Java Neural Network Framework Neuroph
Home Page: http://neuroph.sourceforge.net
In order to make Neuroph project easier to use, navigate and maintain this repo has been splitted in two repos which contain Neuroph Framework and Neuroph Studio GUI. These two repos will be active and should be used. This old repo here will remain active for some time, but should not be used for new development and will not include latest additions.
Currently for the Input Layer in Perceptron#createNetwork are used Instances of org.neuroph.core.Neuron (default).
Because the input layer has no Input Connections in Neuron#calculate WeightedSum-InputFunction is generating always 0.
Initialisation of input layerin Perceptron#createNetwork should look like this:
// init neuron settings for input layer
...
inputNeuronProperties.setProperty("neuronType", InputNeuron.class);
....
In version 2.9, getVersion() in org.neuroph.util.Neuroph returns 2.8. serialVersionUID also outdated, leading to this error:
java.io.InvalidClassException: org.neuroph.core.NeuralNetwork; local class incompatible: stream classdesc serialVersionUID = 7, local class serialVersionUID = 6
I am not positive on why constructing a relatively large NN would take so much time.
To construct a small NN, for example
MultiLayerPerceptron neuralNet = new MultiLayerPerceptron(10000, 20);
It takes about 7 seconds to finish this single line of code.
However, if we simply expands the number of input neurons, i.e.
MultiLayerPerceptron neuralNet = new MultiLayerPerceptron(100000, 20);
It takes more than 15 mins to complete construction of this MLP NN. (My computer has a quad-core i5 Skylake processor.)
In the example above, it's about 20_10^5 = 2_10^6 connections. From my perception, it seems to be an "easy" and quick task for the computer. I am not sure why the construction would take so much time.
@neuroph Any comments or help would be greatly appreciated!
hello
i have tried the image recognition using the neuroph studio
all buttons are getting disabled after I loaded the image directory to be recognized and the junk directory
is this a bug in Neuroph studio or am i doing something wrong
When attempting to build the pom.xml in neuroph-2.9 using:
mvn clean package
I am getting quite a few test failures. Are the tests out of date and should I use -DskipTests or am I not using the correct maven commands?
Running org.neuroph.contrib.AppTest
Tests run: 1, Failures: 0, Errors: 0, Skipped: 0, Time elapsed: 0.019 sec
Running org.neuroph.contrib.ConvolutionLayerTest
Tests run: 3, Failures: 0, Errors: 0, Skipped: 0, Time elapsed: 0.019 sec
Running org.neuroph.contrib.PoolingLayerTest
Tests run: 3, Failures: 1, Errors: 0, Skipped: 0, Time elapsed: 0.01 sec <<< FAILURE!
testPoolingTwoLayersWithOneFeatureMap(org.neuroph.contrib.PoolingLayerTest) Time elapsed: 0.01 sec <<< FAILURE!
java.lang.AssertionError: expected:<4> but was:<1>
at org.junit.Assert.fail(Assert.java:88)
at org.junit.Assert.failNotEquals(Assert.java:743)
at org.junit.Assert.assertEquals(Assert.java:118)
at org.junit.Assert.assertEquals(Assert.java:555)
at org.junit.Assert.assertEquals(Assert.java:542)
at org.neuroph.contrib.PoolingLayerTest.testPoolingTwoLayersWithOneFeatureMap(PoolingLayerTest.java:114)
Tests run: 3, Failures: 2, Errors: 0, Skipped: 0, Time elapsed: 0.009 sec <<< FAILURE!
testUpdateNeuronWeightsInOnlineMode(org.neuroph.nnet.learning.LMSTest) Time elapsed: 0.006 sec <<< FAILURE!
