"SVM set failed" exception

[javiermaldonadoc]

Hello,
I'm using a I'm using the MultiClassLibSVM with 5 class dataset. But in the end run throw an exception "SVM set failed". Also, the process is very slow, about two hours to process a training dataset of 125K rows and 41 columns (NSL-KDD dataset).
As additional information: I'm tried the same dataset with random forest and CART with no problem.

This is the function with the corresponding output.
SVM-output.txt

[karlnapf]

Hi Shogun stores data in column major format. This mean your data is interpreted as 41 vectors of dimension 125k, which probably causes problems. Try transposing.... Best H

On Thu, 14 May 2020 at 21:47, javiermaldonadoc ***@***.***> wrote: Hello, I'm using a I'm using the MultiClassLibSVM with 5 class dataset. But in the end run throw an exception "SVM set failed". Also, the process is very slow, about two hours to process a training dataset of 125K rows and 41 columns (NSL-KDD dataset). As additional information: I'm tried the same dataset with random forest and CART with no problem. This is the function with the corresponding output. SVM-output.txt <https://github.com/shogun-toolbox/shogun/files/4630785/SVM-output.txt> — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#5034>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAFKVP6STJZOZIPVYT6OIFLRRRKFRANCNFSM4NBAR4UA> .

-- Sent from my phone

[javiermaldonadoc]

Thanks for your response,

I've made the transposition and give me this error:
Created. Training...[05/14/20 17:31:27 info] 125973 trainlabels, 5 classes
[05/14/20 17:31:27 error] Number of training vectors does not match number of labels
terminate called after throwing an instance of 'shogun::ShogunException'
what(): Number of training vectors does not match number of labels

And when i try to run random forest and CART with this transposition, give me this error:
terminate called after throwing an instance of 'fmt::v6::format_error'
what(): argument index out of range
Aborted (core dumped)

Before transposing, random forest and CART, works...
Any idea? Thanks!

Hi Shogun stores data in column major format. This mean your data is interpreted as 41 vectors of dimension 125k, which probably causes problems. Try transposing.... Best H
On Thu, 14 May 2020 at 21:47, javiermaldonadoc @.***> wrote: Hello, I'm using a I'm using the MultiClassLibSVM with 5 class dataset. But in the end run throw an exception "SVM set failed". Also, the process is very slow, about two hours to process a training dataset of 125K rows and 41 columns (NSL-KDD dataset). As additional information: I'm tried the same dataset with random forest and CART with no problem. This is the function with the corresponding output. SVM-output.txt https://github.com/shogun-toolbox/shogun/files/4630785/SVM-output.txt — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#5034>, or unsubscribe https://github.com/notifications/unsubscribe-auth/AAFKVP6STJZOZIPVYT6OIFLRRRKFRANCNFSM4NBAR4UA .
-- Sent from my phone

[javiermaldonadoc]

By executing
features_train = create(f_feats_train);
cout<<"Features Train: "<<features_train->get_num_vectors()<<endl;
features_test = create(f_feats_test);
cout<<"Features Test: "<<features_test->get_num_vectors()<<endl;

I got these results:
Transposing:
Features Train: 41
Features Test: 41

Without transposing:
Features Train: 125973
Features Test: 22543

The last results is consistent with the number of cases (rows) and the first is consistent with the number of features (columns)... This information could help?

Thank you very much!

[karlnapf]

The error message seems to be pretty clear to me?
Number of training vectors does not match number of labels
How many labels do you pass?

[karlnapf]

@gf712 the error he gets on cart seems to have to do with the fmt lib crashing? :D

[gf712]

Yes, seems like there is a formatting error.. @vigsterkr @theartful can we switch on the compile time checks of fmt to avoid this type of error at runtime? I am talking about https://fmt.dev/latest/api.html#format-api

[gf712]

@javiermaldonadoc get_num_vectors gets you the number of examples, not the feature count (from the docs "get number of examples/vectors, possibly corresponding to the current subset")

[ghost]

@gf712 FMT_STRING requires a constexpr string literal. I don't know how to make it work without either using a macro or changing every format string at call site to call FMT_STRING. https://gist.github.com/theartful/01060830399985b72c817f39de8be7ed

[gf712]

@gf712 FMT_STRING requires a constexpr string literal. I don't know how to make it work without either using a macro or changing every format string at call site to call FMT_STRING. https://gist.github.com/theartful/01060830399985b72c817f39de8be7ed

Hmmm that is annoying.. I guess we just have to look for the bug then. Or we go back to using macros to do these checks. @karlnapf @vigsterkr ?

[karlnapf]

no macros please :)
I guess this should be relatively easy to debug. I have the feeling it probably is due to a really long sequence being requested or something

[javiermaldonadoc]

@javiermaldonadoc get_num_vectors gets you the number of examples, not the feature count (from the docs "get number of examples/vectors, possibly corresponding to the current subset")

Yes, that's correct and i'm checking that using these function jointly with transpose function. I have 41 features and 125K of examples in training and 22K in testing and is consistent with the given numbers by using this function.

[javiermaldonadoc]

The error message seems to be pretty clear to me?
Number of training vectors does not match number of labels
How many labels do you pass?

I checked that those numbers are the same 125K in both cases, labels and vectors.

[karlnapf]

strange. Could you post a (preferably minimal standalone with synthetic data) example to reproduce this issue? Maybe there is a problem in the multiclass codes ...

[javiermaldonadoc]

strange. Could you post a (preferably minimal standalone with synthetic data) example to reproduce this issue? Maybe there is a problem in the multiclass codes ...

Thank you!
Here is a link to the used datasets. https://drive.google.com/open?id=1Cz0ju3xfpw70Ve8PJ9AeQcZt_RPR3Ziq
You could use the small one (KDD-Test*), is the smallest of all datasets. An my code https://pastebin.com/HMAS9dEL, here is a full functional example (load data, create features and labels sets, training and testing SVM). As a reference, i'm using the NSL-KDD dataset https://www.unb.ca/cic/datasets/nsl.html

Hope this work for you! Again, thank you very much! I really appreciate that!

[stale]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.

[stale]

This issue is now being closed due to a lack of activity. Feel free to reopen it.