转自http://www.codeproject.com/KB/recipes/neural_network.aspx
Introduction
I'd like to present a console based implementation of the backpropogation neural network C++ library I developed and used during my research in medical data classification and the CV library for face detection: Face Detection C++ library with Skin and Motion analysis. There are some good articles already present at The CodeProject, and you may consult them for the theory. In my code, I present the necessary features as input data preprocessing in the input layer with Minmax, Zscore, Sigmoidal, and Energy normalization. These parameters are obtained from the training set, and then used for preprocessing every incoming vector for classification. The console supports training data random separation to train, validation, and test sets before backpropagation training. Random separation allows to obtain a representative train set comparing performance on validation and test parts. A validation set is useful for preventing over-fitting by estimating the performance on that set. At the end of the backpropagation session, I save both network configurations, the one with the best performance on the validation set and the last training epoch configuration. For performance estimation, I use sensitivity, specificity, positive predictivity, negative predictivity, and accuracy metrics. For validation performance estimation in the case of biased data distribution (for example, in my Face Detection article, you may find that there are a lot more non-faces than faces, 19:1 ratio) I provide a geometric mean and F-measure metrics to support the scenario. To support a large amount of data vectors, I provide File Mapping based data load. That allows to map your hundreds of megs of training data to memory in 1 msec and start your training session immediately. For relatively small amounts of data, you may use a text file format. Finally, the console implementation is easier to use, you avoid a lot of mouse clicking in GUI applications, and may automate the process with batch files for choosing the right network topology, the best performance on the validation and test sets, and so on.
Background
Have a look at the CodeProject neural network articles. I also used tutorials from the generation5 site. For biased data distribution problems, I used: Evaluation of classifiers for an uneven class distribution problem. Have a look there to understand geometric mean and F-measure metrics.
Using the code
The help line to console is shown below:
argv[1] t-train
argv[2] network conf file
argv[3] cls1 files [0.9]
argv[4] cls2 files [0.1]
argv[5] epochs num
argv[6] [validation class]
argv[7] [test class]
argv[8] [validation TH 0.5]
argv[9] [vld metric mse]
argv[10] [norm]: [0-no], 1-minmax, 2-zscore, 3-softmax, 4-energy
argv[11] [error tolerance cls] +- 0.05 default
argv[1] r-run
argv[2] network conf file
argv[3] cls files
argv[4] [validation TH 0.5]
argv[5] [norm]: [0-no], 1-minmax, 2-zscore, 3-softmax, 4-energy
ann1dn.exe t net.nn cls1 cls2 3000 [tst.txt][val.txt]
[TH [0.5]][val type [mse]] [norm [0]] [err [0.05]]
ann1dn.exe r net.nn testcls [TH [0.5]] [norm [0]]
metrics: [0 - mse]
1 - AC
2 - sqrt(SE*SP)
3 - sqrt(SE*PP)
4 - sqrt(SE*SP*AC)
5 - sqrt(SE*SP*PP*NP*AC)
6 - F-measure b=1
7 - F-measure b=1.5
8 - F-measure b=3
The minimal number of parameters to start a training session:
>ann1dn.exe t network.nn data1_file data2_file 1000
It will use the network.nn file as a neural network, and load data form data1_file and data2_file, which represents data vectors from positive and negative classes, and train it for 1000 epochs.
The neural network file format is described in my Face Detection article. To start with random initialized weights before the training session, you need to provide only the number of layers and the number of neurons per layer in that file. For example, in the demo zip, you will find a iris.nn file:
3 4 8 1
Three layers, and 4, 8, and 1 neurons per layer.
It supports two data file formats. The text one is:
vector_name class_mark x1 x2 x3 ... xN ...
vector_name
is the name of your particular vector; it should not start with a numeric element, only with a letter.class_mark
is the non zero number corresponding to the class attribute: 1, 2, etc. In the console application, I use only 1 as positive class with a 0.9 desired output, and 2 as negative with 0.1 as the desired output. The next line contains the vector entries in integer or floating point format.
In the demo zip, the IRIS data is organized using a text format with four-dimensional entries:
virgi1 2 64 28 56 22 virgi2 2 67 31 56 24 virgi3 2 63 28 51 15 virgi4 2 69 31 51 23 virgi5 2 65 30 52 20 virgi6 2 65 30 55 18 ... setosa1 1 50 33 14 2 setosa2 1 46 34 14 3 setosa3 1 46 36 10 2 setosa4 1 51 33 17 5 setosa5 1 55 35 13 2 ...
