Machine Learning (マシンラーニング) in this project, it implemented 3 layers ( Input Layer, Hidden Layer and Output Layer ) neural network (ニューラルネットワーク) and it named Back Propagation Neural Network (BPN). This version implemented QuickProp theory and Kecman's theory. KRBPN can be used in products recommendation (おすすめの商品), user behavior analysis (ユーザーの行動分析), data mining (データマイニング) and data analysis (データ分析).
This project designed for quickly operation for mobile device perform the basic data analysis. If you need help to know how to use this network, just ask me via email.
platform :ios, '7.0'
pod "KRBPN", "~> 2.0.2"#import "KRBPN.h"
@interface ViewController ()
@property (nonatomic, strong) KRBPN *_krBPN;
@end
@implementation ViewController
@synthesize _krBPN;
-(void)useSample1
{
_krBPN.activeFunction = KRBPNActivationBySigmoid;
//各輸入向量陣列值 & 每一筆輸入向量的期望值( 輸出期望 )，因使用 S 形轉換函數，故 Input 值域須為 [0, 1]，輸出目標為 [0, 1]
//Add the patterns, the weights connect with hidden layer, the output targets
[_krBPN addPatterns:@[@1, @0.1, @0.5, @0.2] outputGoals:@[@0.7f, @0.8f]]; //Pattern 1
[_krBPN addPatterns:@[@0, @1, @0.3, @0.9] outputGoals:@[@0.1f, @0.1f]]; //Pattern 2
[_krBPN addPatterns:@[@1, @0.3, @0.1, @0.4] outputGoals:@[@0.95f, @0.9f]]; //Pattern 3
//輸入層各向量值到隱藏層神經元的權重 ( 連結同一個 Net 的就一組一組分開，有幾個 Hidden Net 就會有幾組 )
//Add pattern-weights in Input layer to Hidden Layer
[_krBPN addPatternWeights:@[@0.2, @-0.3]]; //W15, W16
[_krBPN addPatternWeights:@[@0.4, @0.1]]; //W25, W26
[_krBPN addPatternWeights:@[@-0.5, @0.2]]; //W35, W36
[_krBPN addPatternWeights:@[@-0.1, @0.3]]; //W45, W46
//隱藏層神經元的偏權值 & 隱藏層神經元到輸出層神經元的權重值
//Add Hidden Layer biases of nets and the output weights in connect with output layer.
[_krBPN addHiddenLayerNetBias:-0.4f outputWeights:@[@-0.3f, @0.2f]]; //Net 5 bias, W57, W58 to output layer
[_krBPN addHiddenLayerNetBias:0.2f outputWeights:@[@-0.2f, @0.5f]]; //Net 6 bias, W67, W68 to output layer
//輸出層神經元偏權值, Net 7, 8
//Add the output layer biases
[_krBPN addOutputBiases:@[@0.0f, @0.1f]];
__block typeof(_krBPN) _weakKrBPN = _krBPN;
//訓練完成時( Training complete )
[_krBPN setTrainingCompletion:^(BOOL success, NSDictionary *trainedInfo, NSInteger totalTimes){
if( success )
{
NSLog(@"Training done with total times : %i", totalTimes);
NSLog(@"TrainedInfo 1 : %@", trainedInfo);
//Start in checking the network is correctly trained.
NSLog(@"======== Start in Verification ========");
[_weakKrBPN setTrainingCompletion:^(BOOL success, NSDictionary *trainedInfo, NSInteger totalTimes){
NSLog(@"Training done with total times : %i", totalTimes);
NSLog(@"TrainedInfo 2 : %@", trainedInfo);
}];
[_weakKrBPN recoverNetwork];
[_weakKrBPN directOutputAtInputs:@[@1, @0.1, @0.5, @0.2]];
}
}];
[_krBPN training];
}
//Only setups patterns and output goals, and 1 output.
