patrickzib / sfa Goto Github PK

HI,

The code in the README seems to be different than how the usage is in the test cases.
For example:

Score score = weasel.fit(trainSamples);
Score should be Classifier.Score

Is this correct? If so, should the README be updated?

After training, wouldn't it be a good idea to save a model so we could load it later?

Idealy, we need to add a static load() and save() methods to Classifier so it would work with all classifiers implemented in the project.

I will be happy to help.

Hello patrick, thanks for your effort,
I am adding the issue here just not to forget, if possible to enable back seach by radius on SFATrie instead of just searching for the K nearest neighbors.
Cheers!

Dear authors,

First of all, congratulation for your work!

Second, is there a version of TEASER algorithm that can be used in python?

Thank you and best regards,

Andrea

Great Library. Is there a python version?

I have my data in csv format.
How do I get the code to read in the csv files?

The only way I have been able to get it to work so far is by doing the following:

open CBF in notepad, then paste in my data, then save
I did this for both train and test

can you classify categorical data time series
per
Classification and Accuracy: WEASEL and the BOSS ensemble classifier offer state-of-art classification accuracy [2], [3], [4].

I am getting errors on about 15 data sets out of 85 data sets of UCR archive.
I run in ItelliJ IDEA with -Xmx2048m
For example:

Done reading from /home/p/augment/UCR_TS_Archive_2015/Phoneme/Phoneme_TEST samples 1896 queryLength 1024
Done reading from /home/p/augment/UCR_TS_Archive_2015/Phoneme/Phoneme_TRAIN samples 214 queryLength 1024
Exception in thread "main" com.carrotsearch.hppc.BufferAllocationException: Not enough memory to allocate buffers for rehashing: 65,536 -> 131,072
at com.carrotsearch.hppc.IntIntHashMap.allocateBuffers(IntIntHashMap.java:1144)
at com.carrotsearch.hppc.IntIntHashMap.allocateThenInsertThenRehash(IntIntHashMap.java:1169)
at com.carrotsearch.hppc.IntIntHashMap.put(IntIntHashMap.java:177)
at com.carrotsearch.hppc.IntIntHashMap.putOrAdd(IntIntHashMap.java:256)
at sfa.transformation.WEASEL.createBagOfPatterns(WEASEL.java:164)
at sfa.classification.WEASELClassifier.predict(WEASELClassifier.java:130)
at sfa.classification.WEASELClassifier.score(WEASELClassifier.java:124)
at sfa.classification.WEASELClassifier.eval(WEASELClassifier.java:95)
at sfa.MyUCRmain.main(MyUCRmain.java:61)
Caused by: java.lang.OutOfMemoryError: Java heap space

I learned Java many years ago, without using the tools that this project uses (Gradle, Maven). It would be great to have instructions on how to build the project, including IDE preference, and what software versions are needed (JDK 10? Gradle version >= 4.8)? Some things that come to mind include:

• Installing Gradle (there are multiple ways)
• How to import the project
• Gradle configuration

In the MFT class, the fft object is initialised with the windowSize when the MFT instance is initialised, however when the array size is set later:

int arraySize = Math.max(l+this.startOffset, this.windowSize);

and when this is used in line 107:

mftData = Arrays.copyOf(timeSeries.getData(), arraySize);
this.fft.realForward(mftData);
mftData[1] = 0;

if arraySize is larger than the windowSize, then it leaves some of the mftData not transformed and is then used later on in the rest of the MFT and in the classification.

I saw this happen at windowSize = 5, l = 4 and so arraySize = 6

This has an implication on the creation of words especially at small window sizes, however I am not sure how often this occurs or how much it affects the final results

Hallo

Thank you for your fascinating work!
Which structure have the datasets in the datasets folder? They don't look like original UCI datasets...
What are the columns?

For example, DigitShapeRandom:
1 1 1 0.3421972205305417 -1.594004942648406
1 2 1 0.3490627644881473 -1.4250704156116172
1 3 1 0.3353316765729366 -1.2647991976536381
1 4 1 0.3559283084457524 -1.0915330160774444
1 5 1 0.3559283084457524 -0.9225984890406556
1 6 1 0.35249553646694987 -0.7709905801614865
1 7 1 0.3490627644881473 -0.5847294349670782
1 8 1 0.37309216833976566 -0.4634431078637429
1 9 1 0.37309216833976566 -0.35515174437862146

First column seems to be the class. And the rest?

Thank you very much for your time.