The binary floating point file format is expedient when you have a large amount of data. The data is saved in a separate file as a sequence of floating point numbers in binary format, using 4 bytes per floating point number:
file1.dat (2x3 matrix)
[x11] [x12] [x13] [x21] [x22] [x23]
And, the dimensions of the data matrix are saved in the file with the same name but with a .hea extension.
file1.hea
2 3
In the previous example, the file file1.dat contains two three-dimensional vectors.
In that case, your data1_file or data2_file may contain the entries with the full path to the files and the class marks:
fullpath\file1.dat 1 fullpath\file2.dat 1 ...
The next parameters to the console application for backprop training are optional. You may use them for validation and testing of your network, for input data normalization, and error limits during training process.
>ann1dn t network.nn data1_file data2_file 1000 vld_file tst_file 0.5 2 2
This command line demonstrates that you use your validation and test sets in the vld_file and tst_file files in text or binary format, as described above, with a validation threshold 0.5 (that is, the network output greater than 0.5 attributes a data vector to a positive class), with geometric mean of sensitivity and specificity as the performance metric for validation stopping and Zscore normalization. The allowed validation metrics are specified at the end of the console help line. If your vld_file or tst_file are empty files, then the corresponding data set will be composed of randomly selected entries from your training set, with 25% records from each class.
The last eleventh argument to the console is the error tolerance. If the difference between the desired output and the mean network output for positive and negative classes is less than the error for 10 consecutive epochs, the training stops. During the backpropagation training, if the difference between the desired output and the network output for a particular vector is less than the error you specify, the network weights are not adjusted. This allows to correct the network weight connections for the data vectors which are not yet 'memorized'.
The next example demonstrates the sample training session for the IRIS data:
>ann1dn t iris.nn setosa_versi.dat virgi.dat 200 void void 0.5 2 3 loading data... cls1: 100 cls2: 50 files loaded. size: 4 samples validaton size: 25 12 validaton size: 26 13 normalizing minmax... training... epoch: 1 out: 0.555723 0.478843 max acur: 0.92 (epoch 1) se:84.00 sp:100.00 ac:89.19 epoch: 2 out: 0.582674 0.400396 max acur: 0.92 (epoch 1) se:84.00 sp:100.00 ac:89.19 epoch: 3 out: 0.626480 0.359573 max acur: 0.92 (epoch 3) se:84.00 sp:100.00 ac:89.19 epoch: 4 out: 0.655483 0.326918 max acur: 0.94 (epoch 4) se:96.00 sp:91.67 ac:94.59 epoch: 5 out: 0.699125 0.323879 max acur: 0.94 (epoch 5) se:88.00 sp:100.00 ac:91.89 epoch: 6 out: 0.715539 0.299085 max acur: 0.94 (epoch 6) se:88.00 sp:100.00 ac:91.89 epoch: 7 out: 0.733927 0.292526 max acur: 0.96 (epoch 7) se:92.00 sp:100.00 ac:94.59 epoch: 8 out: 0.750638 0.278721 max acur: 0.98 (epoch 8) se:96.00 sp:100.00 ac:97.30 epoch: 9 out: 0.774599 0.277550 max acur: 0.98 (epoch 8) se:96.00 sp:100.00 ac:97.30 epoch: 10 out: 0.774196 0.256110 max acur: 0.98 (epoch 8) se:96.00 sp:100.00 ac:97.30 epoch: 11 out: 0.793877 0.260753 max acur: 0.98 (epoch 8) se:96.00 sp:100.00 