junit.framework.AssertionFailedError: expected:<0.09707076026863733> but was:<0.047070760268637324>
at junit.framework.Assert.fail(Assert.java:50)
at junit.framework.Assert.failNotEquals(Assert.java:287)
at junit.framework.Assert.assertEquals(Assert.java:67)
at junit.framework.Assert.assertEquals(Assert.java:74)
at org.neuroph.nnet.learning.LMSTest.testUpdateNeuronWeightsInOnlineMode(LMSTest.java:43)
at sun.reflect.NativeMethodAccessorImpl.invoke0(Native Method)
at sun.reflect.NativeMethodAccessorImpl.invoke(NativeMethodAccessorImpl.java:62)
at sun.reflect.DelegatingMethodAccessorImpl.invoke(DelegatingMethodAccessorImpl.java:43)
at java.lang.reflect.Method.invoke(Method.java:483)
at junit.framework.TestCase.runTest(TestCase.java:168)
at junit.framework.TestCase.runBare(TestCase.java:134)
at junit.framework.TestResult$1.protect(TestResult.java:110)
at junit.framework.TestResult.runProtected(TestResult.java:128)
at junit.framework.TestResult.run(TestResult.java:113)
at junit.framework.TestCase.run(TestCase.java:124)
at junit.framework.TestSuite.runTest(TestSuite.java:243)
at junit.framework.TestSuite.run(TestSuite.java:238)
at org.junit.internal.runners.JUnit38ClassRunner.run(JUnit38ClassRunner.java:83)
at org.apache.maven.surefire.junit4.JUnit4Provider.execute(JUnit4Provider.java:236)
at org.apache.maven.surefire.junit4.JUnit4Provider.executeTestSet(JUnit4Provider.java:134)
at org.apache.maven.surefire.junit4.JUnit4Provider.invoke(JUnit4Provider.java:113)
at sun.reflect.NativeMethodAccessorImpl.invoke0(Native Method)
at sun.reflect.NativeMethodAccessorImpl.invoke(NativeMethodAccessorImpl.java:62)
at sun.reflect.DelegatingMethodAccessorImpl.invoke(DelegatingMethodAccessorImpl.java:43)
at java.lang.reflect.Method.invoke(Method.java:483)
at org.apache.maven.surefire.util.ReflectionUtils.invokeMethodWithArray(ReflectionUtils.java:189)
at org.apache.maven.surefire.booter.ProviderFactory$ProviderProxy.invoke(ProviderFactory.java:165)
at org.apache.maven.surefire.booter.ProviderFactory.invokeProvider(ProviderFactory.java:85)
at org.apache.maven.surefire.booter.ForkedBooter.runSuitesInProcess(ForkedBooter.java:103)
at org.apache.maven.surefire.booter.ForkedBooter.main(ForkedBooter.java:74)
Hi folks,
I just might not have found it, but is there possibility to outsource the learning process to a new thread?
I read something about learnInNewThread(), but tbh I'm not quite sure how it works.
I'd like to learn multiple nets at once, and also poll their current error while learning.
thanks and regards
The Neuroph home page mentions that a port of it is available in C#.
Can you add a link to where we can find the source?
Naming the repository "master" instead of e.g. "neuroph" leads to a directory "master" when cloning the repository. This is unsual, because one won't find the project by name.
Having a sub-project-dir "neuroph-2.9" is unsual too, because this leads to renaming the directory with each release. Instead of using version numbers in directory names, tags or branches would do the versioning stuff a lot better. For just marking releases, tags are optimal. If ever a bugfix release for a particular release is needed, one could create a branch based on the release tag.
[ERROR] 'dependencies.dependency.version' for android:android:jar is missing. @ org.neuroph:neuroph-imgrec:[unknown-version], /Users/kees/workspaces/iwoz/neuroph/neuroph-2.9/ImageRec/pom.xml, line 21, column 19
This dependency in the ImageRec module is causing the trouble:
<dependency>
<groupId>android</groupId>
<artifactId>android</artifactId>
</dependency>
Removing this causes compilation errors in ImageAndroid.java
Changing the depency to the following makes the project compile again:
<dependency>
<groupId>com.google.android</groupId>
<artifactId>android</artifactId>
<version>1.5_r4</version>
</dependency>
Neuroph does not seem to be present at maven central. Why? It would be much easier to use it in projects.
