Unreasonably long running time on TIRA when using DL-models

[Fredrik Johansson]

Hi,

Is there anyone else who have experiences of running some kind of deep learning models on TIRA, and if so, approximately how long does it take you to run it on the article-training and article-test data? I am running a kind of RNN model trained in pytorch (which runs quite fast on a GPU) and it took me almost 20 hours to run (i.e. predict, not train) on the by-article training data on the TIRA-CPU. I know that pytorch models can be rather slow to run on the CPU, but 20 hours is totally insane, especially compared to the time it took to run more shallow models before the early-bird deadline. The model is a around 200MB so it is large but it is only loaded once in the code and then used to call predict() one time per new article. It would probably be faster to collect the whole dataset and then predict larger batches but I find it totally unreasonably that it should take so long to run anyway.

All kinds of input or suggestions on what to do in order to speed up things would be highly appreciated.

Best regards,

Fredrik

[Fredrik Johansson]

Sorry for spamming with two almost similar posts, I thought the first one disappeared in cyberspace but I now realize that it just was waiting for approval.

[Martin Potthast]

Hi Fredrik,

there's no problem to increase the amount of resources for your virtual machine --- within reason. Please estimate the minimum requirements you can get by with, let us know via the organizers mailing list. There's no need, though, to increase resources up to a point where the process runs in minutes; we can accommodate long running times, and running time is not an factor with respect to ranking of submissions. I understand though, that you'd like to have a short(er) turnaround.

Best,

Martin

--
--
You received this message because you are subscribed to the Google Group "PAN".
Visit this group at http://groups.google.com/group/pan-workshop-series
To unsubscribe send email to pan-workshop-se...@googlegroups.com.
---
You received this message because you are subscribed to the Google Groups "PAN Workshop Series on Digital Text Forensics" group.
To unsubscribe from this group and stop receiving emails from it, send an email to pan-workshop-se...@googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

[Fredrik Johansson]

Thanks a lot for the answer. I will try to make some kind of estimate and get back via the organizers mailing list.