ac:97.30 epoch: 12 out: 0.806802 0.245758 max acur: 0.98 (epoch 8) se:96.00 sp:100.00 ac:97.30 epoch: 13 out: 0.804381 0.228810 max acur: 0.98 (epoch 13) se:96.00 sp:100.00 ac:97.30 epoch: 14 out: 0.814079 0.218740 max acur: 0.98 (epoch 13) se:96.00 sp:100.00 ac:97.30 epoch: 15 out: 0.827635 0.223827 max acur: 0.98 (epoch 13) se:96.00 sp:100.00 ac:97.30 epoch: 16 out: 0.832102 0.210360 max acur: 0.98 (epoch 13) se:96.00 sp:100.00 ac:97.30 epoch: 17 out: 0.840352 0.213165 max acur: 0.98 (epoch 17) se:96.00 sp:100.00 ac:97.30 epoch: 18 out: 0.848957 0.201766 max acur: 0.98 (epoch 18) se:96.00 sp:100.00 ac:97.30 epoch: 19 out: 0.844319 0.188338 max acur: 0.98 (epoch 19) se:96.00 sp:100.00 ac:97.30 epoch: 20 out: 0.856258 0.184954 max acur: 0.98 (epoch 19) se:96.00 sp:100.00 ac:97.30 epoch: 21 out: 0.853244 0.178349 max acur: 0.98 (epoch 19) se:96.00 sp:100.00 ac:97.30 epoch: 22 out: 0.867145 0.185852 max acur: 0.98 (epoch 22) se:96.00 sp:100.00 ac:97.30 epoch: 23 out: 0.863079 0.171684 max acur: 0.98 (epoch 23) se:96.00 sp:100.00 ac:97.30 epoch: 24 out: 0.870108 0.170253 max acur: 0.98 (epoch 24) se:96.00 sp:100.00 ac:97.30 epoch: 25 out: 0.873538 0.164185 max acur: 0.98 (epoch 25) se:96.00 sp:100.00 ac:97.30 epoch: 26 out: 0.871584 0.150496 max acur: 1.00 (epoch 26) se:100.00 sp:100.00 ac:100.00 epoch: 27 out: 0.879310 0.161155 max acur: 1.00 (epoch 26) se:100.00 sp:100.00 ac:100.00 epoch: 28 out: 0.879986 0.154784 max acur: 1.00 (epoch 26) se:100.00 sp:100.00 ac:100.00 epoch: 29 out: 0.880308 0.139083 max acur: 1.00 (epoch 26) se:100.00 sp:100.00 ac:100.00 epoch: 30 out: 0.890360 0.149518 max acur: 1.00 (epoch 26) se:100.00 sp:100.00 ac:100.00 epoch: 31 out: 0.888561 0.145144 max acur: 1.00 (epoch 26) se:100.00 sp:100.00 ac:100.00 epoch: 32 out: 0.880072 0.129197 max acur: 1.00 (epoch 32) se:100.00 sp:100.00 ac:100.00 epoch: 33 out: 0.896553 0.139937 max acur: 1.00 (epoch 32) se:100.00 sp:100.00 ac:100.00 epoch: 34 out: 0.893467 0.137607 max acur: 1.00 (epoch 32) se:100.00 sp:100.00 ac:100.00 epoch: 35 out: 0.893400 0.125793 max acur: 1.00 (epoch 32) se:100.00 sp:100.00 ac:100.00 epoch: 36 out: 0.905036 0.139306 max acur: 1.00 (epoch 32) se:100.00 sp:100.00 ac:100.00 epoch: 37 out: 0.900872 0.118167 max acur: 1.00 (epoch 32) se:100.00 sp:100.00 ac:100.00 epoch: 38 out: 0.909384 0.134014 max acur: 1.00 (epoch 32) se:100.00 sp:100.00 ac:100.00 training done. training time: 00:00:00:031 classification results: maxacur.nn train set: 49 25 sensitivity: 100.00 specificity: 100.00 +predictive: 100.00 -predictive: 100.00 accuracy: 100.00 validation set: 25 12 sensitivity: 100.00 specificity: 100.00 +predictive: 100.00 -predictive: 100.00 accuracy: 100.00 test set: 26 13 sensitivity: 88.46 specificity: 92.31 +predictive: 95.83 -predictive: 80.00 accuracy: 89.74 classification results: iris.nn train set: 49 25 sensitivity: 97.96 specificity: 100.00 +predictive: 100.00 -predictive: 96.15 accuracy: 98.65 validation set: 25 12 sensitivity: 96.00 specificity: 100.00 +predictive: 100.00 -predictive: 92.31 accuracy: 97.30 test set: 26 13 sensitivity: 88.46 specificity: 100.00 +predictive: 100.00 -predictive: 81.25 accuracy: 92.31
The network configuration is saved to maxacur.nn corresponding to the best performance on the validation set, and the last epoch configuration is saved to iris.nn. At the end, you may compare the results for them.
