methods outlined in Using PEAR have failed.
If you wish to compile FANN-PHP into PHP itself, you can. First, uncompress the package into the ext subdirectory of your copy of the PHP source code, and rename the directory to ext/fann (from fann-x.x.x).
Next, you must rebuild the configure script- to do so, run ./buildconf from the PHP source directory.
From here on, the procedure is similar to when you built PHP originally- run ./configure with your desired options, plus --with-fann.
Finally, run make and make install. Note that you will probably need to be root for make install to work.
This method may require flex and bison to work- more information can be obtained at http://www.php.net/anoncvs.php
mixed fann_create(mixed data, float connection_rate, float learning_rate);
fann_create will create an artificial neural network using the data given.
If the first parameter is an array, fann_create will use the data and structure of the
array, as well as connection_rate and learning_rate.
If fann_create is called with a sole string argument, it will attempt to load an ANN
created with fann_save from the file at filename.
fann_create will return the artificial neural network on success, or FALSE if it fails.
Example 6-1. fann_create from scratch
<?php
$ann = fann_create(
/* Layers. In this case, three layers-
* two input neurons, 4 neurons on a
* hidden layer, and one output neuron. */
array(2, 4, 1),
1.0,
0.7);
?>
Example 6-2. fann_create loading from a file
<?php
$ann = fann_create("http://www.example.com/ann.net");
?>
See also fann_save.
This function appears in FANN-PHP >= 0.1.0.
bool fann_train(resource ann, mixed data, int max_iterations, double desired_error, int iterations_between_reports);
fann_train will train ann on the data supplied, returning TRUE
on success or FALSE on failure.
Resources is an artificial neural network returned by fann_create.
data must be either an array of training data, or the URI of a properly formatted training file.
fann_train will continue training until desired_error is
reached, or max_iterations is exceeded.
If iterations_between_reports is set, fann_create will output a
short progress report every iterations_between_reports. Default is 0 (meaning no
reports).
Example 6-1.
fann_create from training data
<?php
$ann = fann_create(array(2, 4, 1), 1.0, 0.7);
if ( fann_train($ann,
array(
array(
array(0,0), /* Input(s) */
array(0) /* Output(s) */
),
array(
array(0,1), /* Input(s) */
array(1) /* Output(s) */
),
array(
array(1,0), /* Input(s) */
array(1) /* Output(s) */
),
array(array(1,1), /* Input(s) */
array(0) /* Output(s) */
)
),
100000,
0.00001,
1000) == FALSE) {
exit('Could not train $ann.');
}
?>
This function appears in FANN-PHP >= 0.1.0.
bool fann_save(resource ann, string filename);
fann_save will save ann to filename,
returning TRUE on success or FALSE on failure.
See also fann_create.
This function appears in FANN-PHP >= 0.1.0.
mixed fann_run(resource ann, array input);
fann_run will run input through ann,
returning an an output array on success or FALSE on failure.
Example 6-1.
fann_runExample
<?php
if ( ($ann = fann_create("http://www.example.com/ann.net")) == FALSE )
exit("Could not create ANN.");
if ( fann_train($ann, "http://www.example.com/train.data", 100000, 0.00001) == FALSE )
exit("Could not train ANN.");
if ( ($output = fann_run($ann, array(0, 1))) == FALSE )
exit("Could not run ANN.");
else
print_r($output);
?>
This function appears in FANN-PHP >= 0.1.0.
void fann_randomize_weights(resource ann, float minimum, float maximum);
fann_randomize_weights will randomize the weights of all neurons in
ann, effectively resetting the network.
See also: Adjusting Parameters,
fann_init_weights
This function appears in FANN-PHP >= 0.1.0.
void fann_init_weights(resource ann, mixed training_data);
This function behaves similarly to fann_randomize_weights.
It will use the algorithm developed by Derrick Nguyen and Bernard Widrow [Nguyen and Widrow, 1990]
to set the weights in such a way as to speed up training.
The algorithm requires access to the range of the input data (ie, largest and smallest input), and therefore accepts a second argument, data, which is the training data that will be used to train the network.
See also: Adjusting Parameters,
fann_randomize_weights
This function appears in FANN-PHP >= 0.1.0.
float fann_get_MSE(resource ann);
fann_get_MSE will return the mean squared error (MSE) of ann,
or 0 if it is unavailable.
This function appears in FANN-PHP >= 0.1.0.
int fann_get_num_input(resource ann);
fann_get_num_inputwill return the number of input neurons in
ann.
See also fann_get_num_output,
fann_get_total_neurons.
This function appears in FANN-PHP >= 0.1.0.
int fann_get_num_output(resource ann);
fann_get_num_output will return the number of output neurons in
ann.