-(void)useSample2
{
_krBPN.activeFunction = KRBPNActivationByTanh;
[_krBPN addPatterns:@[@1, @0.1, @0.5, @0.2] outputGoals:@[@0.7f]]; //Pattern 1
[_krBPN addPatterns:@[@-0.5, @0.8, @-0.3, @0.9] outputGoals:@[@-0.1f]]; //Pattern 2
[_krBPN addPatterns:@[@1, @0.3, @0.1, @0.4] outputGoals:@[@0.9f]]; //Pattern 3
__block typeof(_krBPN) _weakKrBPN = _krBPN;
//訓練完成時( Training complete )
[_krBPN setTrainingCompletion:^(BOOL success, NSDictionary *trainedInfo, NSInteger totalTimes){
if( success )
{
NSLog(@"Training done with total times : %i", totalTimes);
NSLog(@"TrainedInfo 1 : %@", trainedInfo);
//Start in checking the network is correctly trained.
NSLog(@"======== Start in Verification ========");
[_weakKrBPN setTrainingCompletion:^(BOOL success, NSDictionary *trainedInfo, NSInteger totalTimes){
NSLog(@"Training done with total times : %i", totalTimes);
NSLog(@"TrainedInfo 2 : %@", trainedInfo);
}];
[_weakKrBPN recoverNetwork];
[_weakKrBPN directOutputAtInputs:@[@-0.5, @0.8, @-0.3, @0.9]];
}
}];
[_krBPN trainingByRandomSettings];
}
//To learn and verify numbers 0 to 9. And only setups patterns and output goals, and 10 outputs.
-(void)useSample3
{
_krBPN.activeFunction = KRBPNActivationBySigmoid;
//1
[_krBPN addPatterns:@[@0, @0, @0, @0,
@0, @0, @0, @0,
@0, @0, @0, @0,
@0, @0, @0, @0,
@0, @0, @0, @0,
@0, @0, @0, @0,
@0, @0, @0, @1,
@1, @1, @1, @1,
@1, @1, @1, @1]
outputGoals:@[@1, @0, @0, @0, @0, @0, @0, @0, @0, @0]];
//2
[_krBPN addPatterns:@[@1, @0, @0, @0,
@1, @1, @1, @1,
@1, @1, @0, @0,
@0, @1, @0, @0,
@0, @1, @1, @0,
@0, @0, @1, @0,
@0, @0, @1, @1,
@1, @1, @1, @1,
@0, @0, @0, @1]
outputGoals:@[@0, @1, @0, @0, @0, @0, @0, @0, @0, @0]];
//3
[_krBPN addPatterns:@[@1, @0, @0, @0,
@1, @0, @0, @0,
@1, @1, @0, @0,
@0, @1, @0, @0,
@0, @1, @1, @0,
@0, @0, @1, @0,
@0, @0, @1, @1,
@1, @1, @1, @1,
@1, @1, @1, @1]
outputGoals:@[@0, @0, @1, @0, @0, @0, @0, @0, @0, @0]];
//4
[_krBPN addPatterns:@[@1, @1, @1, @1,
@1, @0, @0, @0,
@0, @0, @0, @0,
@0, @1, @0, @0,
@0, @0, @0, @0,
@0, @0, @1, @0,
@0, @0, @0, @1,
@1, @1, @1, @1,
@1, @1, @1, @1]
