Training Fix

[kylase]

• Fixed Function uses old way to load word vectors #9
• Populate missingtags for training

[cmkumar87]

Hi Yuan Chuan “Populate missingtags for training” Are you saying that in the original implementation we missed to add some of the parscit tags that is there in the training data? Muthu

On Tue, Nov 13, 2018 at 12:00 AM Yuan Chuan Kee ***@***.***> wrote: - Fixed #9 <#9> - Populate missingtags for training ------------------------------ You can view, comment on, or merge this pull request online at: #11 Commit Summary - #9 Deprecate load_word2vec_format - Use inputs instead of input as input is a keyword - Use print function - import print_function - Populate tags File Changes - *M* loader.py <https://github.com/WING-NUS/Neural-ParsCit/pull/11/files#diff-0> (19) - *M* train.py <https://github.com/WING-NUS/Neural-ParsCit/pull/11/files#diff-1> (18) - *M* utils.py <https://github.com/WING-NUS/Neural-ParsCit/pull/11/files#diff-2> (27) Patch Links: - https://github.com/WING-NUS/Neural-ParsCit/pull/11.patch - https://github.com/WING-NUS/Neural-ParsCit/pull/11.diff — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#11>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ABmJzDK6XP4C0cInyaMFO0Atil2RSAaIks5uuZsEgaJpZM4YZ8Ib> .

-- Cheers! Muthu

[kylase]

@cmkumar87, not in that aspect. The upstream code has tags as a key for each piece of data, however somehow it has been missing.

[cmkumar87]

@kylase I saw that the change adds tag to id to the load dataset module. This change seems sensitive to me; perhaps @animeshprasad can weigh in. Was this a bug or just an enhancement to make training easier in some way?

• Did yo try retraining with the new code on the cora dataset atleast? Do you get the same published results?

[kylase]

It is a critical bug. If you compare that specific file and line to the original code, you will find that the labels (tags) are not provided to the training.

I don’t know how it managed to run previously, but the git log shows it is non-existent before I take over it.

[cmkumar87]

@kylase that is weird. The file from the first modifications of the Named Entity Tagger contains the tagging scheme. See this commit: 590de7c

But I dont' see it in the WING_NUS/Neural-ParsCit. So @animeshprasad may have changed it for some reason. Perhaps the functionality was moved to a different function? Can you trace the commits to check for the modifications to this file?

This is some of the history I see,
Original NER code: https://github.com/glample/tagger/blob/master/loader.py
tags are there.
In this commit where a completely new version of the file is being uploaded it is not there:
e659829

[kylase]

@cmkumar87, refer to this blame and look for prepare_dataset. This is the commit that removes it.

Yes, I looked at the original code and then realised that it's missing and hence I put it back. The training has been failing because it becomes an unsupervised dataset.

[cmkumar87]

Hi Yuan Chuan Yes, I see in one of the commits the tagids line is being removed from prepare dataset method. I am checking with you again because what you are saying pretty damning. With an error like that, the training file won't have labels and parser would not learn anything! More likely, the objective function that calculates loss would through an error since it's not able to see the target/label. Yet, we have state of the art results with a lot experiments run over the dataset! So, probably the functionality was moved elsewhere. Cheers! Muthu

…

On Tue, 13 Nov 2018 at 20:04, Yuan Chuan Kee ***@***.***> wrote: @cmkumar87 <https://github.com/cmkumar87>, refer to this blame <https://github.com/WING-NUS/Neural-ParsCit/blame/bb9b9a002582a6619e8fb6f14956e12fddc19608/loader.py> and look for prepare_dataset. This is the commit that removes it. — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#11 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ABmJzL6qjLCPoE22aHmj9ebMBXtLXf97ks5uurVdgaJpZM4YZ8Ib> .

[kylase]

I have no idea that happened between the code that was run for the paper and the commit. I have been looking at the training code and it seems to have been commented out components other than the training with the training dataset.

Now I am working on restoring the testing portion then follow by the cross-validation.

[cmkumar87]

@kylase Normally the training code and testing code is factored and are executed conditionally based on a command line args passed.
Are you saying that this is not how n-parscit is written? Did the developer remove / comment out the training code to allow the parser to operate under the test mode?

[knmnyn]

Has this issue been resolved? Seems from the comments that it is still unresolved.

[cmkumar87]

Update on this:

• This codebase was forked from a version that is different from the one that was used to run the experiments for Prasad, Kaur, Kan, 2018, IJDL.
• The correct version in question was identified to be on a local server. @kylase is trying to run this; he has been facing out-of-memory errors on the wing internal gpu server possibly due to 'some' theano misconfiguration. Will let him say more.