See also fann_get_num_input,
fann_get_total_neurons.
This function appears in FANN-PHP >= 0.1.0.
int fann_get_total_neurons(resource ann);
fann_get_total_neuronswill return the total number of neurons in
ann.
See also fann_get_num_input,
fann_get_num_output.
This function appears in FANN-PHP >= 0.1.0.
int fann_get_total_connections(resource ann);
fann_get_total_connections will return the total number of connections in
ann.
This function appears in FANN-PHP >= 0.1.0.
float fann_get_learning_rate(resource ann);
fann_get_learning_rate will return the learning rate of ann.
See also fann_set_learning_rate.
This function appears in FANN-PHP >= 0.1.0.
int fann_get_activation_function_hidden(resource ann);
fann_get_activation_function_hidden will return the activation function for the hidden
neurons in ann.
See also fann_set_activation_function_hidden.
This function appears in FANN-PHP >= 0.1.0.
int fann_get_activation_function_output(resource ann);
fann_get_activation_function_output will return the activation function for the output
neurons in ann.
See also fann_set_activation_function_output.
This function appears in FANN-PHP >= 0.1.0.
float fann_get_activation_steepness_hidden(resource ann);
fann_get_activation_steepness_hidden will return the steepness of the activation
function for the hidden neurons in ann.
See also
fann_set_activation_steepness_hidden.
This function appears in FANN-PHP >= 0.1.0.
float fann_get_activation_steepness_output(resource ann);
fann_get_activation_steepness_output will return the steepness of the activation
function for the output neurons in ann.
See also fann_set_activation_steepness_output.
This function appears in FANN-PHP >= 0.1.0.
float fann_set_learning_rate(resource ann);
fann_set_learning_rate will return the learning rate of ann.
See also fann_set_learning_rate.
This function appears in FANN-PHP >= 0.1.0.
void fann_set_activation_function_hidden(resource ann, int activation_function);
fann_set_activation_function_hidden sets the activation function for the hidden
neurons to activation_function, which must be one of the supported activation
functions.
See also fann_get_activation_function_hidden.
This function appears in FANN-PHP >= 0.1.0.
void fann_set_activation_function_output(resource ann, int activation_function);
fann_set_activation_function_output sets the activation function for the output
neurons to activation_function, which must be one of the supported activation
functions.
See also fann_get_activation_function_output.
This function appears in FANN-PHP >= 0.1.0.
void fann_set_activation_steepness_hidden(resource ann, float steepness);
fann_set_activation_steepness_hiddensets the steepness of the activation function
hidden neurons to steepness.
See also fann_get_activation_steepness_hidden.
This function appears in FANN-PHP >= 0.1.0.
void fann_set_activation_steepness_output(resource ann, float steepness);
fann_set_activation_steepness_output sets the steepness of the activation function
output neurons to steepness.
See also fann_get_activation_steepness_output.
This function appears in FANN-PHP >= 0.1.0.
These functions allow you to interact with the FANN library from Python.
This extension requires the FANN library, version 1.1.0 or later.
This python binding is provided by Vincenzo Di Massa (hawk.it@tiscalinet.it) and updated by Gil Megidish (gil@megidish.net)
Make sure to make and install the fann library first. Make sure that you have swig and python development files installed. Perhaps change the include directory of python. Then run 'make' to compile in the python directory.
Copy the generated _fann.so and fann.py files to python modules or into working directory.
After the install, just import fann and all the C functions will be available to your python code.
These functions allow you to interact with the FANN library from Delphi.
This extension requires the FANN library, version 1.2.0 or later.
This extension can be downloaded from the FANN library download section.
This Delphi binding is provided by Maurício Pereira Maia (mauricio@uaisol.com.br)
Make sure to make and install the fann library first. Put the file fannfloat.dll in your PATH. (If you want to use the fixed version you should define FIXEDFANN on fann.pas). Include fann.pas in your project and in your unit uses clause, and have fun! See the XorConsole sample for more details.
TFannNetwork is a Delphi component that encapsulates the Fann Library. You do not have to install TFannNetwork to use Fann on Delphi, but it will make the library more Delphi friendly. Currently it has only a small subset of all the library functions, but I hope that will change in the near future.
To install TFannNetwork you should follow all the previous steps and copy the FannNetwork.pas and Fann.dcr to your Delphi Library PATH. Choose Component/Install Component. In the Unit file name field, click on Browse and point to the fannnetwork.pas file. By default Delphi will install in the Borland User Components package, it might be changed using Package file name field or Into new package page. Click on Ok. A confirmation dialog will be shown asking if you want to build the package. Click on Yes. You have just installed TFannNetwork, now close the package window (Don't forget to put Yes when it ask if you want to save the package). See the FannNetwork.pas file or the Xor Sample.