outputGoals:@[@0, @0, @0, @1, @0, @0, @0, @0, @0, @0]];
//5
[_krBPN addPatterns:@[@1, @1, @1, @1,
@1, @0, @0, @0,
@1, @1, @0, @0,
@0, @1, @0, @0,
@0, @1, @1, @0,
@0, @0, @1, @0,
@0, @0, @1, @1,
@0, @0, @0, @1,
@1, @1, @1, @1]
outputGoals:@[@0, @0, @0, @0, @1, @0, @0, @0, @0, @0]];
//6
[_krBPN addPatterns:@[@1, @1, @1, @1,
@1, @1, @1, @1,
@1, @1, @0, @0,
@0, @1, @0, @0,
@0, @1, @1, @0,
@0, @0, @1, @0,
@0, @0, @1, @1,
@0, @0, @0, @1,
@1, @1, @1, @1]
outputGoals:@[@0, @0, @0, @0, @0, @1, @0, @0, @0, @0]];
//7
[_krBPN addPatterns:@[@1, @0, @0, @0,
@0, @0, @0, @0,
@0, @1, @0, @0,
@0, @0, @0, @0,
@0, @0, @1, @0,
@0, @0, @0, @0,
@0, @0, @0, @1,
@1, @1, @1, @1,
@1, @1, @1, @1]
outputGoals:@[@0, @0, @0, @0, @0, @0, @1, @0, @0, @0]];
//8
[_krBPN addPatterns:@[@1, @1, @1, @1,
@1, @1, @1, @1,
@1, @1, @0, @0,
@0, @1, @0, @0,
@0, @1, @1, @0,
@0, @0, @1, @0,
@0, @0, @1, @1,
@1, @1, @1, @1,
@1, @1, @1, @1]
outputGoals:@[@0, @0, @0, @0, @0, @0, @0, @1, @0, @0]];
//9
[_krBPN addPatterns:@[@1, @1, @1, @1,
@1, @0, @0, @0,
@0, @1, @0, @0,
@0, @1, @0, @0,
@0, @0, @1, @0,
@0, @0, @1, @0,
@0, @0, @0, @1,
@1, @1, @1, @1,
@1, @1, @1, @1]
outputGoals:@[@0, @0, @0, @0, @0, @0, @0, @0, @1, @0]];
//0
[_krBPN addPatterns:@[@1, @1, @1, @1,
@1, @1, @1, @1,
@1, @1, @0, @0,
@0, @0, @0, @0,
@0, @1, @1, @0,
@0, @0, @0, @0,
@0, @0, @1, @1,
@1, @1, @1, @1,
@1, @1, @1, @1]
outputGoals:@[@0, @0, @0, @0, @0, @0, @0, @0, @0, @1]];
__block typeof(_krBPN) _weakKrBPN = _krBPN;
//訓練完成時( Training complete )
[_krBPN setTrainingCompletion:^(BOOL success, NSDictionary *trainedInfo, NSInteger totalTimes){
if( success )
{
NSLog(@"Training done with total times : %i", totalTimes);
NSLog(@"TrainedInfo 1 : %@", trainedInfo);
//Start in checking the network is correctly trained.
NSLog(@"======== Start in Verification ========");
[_weakKrBPN setTrainingCompletion:^(BOOL success, NSDictionary *trainedInfo, NSInteger totalTimes){
NSLog(@"Training done with total times : %i", totalTimes);
NSLog(@"TrainedInfo 2 : %@", trainedInfo);
}];
[_weakKrBPN recoverNetwork];
//Verified number " 7 ", and it has some defects.