Regards
Ilja

The getEuclidean Distance gives incorrect results.
It should depends whether the timeseries is normalized or not and whether the query is normalized or not.
In case of normalization, both should be de-normalized (value * std + avg).
The current implementation normalize again !

As well, for the epsilon radius search, std needs to be taken into account
Cheers!
Assaad

MTSClassificationTest fails line 443 because offset is out of bounds.
This is caused by changes in commit 97aee8e.

Reverting just this loop to the previous code works.

In tests cases, no assertions are used. Only printing is insufficient to prove the test case is working properly. A good example on using assertion is here. Have a look at assertEquals and assertTrue.

Hi Patrick:

Great effort has been put into this library later. Such a great work. Do you have any plans for a v0.1 version release?

We should also aim at uploading that release's artifact to the Maven's central repository. This will make it easier for new comers to use it in their projects and give the library an industrial legitmacy.

Whilst refactoring ParallelFor is all static and when method shutdown() is called, there is no way of reinitialising it back.

This makes me wonder why it's all static and why it has a shutdown() method.

Hi Patrick,

I recently read your paper on WEASEL+MUSE, it's a very good work.

I'm now trying to run WEASEL+MUSE, on UCR multivariate datasets (30 datasets), I think it is not possible to run on ts formatted dataset. Is there a way to run the code from txt files? Am I missing something here?

When working with WEASEL, I noticed save() method generates a seriously large file to represent the classifier model; approximately 300 MB model for a moderately small dataset. I'm not sure exactly why but my first guess it's saving training data as it's linked somewhere in the dependencies of the objects.

Do you have any idea?

Shall we have a contributors.md file? Something similar to this?

Forgive the title, it sounds like title of a spam email :)

http://timeseriesclassification.com/ has a nice board comparing time series classification algorithms. Why don't we see WEASEL in there?

Nice work. In order to get this code standardised, why not using Maven? It's the most acceptable packaging, build and automation tool for Java. It will also help with wiring unit test cases for the project.

I'll be happy to help ;)

I am not able to load any dataset (any type arff or csv). For example: TwoLeadECG dataset I got this exception.

Exception in thread "main" java.lang.NumberFormatException: For input string: "-0.04372,0.003764,0.039377,0.1106,0.18183,0.30054,0.24118,0.20557,0.16996,0.074989,0.063118,0.015635,-0.031849,-0.067461,-0.079332,-0.091203,-0.15056,-0.22178,-0.24552,-0.2574,-0.2574,-0.24552,-0.23365,-0.24552,-0.28114,-0.2574,-0.2574,0.039377,0.098731,-0.70849,-1.7769,-2.3585,-2.5366,-2.5841,-2.4298,-2.0974,-1.9905,-1.9193,-1.8125,-1.765,-1.5869,-1.2545,-0.87468,-0.57791,-0.3761,-0.15056,0.18183,0.46673,0.57357,0.66853,0.69227,0.69227,0.73976,0.7635,0.81098,0.8466,0.88221,0.95343,1.0009,1.0484,1.1078,1.1671,1.2383,1.2739,1.2858,1.3333,1.357,1.3333,1.2739,1.1909,1.0603,0.95343,0.81098,0.69227,0.57357,0.44299,0.33615,0.21744,0.14621,0.08686,0.027506,-0.008107,1" at sun.misc.FloatingDecimal.readJavaFormatString(Unknown Source) at sun.misc.FloatingDecimal.parseDouble(Unknown Source) at java.lang.Double.parseDouble(Unknown Source) at java.lang.Double.valueOf(Unknown Source) at sfa.timeseries.TimeSeriesLoader.loadDataset(TimeSeriesLoader.java:53)

Am I missing something or there is a bug in the loader class?
Thanks

When running WEASEL on time series data whose lengths vary, the following error occurs (not always, but sometimes).