To have all required dependencies in the libs folder of the neuroph zip distribution, logback-core-1.0.13.jar should be added. Otherwise the examples from the pdf docs do not run.
Although eval package was included in neuroph-core source code 10 months ago, it's missing in http://neuroph.sourceforge.net/maven2/org/neuroph/neuroph-core/2.94/neuroph-core-2.94.jar
public double getDerivative(double net) {
return 1d * (1d - output);
}
Shouldn't this be rather output * (1d - output); ?
With even small size data set (2000 entries with 3 input parameters and 2 output parameters),
and a very minimal size NN (3 input, 5 intermediate, 2 output neurons)
The training hanged in ubuntu as well as windows version.
Netbeans errors keep on cropping like 'GC overhead exceeded' or 'too much data on clipboard'.
I understand it has not much to do with neuroph, but with netbeans platform.
Any fix ?
I have trained a network using RGBImageRecognitionTrainingSample. When I try to use the network with ImageRecognitionSample the code fails here with a NullPointerException:
ImageRecognitionPlugin imageRecognitionPlugin =
(ImageRecognitionPlugin) neuralNetwork.getPlugin(ImageRecognitionPlugin.class);
imageRecognitionPlugin is null although in my debugger I see that the plugins collection in the network is filled.
If you create a layer, add some neurons too it (or initialize it with a set number of neurons, as I am) and then remove some of them, the layer's getNeurons() array leaves trailing nulls. This is because the internal array isn't shrunk down to the size of the list. It simply doesn't use that section of the array. The problem comes when anything uses the Layer.getNeurons() method, and naively assumes that the elements will be non-null.
Here's the proper implementation and the flawed one, as they appear side by side in the code:
public Object[] toArray() {
return Arrays.copyOf(elementData, size);
}
public final E[] asArray() {
return elementData;
}Note that the proper implementation only copies size elements, whereas the flawed one just throws the entire array at you.
Greetings,
I'm trying to create an Android application that uses neural network for image recognition.
I've encountered many problems with various Neuroph versions while trying to implement it in my application.
Problems mostly revolve around java.awt classes because they are not available in Android framework.
In the neuroph-core-2.92.jar file I've seen some commented lines with methods for usage in Android framework and would like to know when it is going to be fully implemented and how could one help to implement it faster.
If it's not going to be implemented soon, what can I do to temporarily convince core classes to use Android framework's classes?
The Android app example on your page works well out of the box, but for some reason, I cannot replicate the procedure of opening a neural network from resources.
Thank you for your work, cheers
Marko
1.) Broken references to wrapped jars.
2.) After fixing build fails:
The module /home/.../NeurophStudio/build/cluster/modules/org-jzy3d.jar has no public packages and so cannot be compiled against
BUILD FAILED (total time: 22 seconds)
"learn" or "calculate" returns this error after using any remove connection function:
java.lang.NullPointerException
at org.neuroph.core.input.WeightedSum.getOutput(WeightedSum.java:34)
at org.neuroph.core.Neuron.calculate(Neuron.java:145)
at org.neuroph.core.Layer.calculate(Layer.java:259)
at org.neuroph.core.NeuralNetwork.calculate(NeuralNetwork.java:313)
Please,
Separate IDE and library sources into two repositories.
It will help to download source of library.
We want to make use of ur library in one of our systems.
The issue we came across is that, our architecture assumes that the learners are leaned in online fashion (not batch, i.e. the classifier seems an instances, learns it, moves to the next instance. That said we can still iterate over the data multiple times).
I was checking your examples for multi-layer perceptron, which accepts data as batch:
// create multi layer perceptron
MultiLayerPerceptron myMlPerceptron = new MultiLayerPerceptron(TransferFunctionType.TANH, 2, 3, 1);
// learn the training set
myMlPerceptron.learn(trainingSet);However I know somewhere (?) internally somewhere you learn the data instance by instance. Anyone can help me find it?
Essentially I am looking for something like this:
// create multi layer perceptron
MultiLayerPerceptron myMlPerceptron = new MultiLayerPerceptron(TransferFunctionType.TANH, 2, 3, 1);
// made up code:
for (someInstance : trainingSet) {
// learn the training set
myMlPerceptron.learnOneInstance(someInstance);
}I'm specifically wondering if multithreading and deep network implementations are still panned.
Because the Javadoc does not list NeuralNetwork.createFromFile() as throwing any exceptions and the exceptions that are thrown are RuntimeExceptions, it becomes difficult to know when to catch exceptions. For example, it may be assumed that running NeuralNetwork.createFromFile() on a nonexistent file will not produce an exception since the compiler does not ask the developer to catch any exceptions. This makes debugging difficult when trying to catch all instances of NeurophException since they are never documented or caught by the compiler.
I'm using the DecimalScaleNormalizer class for normalize my dataset, the inputs are normalize correctly but the output not.
Debbuging the normalize method I see that the class finds the correct maxOut value but in the method findScaleVectors():
this.scaleFactorOut = new double[this.maxOut.length];
for(i = 0; i < this.scaleFactorOut.length; ++i) {
this.scaleFactorOut[i] = 1.0D;
}
for(i = 0; i < this.maxOut.length; ++i) {
while(this.maxIn[i] > 1.0D) {
this.maxOut[i] /= 10.0D;
this.scaleFactorOut[i] *= 10.0D;
}
}
the value this.scaleFactorOut is 0 and then the scaleFactorOut is 1 and not 10.000 how is expect
The problem is the while, cause it's asking for the maxIn vector instead of maxOut vector
The current implementation does not converge when training logic OR function with perceptron learning. Below is the reason I found.
The reason to update threshold is to avoid having a bias neuron according to http://stackoverflow.com/questions/6554792/whats-the-point-of-the-threshold-in-a-perceptron
x1.w1 + x2.w2 = threshold
to
x1.w1 + x2.w2 - 1.threshold = 0
where bias is -threshold. So the way to update threshold should be
threshold = threshold + this.learningRate * neuronError;
Some dependencies (e.g. android:android) cannot be resolved via maven central. And there is no documentation where to get them from. This way it is very difficult to build neuroph via maven.
Furthermore at least some tests fail if executed via maven, e.g. GaussianTest.
The NeuralNetworkEventListener interface of org.neuroph.core.events package used FunctionalInterface annotation, which is included in jdk1.8. So the project should be compiled with jdk1.8.
But inside POM file, which was indicated that
<maven.compiler.source>1.7</maven.compiler.source>
<maven.compiler.target>1.7</maven.compiler.target>this should be solved, and I suggest that, don't use FunctionalInterface to keep compatible with jdk1.7
Im trying to create an app using this library. Is it possible for me to implement this library with android studio?
The dimensions for the weights array are interchanged in the initWeights(double min, double max) method:
public void initWeights(double min, double max) {
weights = new Weight[height][width];
for (int i = 0; i < height; i++) {
for (int j = 0; j < width; j++) {
Weight weight = new Weight();
weight.randomize(min, max);
weights[i][j] = weight;
}
}
}
so only square shaped kernel would work. Otherwise an ArrayIndexOutOfBoundsException will be thrown.
A fixed Version could be:
public void initWeights(double min, double max) {
weights = new Weight[width][height];
for (int i = 0; i < height; i++) {
for (int j = 0; j < width; j++) {
Weight weight = new Weight();
weight.randomize(min, max);
weights[j][i] = weight;
}
}
}
import org.neuroph.util.NeurophArrayList; fails. I am not sure if NeurophArrayList is deprecated or it is a new addition? It is breaking everything in core module.
