Example for learning/predicting multiclass on own data

[tklein23]

This task is about creating examples in any of the available interfaces to train a multiclass SVM and to predict multiclass labels using the learned SVM:

$ python train_multiclass_svm.py train.data train.labels multiclass-svm.model
$ python predict_multiclass_svm.py multiclass-svm.model eval.data predicted.labels 

The example should be as simple as possible. The goals are:

• applying algorithms to own data without a need to change the scripts/sources
• providing a simple examples for new developers to start own scripts/experiments

Additionally, we could try to provide a script, that evaluates the outcome of the above scripts with. For example:

$ python evaluate_multiclass_labels.py eval.label predicted.labels 

Disclaimer: This task could also be solved with other interfaces than python and for other algorithms that multilabel SVMs.

[PirosB3]

I would love to do this task, can I start?
This would be also good experience for me because it somewhat regards my SoC proposal.

Thanks

[PirosB3]

Hi @tklein23
What type of format would you like the *.data and *.labels? I was thinking it would be a good idea to add something standard like svmlight. In this way, we reduce the files from 2 to 1 (svmlight contains both target and features) and provide a tool that is compatible with standard formats.

.=. : : ... : #

What do you think?

Thanks,
Dan

[tklein23]

Hey Dan! Feel free to take this task!

I totally agree that we should use standard formates. I think we can go with SVMlighs format, like shown in data/toy/7class_example4_train.light. No need to split it into two files.

The goal is simply to have something that can be applied easily to own data (without touching source). Btw., you're not limited to Python - feel free to do it in C++ if you like to.

[PirosB3]

@tklein23 I have started working on this new feature. When you have a second can you see my initial commit? it's no where near to a production version, but if there is something I am doing wrong please let me know.

Also, wouldn't it be better to have a single file that does training+evaluation? (like you said: evaluate_multiclass_labels.py, but without having the other two scripts) there is a lot of reusable functionality between the two.

Let me know,
Dan

[vigsterkr]

@PirosB3 we would like you to send a PR (pull request) instead of asking people to check on your forked repository... it is essential that you start working with PRs as that's how we do development during the whole GSoC.

even if your code is not ready it's ok to send a PR as we'll discuss things in that PR and then you can change and add more commits to the PR obviously...

[PirosB3]

Perfect!
I will send a PR

2014-03-07 13:31 GMT+00:00 Viktor Gal notifications@github.com:

@PirosB3 https://github.com/PirosB3 we would like you to send a PR
(pull request) instead of asking people to check on your forked
repository... it is essential that you start working with PRs as that's how
we do development during the whole GSoC.

even if your code is not ready it's ok to send a PR as we'll discuss
things in that PR and then you can change and add more commits to the PR
obviously...

Reply to this email directly or view it on GitHubhttps://github.com//issues/1868#issuecomment-37023878
.

---

PirosB3

https://github.com/PirosB3 http://pirosb3.com

[tklein23]

You wrote: Also, wouldn't it be better to have a single file that does training+evaluation? (like you said: evaluate_multiclass_labels.py, but without having the other two scripts) there is a lot of reusable functionality between the two.

I think it's better to have individual scripts for training, predicting and evaluation. Evaluation for example is not limited to a specific learning algorithm; you can use it to evaluate everything that outputs multiclass-labels.

Anyway, if you see reusable code, try to use methods/includes/whatever-codeblocks for it.

[karlnapf]

As for format, keep in mind we can serialise objects in Shogun