mxnet-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Zhao, Patric" <patric.z...@intel.com>
Subject RE: Project Proposal for fused CPU RNN OPs to the release 1.3
Date Sat, 23 Jun 2018 02:40:18 GMT
Hello Steffen,

Really thanks to look into our proposal. I totally understand your concern that the quality
is the most important thing.
We will pay much attention on it.

Regarding RNN Ops, the new OP provides about 2-3X performance boost (the performance section
of proposal).
Most importantly, it makes the gluon RNN/NLP model can be hybridized to symbolic models in
both CPU and GPU
(previously it only works on GPU w/o the fused CPU Ops).
For the correctness, the unit tests was added w/ the PR and we also tested the accuracy with
the real cases. 
I will update the doc with more information about the correctness and our experiment results.

Regarding MKL-DNN integration,  the issues (bugs and some corner cases) has gradually emerged
along with more users switching to it. 
Absolutely, it is really important to make it stable and we really care about this.
Zheng Da, Alex and our team are working on the known issues and have already fixed lots of
them. 
Furthermore, a bunch of unit test, including gluon, symbolic, CPP cases, are added to cover
more situations.
Obviously, the MKL-DNN backend trends to more complete and robust. We hope to upgrade it to
GA in the 1.3.

Finally, I think we need some patience for the new features and incubate it to the mature.

Thanks for your suggestions again.

--Patric

> -----Original Message-----
> From: Steffen Rochel [mailto:steffenrochel@gmail.com]
> Sent: Friday, June 22, 2018 10:45 PM
> To: dev@mxnet.incubator.apache.org
> Cc: Lv, Tao A <tao.a.lv@intel.com>; Li, Hao H <hao.h.li@intel.com>; Ye,
> Jason Y <jason.y.ye@intel.com>; Emani, Ashok <ashok.emani@intel.com>
> Subject: Re: Project Proposal for fused CPU RNN OPs to the release 1.3
> 
> Thanks Patric, appreciate  your contributions. I looked at your design
> proposal. I'm missing any statements about validation of correctness and
> performance of the integrated solution. I would suggest to pay more
> attention to this aspect as we struggled in previously releases with the
> quality of the integration. As you know, we still have too many issues on
> MKL-DNN integration to move from experimental to GA stage.
> Regards,
> Steffen
> 
> On Thu, Jun 21, 2018 at 12:09 AM Zhao, Patric <patric.zhao@intel.com>
> wrote:
> 
> > Hi MXNET owner,
> >
> > Recently, we (Intel engineers) have implemented the fused RNN
> > operations
> > (LSTM/GRU/vRNN) for the CPU, including bidirectional, multiple layers,
> > inference/training.
> > The LSTM and GRU PR was merged and vRNN code will be PR soon.
> >
> > The new APIs make the gluon and symbolic models much faster :)
> >
> > Thus, I have added a new row in the 1.3 proposal table and hope the
> > end user can leverage the new feature easily.
> >
> >
> >
> https://cwiki.apache.org/confluence/display/MXNET/Project+Proposals+fo
> > r+next+MXNet+Release
> >
> > Feel free to let me know for any feedbacks and suggestions.
> >
> > BR,
> >
> > Thanks,
> >
> > --Patric
> >
> >
Mime
View raw message