If you are getting in trouble to use your own compiled FANN DLL with Delphi that might be because of the C++ naming mangle that changes between C++ compilers. You will need to make a TDUMP on your dll and changes all the name directives on fann.pas to the correct function names on your dll.
[Blake and Merz, 1998] C. Blake, C. Merz, 1998, UCI repository of machine learning databases, http://www.ics.uci.edu/mlearn/MLRepository.html .
[Darrington, 2003] J. Darrington, 2003, Libann, http://www.nongnu.org/libann/index.html .
[Fahlman, 1988] S.E. Fahlman, 1988, Faster-learning variations on back-propagation, http://www-2.cs.cmu.edu/afs/cs.cmu.edu/user/sef/www/publications/qp-tr.ps , An empirical study.
[LGPL] Free Software Foundation, 1999, GNU Lesser General Public License, Free Software Foundation, http://www.fsf.org/copyleft/lesser.html .
[Heller, 2002] J. Heller, 2002, Jet's Neural Library, http://www.voltar.org/jneural/jneural_doc/ .
[Hertz et al., 1991] J. Hertz, A. Krogh, R.G. Palmer, 1991, Introduction to The Theory of Neural Computing, Addison-Wesley Publishing Company.
[IDS, 2000] ID Software, 2000, Quake III Arena, http://www.idsoftware.com/games/quake/quake3-arena/ .
[Igel and Hüsken, 2000] Christian Igel, Michael Hüsken, 2000, Improving the Rprop Learning Algorithm, http://citeseer.ist.psu.edu/igel00improving.html .
[Kaelbling, 1996] L.P. Kaelbling, M.L. Littman, A.P. Moore, 1996, Reinforcement Learning: A New Survey, Journal of Artificial Intelligence Research, 4, 237-285.
[LeCun et al., 1990] Y. LeCun, J. Denker, S. Solla, R.E. Howard, L.D. Jackel, 1990, Advances in Neural Information Processing Systems II.
[Nguyen and Widrow, 1990] Reinforcement Learning, Derrick Nguyen, Bernard Widrow, 1990, Proc. IJCNN, 3, 21-26, http://www.cs.montana.edu/~clemens/nguyen-widrow.pdf .
[Nissen et al., 2003] S. Nissen, J. Damkjær, J. Hansson, S. Larsen, S. Jensen, 2003, Real-time image processing of an ipaq based robot with fuzzy logic (fuzzy), http://www.hamster.dk/~purple/robot/fuzzy/weblog/ .
[Nissen et al., 2002] S. Nissen, S. Larsen, S. Jensen, 2003, Real-time image processing of an iPAQ based robot (iBOT), http://www.hamster.dk/~purple/robot/iBOT/report.pdf .
[OSDN, 2003] 2003, SourceForge.net, http://sourceforge.net/ .
[Pendleton, 1993] R.C. Pendleton, 1993, Doing it Fast, http://www.gameprogrammer.com/4-fixed.html .
[Prechelt, 1994] L. Prechelt, 1994, Proben1: A set of neural network benchmark problems and benchmarking rules.
[Riedmiller and Braun, 1993] Riedmiller, Braun, 1993, A direct adaptive method for faster backpropagation learning: The RPROP algorithm, 586-591, http://citeseer.nj.nec.com/riedmiller93direct.html .
[Sarle, 2002] W.S. Sarle, , Neural Network FAQ, ftp://ftp.sas.com/pub/neural/FAQ2.html#A_binary .
[Pemstein, 2002] Dan Pemstein, 2002, ANN++, http://savannah.nongnu.org/projects/annpp/ .
[Thimm and Fiesler, High-Order and Multilayer Perceptron Initialization, 1997] Georg Thimm, Emile Fiesler, March 1997, High-Order and Multilayer Perceptron Initialization, IEEE Transactions on Neural Networks, 8, 2, 249-259, http://citeseer.ist.psu.edu/thimm96high.html .
[Thimm and Fiesler, 1997] G Thimm, E Fiesler, 1997, Optimal Setting of Weights, Learning Rate, and Gain, http://citeseer.ist.psu.edu/thimm97optimal.html .
[van Rossum, 2003] P. van Rossum, 2003, Lightweight neural network, http://lwneuralnet.sourceforge.net/ .
[van Waveren, 2001] J.P. van Waveren, 2001, The quake III arena bot, http://www.kbs.twi.tudelft.nl/Publications/MSc/2001-VanWaveren-MSc.html .
[Zell, 2003] A. Zell, 2003, Stuttgart neural network simulator, http://www-ra.informatik.uni-tuebingen.de/SNNS/ .