Skip to content
New issue

Have a question about this project? # for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “#”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? # to your account

SRL taking too much memory #708

Open
qiangning opened this issue Nov 9, 2018 · 1 comment
Open

SRL taking too much memory #708

qiangning opened this issue Nov 9, 2018 · 1 comment

Comments

@qiangning
Copy link
Member

qiangning commented Nov 9, 2018

This is a follow-up issue from iss #656 . In #656 , I reported my observation that our pipeline would fail if one wanted to add Verb SRL via a computer with insufficient memory.

Some data points:

  • 16G Mac (my laptop). Fails.
  • 32G Ubuntu (my desktop). Fails.
  • 32G AWS Linux. Fails.
  • 32G Mac. Works. (Ben's laptop; thanks @Slash0BZ)
  • 64G AWS Linux. Works.
  • Our server, which usually have >100G. Works.

Originally, I thought this was due to the fact that our current Verb_SRL "unnecessarily" requires NER_CONLL, but now I realize:

  • Adding NER_CONLL isn't by mistake, since NER_CONLL is defined explicitly in the feature files. For example, see this line.
    • My first question is: is NER really critical for Verb SRL? @christos-c
    • To evaluate the importance of this NER feature for SRL myself, I have also tried to delete the ne embedding feature in all those feature files and also remove NER_CONLL from this line. Then I have tried to retrain the SRL model via this. However, the trainer failed with some missing TA views like DEPENDENCE:PARSE_STANFORD. @christos-c is there any obvious errors in my procesure?
  • Even if we managed to remove NER_CONLL from SRL, I guess the memory consumption of SRL wouldn't be significantly reduced. I tested myself and also confirmed with @yxd126 that NER_CONLL normally takes roughly 8G memory, which isn't the major reason why SRL is failing on machines with less than 32G memories. So I think the main problem resides in SRL itself.
    • @Slash0BZ Can you share your memory profiling result regarding this issue?
    • @christos-c Is our SRL supposed to be this memory extensive? I'm asking since I tried allennlp srl and it runs smoothly on an AWS machine with 16G memory. This is a critical problem for me since AWS machines with 64G memories are too expensive.
@danyaljj
Copy link
Member

danyaljj commented Nov 9, 2018

(adding @SB299792458 to the thread since he's working on a new version of SRL)

# for free to join this conversation on GitHub. Already have an account? # to comment
Labels
None yet
Projects
None yet
Development

No branches or pull requests

2 participants