Re: LSTM regression (was RE: Include MKLDNN into default mxnet pip package)

Ran benchmark and it addresses issue. Thanks. On 11/28/18, 6:02 PM, "Zhao, Patric" wrote: MKL-DNN v0.17.1 is released https://github.com/intel/mkl-dnn/tree/v0.17.1 I have submitted the PR to pin this release version. Thanks, --Patric > -Original

RE: LSTM regression (was RE: Include MKLDNN into default mxnet pip package)

MKL-DNN v0.17.1 is released https://github.com/intel/mkl-dnn/tree/v0.17.1 I have submitted the PR to pin this release version. Thanks, --Patric > -Original Message- > From: Zhao, Patric [mailto:patric.z...@intel.com] > Sent: Wednesday, November 28, 2018 8:07 PM > To:

RE: Include MKLDNN into default mxnet pip package

+1 for making MKL-DNN default in master branch first for broad testing :) My suggestion is to make MKL-DNN default on the master branch firstly after 1.4.0 releasing branch is cut off. That will help MKL-DNN backend to be widely used and tested by MXNet users who are building MXNet from

Re: Significant performance regression in SpMV

Can you create an issue for this with codesnippet we can run? Alex On 11/27/18, 9:55 PM, "Zheng, Da" wrote: Hello all, I notice there is a significant performance regression in SpMV after this PR (https://github.com/apache/incubator-mxnet/pull/12380). It seems the problem

Build from Source Instruction Changes

I would like to raise a PR for discussion here. In the mxnet installation docs, we currently have three inconsistent ways of compiling the mxnet backend: 1. Use make by passing in the configuration directly make USE_BLAS=openblas 2. Use make with config.mk echo "USE_BLAS=openblas" >>

Re: Include MKLDNN into default mxnet pip package

Thanks for answering Tao. I would like to add that we have the env flag that disables MKLDNN operators if regression occurs. On 11/28/18, 6:05 AM, "Lv, Tao A" wrote: Hi Hagay, thank you for bringing these questions together. I also summarized my opinions here for you easy to check.

Re: using conan to manage Apache Incubator MXNet project dependencies

> Is it possible to have an offline build somehow? For example, if all > dependencies would be stored locally. Probably, that would require some > modifications to the conan file, right? as soon as all dependencies are stored locally, conan will use binaries from the local cache. no modifications

Re: using conan to manage Apache Incubator MXNet project dependencies

Great, thank you for your answers. Is it possible to have an offline build somehow? For example, if all dependencies would be stored locally. Probably, that would require some modifications to the conan file, right? вт, 27 нояб. 2018 г. в 16:59, Konstantin Ivlev : > > Would it be possible to

trouble with foreach operator in conjunction with multiple GPUs

Hello fellow MXNetters We've seen that the subgraph execution mechanism that is used to run things like the foreach operator causes MXExecutorForward to block, instead of just issuing the ops in the normal asynchronous way

RE: Include MKLDNN into default mxnet pip package

Hi Hagay, thank you for bringing these questions together. I also summarized my opinions here for you easy to check. - Make MKL-DNN default in MXNet pip package [Tao]: My suggestion is to make MKL-DNN default on the master branch firstly after 1.4.0 releasing branch is cut off. That will help

LSTM regression (was RE: Include MKLDNN into default mxnet pip package)

Hi Anirudh, The LSTM performance bug is fixed by MKL-DNN and PR in here (https://github.com/apache/incubator-mxnet/pull/13417). I am still working on MKL-DNN team to get a patch release for MXNet 1.4 in 1 or 2 days. Will update the status soon. Thanks everyone. --Patric > -Original

Re: Include MKLDNN into default mxnet pip package

Hey all, I'm also supportive of making MKLDNN the default build for MXNet, but there were a few questions asked in the thread that I am not sure were answered. Would be great if Alex and others who worked on MKLDNN and that are proposing it to be the default can answer them clearly: - What the