[_weakKrBPN directOutputAtInputs:@[@1, @1, @1, @0,
@0, @0, @0, @0,
@0, @1, @0, @0,
@0, @0, @0, @0,
@0, @0, @1, @0,
@0, @0, @0, @0,
@0, @0, @0, @1,
@1, @1, @1, @1,
@1, @1, @1, @1]];
}
}];
[_krBPN trainingByRandomSettings];
//[_krBPN trainingRandomAndSave];
}
//Medical prediction of cancer, 醫療癌症預測
-(void)useSample4
{
_krBPN.activeFunction = KRBPNActivationBySigmoid;
// ER, PR, Hev-2, K, 治療與否, 每 10 年為 0.1 單位 => 存活率
[_krBPN addPatterns:@[@0.9, @0.9, @0.0, @0.15, @0, @0.1f] outputGoals:@[@0.12f]];
[_krBPN addPatterns:@[@0.9, @0.9, @0.0, @0.15, @1, @0.1f] outputGoals:@[@0.12f]];
[_krBPN addPatterns:@[@0.5, @0.4, @0.0, @0.25, @0, @0.1f] outputGoals:@[@0.23f]];
[_krBPN addPatterns:@[@0.5, @0.4, @0.0, @0.25, @1, @0.1f] outputGoals:@[@0.14f]];
[_krBPN addPatterns:@[@0.0, @0.0, @0.0, @0.3, @0, @0.1f] outputGoals:@[@0.35f]];
[_krBPN addPatterns:@[@0.0, @0.0, @0.0, @0.3, @1, @0.1f] outputGoals:@[@0.22f]];
[_krBPN addPatterns:@[@0.0, @0.0, @1.0, @0.3, @0, @0.1f] outputGoals:@[@0.3f]];
[_krBPN addPatterns:@[@0.0, @0.0, @1.0, @0.3, @1, @0.1f] outputGoals:@[@0.12f]];
__block typeof(_krBPN) _weakKrBPN = _krBPN;
//訓練完成時( Training complete )
[_krBPN setTrainingCompletion:^(BOOL success, NSDictionary *trainedInfo, NSInteger totalTimes){
if( success )
{
NSLog(@"Training done with total times : %i", totalTimes);
NSLog(@"TrainedInfo 1 : %@", trainedInfo);
//Start in checking the network is correctly trained.
NSLog(@"======== Start in Verification ========");
[_weakKrBPN setTrainingCompletion:^(BOOL success, NSDictionary *trainedInfo, NSInteger totalTimes){
NSLog(@"Training done with total times : %i", totalTimes);
NSLog(@"TrainedInfo 2 : %@", trainedInfo);
}];
[_weakKrBPN recoverNetwork];
// To predict how many years could be alive with the patients.
[_weakKrBPN directOutputAtInputs:@[@0.6, @0.6, @0.0, @0.25, @0.0, @0.1f]];
[_weakKrBPN directOutputAtInputs:@[@0.6, @0.6, @0.0, @0.25, @1.0, @0.1f]];
}
}];
[_krBPN trainingByRandomSettings];
}
- (void)viewDidLoad
{
[super viewDidLoad];
_krBPN = [KRBPN sharedNetwork];
//_krBPN.delegate = self;
// Learning rate
_krBPN.learningRate = 0.8f;
// Convergence error, 收斂誤差值 ( Normally is 10^-3 or 10^-6 )
_krBPN.convergenceError = 0.001f;
// To limit that iterations
_krBPN.limitIteration = 1000;
// If you wanna use enhanced theory, and to setup that customized learning rate, you could use this as below :
//_krBPN.quickPropFixedRate = 0.5f;
// If you wanna directly use QuickProp method to enhance that network training :
//_krBPN.learningMode = KRBPNLearningModeByQuickPropDynamic;
// If you wanna use Kecman's theory to enhance that network training, please use this :
//_krBPN.learningMode = KRBPNLearningModeByQuickPropFixed;
// If you wanna use Fahlman's QuickProp combined Kecman's Fixed Rate. Yes, in my experiences this usage method is better than others.
_krBPN.learningMode = KRBPNLearningModeByQuickPropSmartHybrid;
// 每一次的迭代( Per iteration-training )
[_krBPN setEachIteration:^(NSInteger times, NSDictionary *trainedInfo){
NSLog(@"Iterations : %i", times);
NSLog(@"Result per Iteration : %@", [trainedInfo objectForKey:KRBPNTrainedOutputResults]);
}];
// Setup anything by yourself, and 2 outputs.
//[self useSample1];
// Only setup patterns and output goals, and 1 output.
//[self useSample2];
// Only setup patterns and output goals, then learning to identify numbers 0 to 9.
[self useSample3];
// Medical prediction of cancer
//[self useSample4];
}V2.0.2
MIT.