To use your trained networks for testing, just run the console with these parameters:
>ann1dn r iris.nn test_data
The test_data file is in text or binary format. You may provide as the fourth argument the threshold (default is 0.5) to obtain the ROC curve on your test set, for example, varying it between 0.0 and 1.0.
Neural Network Classes
The neural network is composed from the following classes:
ANNetwork
ANNLayer
ANeuron
ANLink
The ANNetwork
class contains the implementation of the neural network for users of the library. To avoid protected interface programming for the rest of the classes, I used friend
s. I'll describe the library structure first, and then provide the functions you need to use from the ANNetwork
class to maintain your own implementations.
The ANNetwork
contains an array of ANNLayer
layers. Every layer contains an array of ANeuron
neuron objects, and every neuron contains arrays of ANLink
input and output connections. With that design, you may arrange any desired network structure; however, in my implementation, I provide only feed-forward full connectionist structure.
The basic unit of the neural network is the neuron class, ANeuron
. You may add bias or input connection to it, represented as the ANLink
object.
void ANeuron::add_bias()
void ANeuron::add_input(ANeuron *poutn)
The bias connection always take 1.0f as an input value, as you know. With add_input()
, you add connection to the neuron, supplying with its argument the neuron from the previous layer to which it connects.
void ANeuron::add_input(ANeuron *poutn) { //poutn - Neuron from previous layer ANLink *plnk = new ANLink(this, poutn); inputs.push_back(plnk); if (poutn) poutn->outputs.push_back(plnk); }
So, every neuron 'knows' which neurons from the next layer connect to its output. The ANLink
is like the 'arrow', pointing from the neuron in the previous layer, ANLink::poutput_neuron
to the neuron in the next layer,ANLink::pinput_neuron
.
I organize a full connectionist neural network structure in this way:
void ANNetwork::init_links(const float *avec, const float *mvec, int ifunc, int hfunc) { ANNLayer *plr; //current layer ANNLayer *pprevlr; //previous layer ANeuron *pnrn; //neuron pointer int l = 0; /////input layer/ plr = layers[l++]; swprintf(plr->layer_name, L"input layer"); for (int n = 0; n < plr->get_neurons_number(); n++) { pnrn = plr->neurons[n]; pnrn->function = ifunc; pnrn->add_input(); //one input link for every "input layer" neuron if (avec) pnrn->inputs[0]->iadd = avec[n]; if (mvec) pnrn->inputs[0]->w = mvec[n]; else pnrn->inputs[0]->w = 1.0f; } ///// ////hidden layer's/// 1bias for (int i = 0; i < m_layers_number - 2 ; i++) { //1input [l-2 hidden] 1output pprevlr = plr; plr = layers[l++]; swprintf(plr->layer_name, L"hidden layer %d", i + 1); for (int n = 0; n < plr->get_neurons_number(); n++) { pnrn = plr->neurons[n]; pnrn->function = hfunc; pnrn->add_bias(); for (int m = 0; m < pprevlr->get_neurons_number(); m++) pnrn->add_input(pprevlr->neurons[m]); } } ////// ////output layer/ 1bias pprevlr = plr; plr = layers[l++]; swprintf(plr->layer_name, L"output layer"); for (int n = 0; n < plr->get_neurons_number(); n++) { pnrn = plr->neurons[n]; pnrn->function = hfunc; pnrn->add_bias(); for (int m = 0; m < pprevlr->get_neurons_number(); m++) pnrn->add_input(pprevlr->neurons[m]); } ////// }
The ANeuron
functions that induce the neuron to 'fire', that is to take data from its inputs and process them to its outputs, are:
void ANeuron::input_fire()
void ANeuron::fire()
The first one is used for input layer neurons only. The ANLink
contains an additional term, iadd
, used in normalization. The last one is used for hidden and output layer neurons:
inline void ANeuron::input_fire() { //input layer normalization oval = (inputs[0]->ival + inputs[0]->iadd) * inputs[0]->w; //single input for input layer neuron switch (function) { default: case LINEAR: break; case SIGMOID: oval = 1.0f / (1.0f + exp(float((-1.0f) * oval))); break; } //transfer my output to links connected to my output for (int i = 0; i < get_output_links_number(); i++) outputs[i]->ival = oval; } inline void ANeuron::fire() { //oval = SUM (in[]*w[]) oval = 0.0f; //compute output for Neuron for (int i = 0; i < get_input_links_number(); i++) oval += inputs[i]->ival * inputs[i]->w; switch (function) { default: case LINEAR: break; case SIGMOID: oval = 1.0f / (1.0f + exp(float((-1.0f) * oval))); break; } //transfer my output to links connected to my output for (int i = 0; i < get_output_links_number(); i++) outputs[i]->ival = oval; }
Now, you have some idea of the library internals, and further, I'll describe the ANNetwork
class which you can use for your own implementations.
