SL integration

[danyaljj]

Same as #341 and #247 Closes #308

[kordjamshidi]

ok, let me first update this with master! And refresh my mind with the info of two month ago!

[danyaljj]

Conversation moved here from #385

Interfacing the data structures of LBJava and SL makes a mess for the feature extraction part. Does anyone dare to check with me and help me debug this?

What is the issue exactly?

[kordjamshidi]

I am trying to update this branch, but I have problems because I've made changes to LBjava locally to be able to set weights and I see there are new changes made to LBjava and the version we use in Saul master. So, I need first send a PR to LBjava and upgrade that to be able to compile.

[kordjamshidi]

Lbjava's Learner has changed so much so that not a single line is in its previous place, I failed to merge with my changes, this remains to be resolved.

[kordjamshidi]

It would be good if you could take a look and see if you can merge these since I guess you have restructured the Learner and it is a huge file. @danyaljj

[danyaljj]

Do you have your LBJava modifications on a branch?

[kordjamshidi]

Sorry for the confusion, now just seeng an Email I see Lbjava 24 is in the repo and that is my last version :-), on July 11th. So, the problem is different than what I was thinking.

[kordjamshidi]

So,what are these errors that I get when merging with master: Error:(108, 29) type mismatch; found : edu.illinois.cs.cogcomp.lbjava.util.ExceptionlessInputStream required: edu.illinois.cs.cogcomp.core.datastructures.vectors.ExceptionlessInputStream forward.loadIndexWithId(in) ^

[kordjamshidi]

To summarize: Your help will be to merge just this with Saul/master. Lbjava version is ok. @danyaljj . Seems to be errors due to relocating packages.

[kordjamshidi]

It seems the inference tests do not pass?

[kordjamshidi]

also here I just test a simple list of constrainedClassifiers and for all instances I see this error is reported, 22:01:04 ERROR EntityRelationConstrainedClassifiers$LocConstrainedClassifier$:55 - Warning: Failed to find head could you see if this is related to the changes that recently made by @danyaljj or @bhargav . As mentioned above, it seems setcover tests also are failing. Any ideas?

[kordjamshidi]

It seems there are too many problems to solve with the versions and other changes, to update this PR and get to the actual problem and debugging this semantically. So, please see the issue #386 and the head not found warning which shows with our previous ER examples tests as well.

[bhargav]

@kordjamshidi Looking at the "Head" warning for ER example.

The Semaphore failing tests are all the new SL related ones:

[info] *** 1 TEST FAILED ***
[error] Failed tests:
[error]     edu.illinois.cs.cogcomp.saulexamples.nlp.EntityRelation.weightTest
[error] Error during tests:
[error]     edu.illinois.cs.cogcomp.saulexamples.nlp.EntityRelation.EntityRelationSLTests
[error]     edu.illinois.cs.cogcomp.saulexamples.nlp.EntityRelation.SLTest2
[error]     edu.illinois.cs.cogcomp.saulexamples.nlp.EntityRelation.EntityRelationTests

The tests fail due to:

java.io.FileNotFoundException: ../data/EntityMentionRelation/conll04_small.corp (No such file or directory)

or

[info]   java.io.FileNotFoundException: ./config/DCD.config (No such file or directory)

[kordjamshidi]

the path is fixed but the weightTest is due to SparseNetworkLBP change.

[bhargav]

I still see saul-core/src/main/scala/edu/illinois/cs/cogcomp/saul/classifier/SparseNetworkLBP.scala and saul-core/src/main/scala/edu/illinois/cs/cogcomp/saul/classifier/JoinTrainSparseNetwork.scala files in this branch. I think they were un-deleted when I merged with upstream/master.

@kordjamshidi can you delete both these files try to run the tests? There is some new code that uses SparseNetworkLBP that needs to be changed.

[bhargav]

@kordjamshidi Just saw your other issue about missing public features in SparseNetworkLearner. If it is easier, you can keep the SparseNetworkLBP class and use it only in the ER-SL example for now. We can decide later if we want to have both or not.

[kordjamshidi]

We need to use one thing consistently, since I do not create new ER classifiers for SL, I use the same classifiers of the previous ER.

[kordjamshidi]

I was reviewing this and I think we need to write more an more small tests to see where things go wrong. If someone can help, this would be first step: to add more to the tests. Some of the tests are failing due to the change in the data and need to be corrected too. I have added a README and tried to describe the whole structure of the code and the idea. So far my doubt was that feature indexing and making a global feature vector from all classifiers can be wrong, but it seems that part is correct, added a few tests.

[kordjamshidi]

there is another option also if you want: to review and merge this as it is and then work on the tests. @danyaljj @bhargav

[bhargav]

Will review this sometime before Friday evening. Sorry for the delay.