java.lang.ArrayIndexOutOfBoundsException: 47
	at sfa.timeseries.TimeSeries.calcIncrementalMeanStddev(TimeSeries.java:237)
	at sfa.transformation.MFT.transformWindowing(MFT.java:115)
	at sfa.transformation.SFA.transformWindowing(SFA.java:292)
	at sfa.transformation.SFA.transformWindowingInt(SFA.java:327)
	at sfa.transformation.WEASEL.createWords(WEASEL.java:122)
	at sfa.transformation.WEASEL.access$000(WEASEL.java:21)
	at sfa.transformation.WEASEL$1.run(WEASEL.java:96)
	at sfa.classification.ParallelFor$1.run(ParallelFor.java:31)
	at java.util.concurrent.Executors$RunnableAdapter.call(Executors.java:511)
	at java.util.concurrent.FutureTask.run(FutureTask.java:266)
	at java.util.concurrent.ThreadPoolExecutor.runWorker(ThreadPoolExecutor.java:1142)
	at java.util.concurrent.ThreadPoolExecutor$Worker.run(ThreadPoolExecutor.java:617)
	at java.lang.Thread.run(Thread.java:745)
java.lang.ArrayIndexOutOfBoundsException: 47
	at sfa.timeseries.TimeSeries.calcIncrementalMeanStddev(TimeSeries.java:237)
	at sfa.transformation.MFT.transformWindowing(MFT.java:115)
	at sfa.transformation.SFA.transformWindowing(SFA.java:292)
	at sfa.transformation.SFA.transformWindowingInt(SFA.java:327)
	at sfa.transformation.WEASEL.createWords(WEASEL.java:122)
	at sfa.transformation.WEASEL.access$000(WEASEL.java:21)
	at sfa.transformation.WEASEL$1.run(WEASEL.java:96)
	at sfa.classification.ParallelFor$1.run(ParallelFor.java:31)
	at java.util.concurrent.Executors$RunnableAdapter.call(Executors.java:511)
	at java.util.concurrent.FutureTask.run(FutureTask.java:266)
	at java.util.concurrent.ThreadPoolExecutor.runWorker(ThreadPoolExecutor.java:1142)
	at java.util.concurrent.ThreadPoolExecutor$Worker.run(ThreadPoolExecutor.java:617)
	at java.lang.Thread.run(Thread.java:745)
Exception in thread "main" java.lang.NullPointerException
	at sfa.transformation.WEASEL.createBagOfPatterns(WEASEL.java:151)
	at sfa.classification.WEASELClassifier.predict(WEASELClassifier.java:125)
	at xxxx.execution.WEASELClassifierCrossValidator.main(WEASELClassifierCrossValidator.java:45)

The cause seems to be windowSize in MFT class, which is set from the length of first sample of TimeSeries [] at the line 354 in SFA class.

      if (this.transformation == null) {
        this.transformation = new MFT(samples[0].getLength(), normMean, this.lowerBounding, this.mftUseMaxOrMin);
      }

It should not assume that the length of first sample is the representative, should it?

It feels like it's related #17. I encountered the following two errors (different data sets).
There must be an inconsistency between number of elements of words expected and actually created.

java.lang.NullPointerException
	at sfa.transformation.WEASEL.createBagOfPatterns(WEASEL.java:147)
	at sfa.classification.WEASELClassifier.fitWeasel(WEASELClassifier.java:162)
	at sfa.classification.WEASELClassifier.fit(WEASELClassifier.java:110)
	at xxxxx.execution.WEASELClassifierCrossValidator.main(WEASELClassifierCrossValidator.java:40)
Exception in thread "main" java.lang.NullPointerException
	at sfa.classification.WEASELClassifier.fit(WEASELClassifier.java:113)
	at xxxxx.execution.WEASELClassifierCrossValidator.main(WEASELClassifierCrossValidator.java:40)

java.lang.NullPointerException
	at sfa.transformation.WEASEL.createBagOfPatterns(WEASEL.java:151)
	at sfa.classification.WEASELClassifier.fitWeasel(WEASELClassifier.java:162)
	at sfa.classification.WEASELClassifier.fit(WEASELClassifier.java:110)
	at xxxxx.execution.WEASELClassifierCrossValidator.main(WEASELClassifierCrossValidator.java:40)
Exception in thread "main" java.lang.NullPointerException
	at sfa.classification.WEASELClassifier.fit(WEASELClassifier.java:113)
	at xxxxx.execution.WEASELClassifierCrossValidator.main(WEASELClassifierCrossValidator.java:40)

For example, fitTransform parameters are not super clear what is expected as input
(sorry there is no time to read the entire paper before using the code)

do you know python code for multivariate time series

How many examples should we use for training if using our own data?

Do you have any suggestions? 1000 examples per class?

What would be small? 50 examples per class? Would the algorithms perform worse in this case?

What is the meaning of the output:

CBF;WEASEL;0.967;0.998

Specifically the last two numbers. Are they the train and then test accuracy?

Is there a way to get the data for the histograms?

• For example WEASEL histogram has weights and words.

  • How do we get this data to produce the histograms?