When adding neuroph as a maven dependency, Logback dependencies and the logback.xml config file are included. This can cause, as happened to me, the neuroph logback.xml to overwrite my own logback.xml
For libraries, you should only include slf4j as a dependency.
Nice library, btw!
Method NeuralNetwork#setOutputNeurons doesn't set Neurons - it adds new Neurons. That is a difference ;)
P. S.
If you want to add something outputNeurons.addAll it this case would do.
When I clone the repo and run mvn package, the tanh getDerivative tests fail.
Failed tests: testGetOutput0: expected:<0.0249947929> but was:<0.042888565254513115>
testGetDerivative0: expected:<0.940014848> but was:<0.8233849509308319>
testGetOutput1: expected:<0.0499583> but was:<0.08572357559022631>
testGetDerivative1: expected:<0.786447> but was:<0.37123532456703945>
testGetOutput2: expected:<0.07485969> but was:<0.12845174325067968>
testGetDerivative2: expected:<0.596585> but was:<-0.18777757241307147>
testGetOutput3: expected:<0.09966799> but was:<0.17102031197696171>
testGetDerivative3: expected:<0.419974> but was:<-0.707776976114507>
testGetOutput4: expected:<0.124353> but was:<0.2133773157398819>
testGetDerivative4: expected:<0.280414> but was:<-1.1186837273905166>
testGetOutput5: expected:<0.148885> but was:<0.2554718291942514>
testGetDerivative5: expected:<0.1807066> but was:<-1.4122559381650563>
testGetOutput6: expected:<0.17323515> but was:<0.2972542073284932>
testGetDerivative6: expected:<0.113812> but was:<-1.609214399230635>
testGetOutput7: expected:<0.1973753> but was:<0.3386763119739128>
testGetDerivative7: expected:<0.07065> but was:<-1.7362946813477618>
testGetOutput8: expected:<0.22127846> but was:<0.3796917230669401>
testGetDerivative8: expected:<0.043464> but was:<-1.8163384925297978>
Tests run: 272, Failures: 18, Errors: 0, Skipped: 0
In possibly related news, when I skip the tests (and have jury rigged a few things to compile), the TestTimeSeries training never converges. I'm not sure if it's errors in other parts of the code, errors in the changes I made, or errors in Tanh.java. I think the current implementation of getDerivative is only accurate for amplitude=1 and slope=2, though, according to some equations plugged in to Wolfram Alpha.
I started to focus on your project in 2012 , I can visite your website on sourceforge on that time . But five years later , When I start to research AI with Google'S TensorFlow . Your site came to my mind . It's a good choice some how.
But I can't visit http://neuroph.sourceforge.net/ . (This is because Chinese Great Wall FireWall forbid outer internet).
I will appreciate it if you make website on github .
React
A declarative, efficient, and flexible JavaScript library for building user interfaces.
🖖 Vue.js is a progressive, incrementally-adoptable JavaScript framework for building UI on the web.
Typescript
TypeScript is a superset of JavaScript that compiles to clean JavaScript output.
An Open Source Machine Learning Framework for Everyone
Django
The Web framework for perfectionists with deadlines.
Laravel
A PHP framework for web artisans
Bring data to life with SVG, Canvas and HTML. 📊📈🎉
JavaScript (JS) is a lightweight interpreted programming language with first-class functions.
Some thing interesting about web. New door for the world.
A server is a program made to process requests and deliver data to clients.
Machine learning is a way of modeling and interpreting data that allows a piece of software to respond intelligently.
Some thing interesting about visualization, use data art
Some thing interesting about game, make everyone happy.
Facebook
We are working to build community through open source technology. NB: members must have two-factor auth.
Microsoft
Open source projects and samples from Microsoft.
Google
Google ❤️ Open Source for everyone.
Alibaba
Alibaba Open Source for everyone
D3
Data-Driven Documents codes.
Tencent
China tencent open source team.