You may load the neural network from the file, or arrange its structure specifying the number of layers and neurons per layer:
ANNetwork::ANNetwork(const wchar_t *fname);
ANNetwork::ANNetwork(int layers_number, int *neurons_per_layer);
int nerons_per_layer[4] = {128, 64, 32, 10}; ANNetwork *ann = new ANNetwork(4, neurons_per_layer); ann->init_links(); //feed-forward full connectionist structure ann->randomize_weights();
If you want your custom neural network configuration with recurrent or any other connections, you have to provide your own functions.
The ANNetwork::status()
function returns the class status after construction. Negative values indicate error, 0 - network loaded from file with success, and 1 - random weights.
To train, classify, and save your network, the following functions are provided:
bool ANNetwork::train(const float *ivec, float *ovec, const float *dsrdvec, floaterror = 0.05);
void ANNetwork::classify(const float *ivec, float *ovec);
bool ANNetwork::save(const wchar_t *fname) const;
ivec
and ivec
represent the input vector fed to a neural network and the output vector where it stores its results. Their dimensions should match the number of input and output neurons in the network structure. dsrdvec
is the desired output vector to which it adjusts its connections to match it within the error
tolerance. TheANNetwork::train()
function will return true in case the backpropogation took place, or false if the network output was within error
to the desired vector.
The backpropagation function uses this code:
bool ANNetwork::train(const float *ivec, float *ovec, const float *dsrdvec, float error) // 0.0 - 1.0 learning { float dst = 0.0f; //run network, computation of inputs to output classify(ivec, ovec); for (int n = 0; n < layers[m_layers_number-1]->get_neurons_number(); n++) { dst = fabs(ovec[n] - dsrdvec[n]); if (dst > error) break; } if (dst > error) { backprop_run(dsrdvec); //it was trained return true; } else //it wasnt trained return false; } void ANNetwork::backprop_run(const float *dsrdvec) { float nrule = m_nrule; //learning rule float alpha = m_alpha; //momentum float delta, dw, oval; //get deltas for "output layer" for (int n = 0; n < layers[m_layers_number-1]->get_neurons_number(); n++) { oval = layers[m_layers_number-1]->neurons[n]->oval; layers[m_layers_number-1]->neurons[n]->delta = oval * (1.0f - oval) * (dsrdvec[n] - oval); } //get deltas for hidden layers for (int l = m_layers_number - 2; l > 0; l--) { for (int n = 0; n < layers[l]->get_neurons_number(); n++) { delta = 0.0f; for (int i = 0; i < layers[l]->neurons[n]->get_output_links_number(); i++) delta += layers[l]->neurons[n]->outputs[i]->w * layers[l]->neurons[n]->outputs[i]->pinput_neuron->delta; oval = layers[l]->neurons[n]->oval; layers[l]->neurons[n]->delta = oval * (1 - oval) * delta; } } ////correct weights for every layer/// for (int l = 1; l < m_layers_number; l++) { for (int n = 0; n < layers[l]->get_neurons_number(); n++) { for (int i = 0; i < layers[l]->neurons[n]->get_input_links_number(); i++) { //dw = rule*Xin*delta + moment*dWprv dw = nrule * layers[l]->neurons[n]->inputs[i]->ival * layers[l]->neurons[n]->delta; dw += alpha * layers[l]->neurons[n]->inputs[i]->dwprv; layers[l]->neurons[n]->inputs[i]->dwprv = dw; //correct weight layers[l]->neurons[n]->inputs[i]->w += dw; } } } }
Now, you may compose your own networks and proceed to typical classification tasks used in OCR, computer vision, and so on.
License
This article, along with any associated source code and files, is licensed under The GNU General Public License (GPLv3)