Improvising time for coreferencing about neuralcoref

[pramod2157]

Just tried running it on multiple instances, it works amazing. Thanks @thomwolf

[julien-c]

Hi @Nbharath all the credit goes to @thomwolf 👏

We haven't published the training workflow yet but it is coming in couple of weeks.

[julien-c]

As for the resolution time, I'll let Thomas pitch in!

[thomwolf]

Hi @Nbharath,

• I am currently cleaning up the pyTorch code and taking the occasion to refine a bit the mention detection algorithm. I will publish the code as soon as possible.
• Right now the library uses numpy to compute the matrix multiplication so it can't make use of the GPU. To put it on GPU we would also have to refactor a bit the workflow to parallelize the computation of pair scores. We use it for very small files (dialog) so it is not in my top priority list but if you feel like trying, you are more than welcome ! (maybe using something like cupy would be the simplest way)

[lavinadurgani]

Hello Thomas,

Is the pyTorch code published now ? Will the pyTorch code help to reduce the coreference resolution processing speed if executed on GPU ?

Regards,
Lavina

[arunzz]

Hello

First of all, thanks for the nice library.
The time taken is very long for an average size content. The number of mentions detected are more and we cut short some of the accepted entities and some grammar rules but still the time is too long.
Any tips on how can we reduce it? Please let know.

[thomwolf]

Yes, just wait for the release of version 0.3 (hopefully) next week. Version 0.3 will be about 20 times faster :-)

…

On Mon, May 21, 2018 at 7:03 AM, Arun Kumar ***@***.***> wrote: Hello First of all, thanks for the nice library. The time taken is very long for an average size content. The number of mentions detected are more and we cut short some of the accepted entities and some grammar rules but still the time is too long. Any tips on how can we reduce it? Please let know. — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#4 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AHA0HXPjrjrlYBahpxyyWASsPzh0l8vAks5t0kqRgaJpZM4Pez7j> .

[arunzz]

Glad to know. I will keep waiting then. :)

[dlindbe]

Is there a new ETA for the cython rewrite?

[thomwolf]

Yes, it should be this week, the code pretty much ready (see extension branch), I just have to finish a few edge cases and functions and write the doc (it's now integrated as a spaCy extension so the interface is a bit different). It'll be 50 to 100 times faster on a large doc.

[theSage21]

any way we can help out?

[pramod2157]

@thomwolf when the Cython version will be released?

[thomwolf]

It is released! I have updated the readme!

[arunzz]

It gives an error while loading the new model saying unicode expected not char in line import neuralcoref. Can you pls help

…

On Jun 14, 2018, at 3:13 PM, pramod2157 ***@***.***> wrote: Just tried running it on multiple instances, it works amazing. Thanks @thomwolf — You are receiving this because you commented. Reply to this email directly, view it on GitHub, or mute the thread.