Skip to content

Conversation

@otoalves
Copy link
Contributor

@otoalves otoalves commented Sep 3, 2016

No description provided.

Oto Alves added 30 commits May 12, 2016 14:02
…t of stuff in a silly way, so I deleted/refactored a bunch of code from the previous few days. Most notably, the ConfigurationRanker class was ditched. A lot of what it did, however, was refactored into ConfigurationCollection's static methods.
…y configurations they want in the output. Made some changes to the ranking system. Most notably, the ConfigurationRanker class is back, but looks different from the past version. I just got al of the static ranking methods in ConfigurationCollection and put them into a new class called ConfigurationRanker because it made more sense semantically.
…sometimes not match the N-best-configs feature output and/or smac. Turns out there's a bug on the smac trajectory parsing. I'll refactor that. Making this commit before touching it.
…idation, and ran a bunch of successful tests. Still lacking unit tests, gonna write them.
…orted configurations log to have SMAC's final incumbent as the first one in case of a tie
@larskotthoff
Copy link
Collaborator

Thanks. I'm getting an exception when no configurations were evaluated:

java.lang.IndexOutOfBoundsException: Index: 0, Size: 0
    java.util.ArrayList.rangeCheck(ArrayList.java:653)
    java.util.ArrayList.get(ArrayList.java:429)
    autoweka.ConfigurationCollection.forceFirst(ConfigurationCollection.java:175)
    autoweka.ConfigurationCollection.rank(ConfigurationCollection.java:139)
    weka.classifiers.meta.AutoWEKAClassifier.buildClassifier(AutoWEKAClassifier.java:377)
    weka.gui.explorer.AutoWEKAPanel$11.run(AutoWEKAPanel.java:411)

    at java.util.ArrayList.rangeCheck(ArrayList.java:653)
    at java.util.ArrayList.get(ArrayList.java:429)
    at autoweka.ConfigurationCollection.forceFirst(ConfigurationCollection.java:175)
    at autoweka.ConfigurationCollection.rank(ConfigurationCollection.java:139)
    at weka.classifiers.meta.AutoWEKAClassifier.buildClassifier(AutoWEKAClassifier.java:377)
    at weka.gui.explorer.AutoWEKAPanel$11.run(AutoWEKAPanel.java:411)

and a null pointer exception in another case:

java.lang.NullPointerException
    at autoweka.Ensembler.parseDatasetMetadata(Ensembler.java:80)
    at autoweka.Ensembler.<init>(Ensembler.java:60)
    at weka.classifiers.meta.AutoWEKAClassifier.buildClassifier(AutoWEKAClassifier.java:382)
    at weka.classifiers.evaluation.Evaluation.evaluateModel(Evaluation.java:1527)
    at weka.classifiers.Evaluation.evaluateModel(Evaluation.java:650)
    at weka.classifiers.AbstractClassifier.runClassifier(AbstractClassifier.java:141)
    at weka.classifiers.meta.AutoWEKAClassifier.main(AutoWEKAClassifier.java:203)

Could you have a look please?

@otoalves
Copy link
Contributor Author

otoalves commented Sep 9, 2016

Hi Lars,

Do you know which datasets/seed/constraints caused these exceptions? I might have fixed the issue but it would be safer to run autoweka with the same arguments again to confirm.

@larskotthoff
Copy link
Collaborator

I think I tried this with the yeast dataset from the Auto-WEKA paper. Presumably you tested with all of them to measure the performance improvement with ensembles?

@otoalves
Copy link
Contributor Author

Hi Lars,
The issue when there are no configurations should be fixed now. I hadn't tested with all of them, I had only used a subset which mostly contained the smaller ones, which is probably why I hadn't gotten this exception recently

@larskotthoff
Copy link
Collaborator

Great, thanks! Do you have the performance comparison with and without ensembles as well?

@otoalves
Copy link
Contributor Author

Not yet - I've done a few test runs, but not in a very methodical way. I'm writing a script to run on Parallel over the weekend, so that I can make some pretty graphs with more concrete measures =)
I'm sorry for the delay regarding this, things have been a permanent rush since I arrived here.

@larskotthoff
Copy link
Collaborator

No worries. I've run some small experiments myself and didn't see any improvement with ensembles, so I'm curious to see your experiments.

@larskotthoff
Copy link
Collaborator

Hi Oto, did you have a chance to look at running the experiments? I want to merge this, but would like to see some numbers on how much it improves first.

@otoalves
Copy link
Contributor Author

Hi Lars,
I have them running right now, will be ready in a couple days =)

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants