switch to mxnet-cu92mkl

[ZhennanQin]

No description provided.

[ZhennanQin]

@zhreshold After some investigation, found that it's crashed in mxnet::op::BipartiteMatchingForwardmshadow::cpu. This op doesn't relate to mkl or mkldnn , but possibly has memory access bug exposed by mkl build. Do you know any risk for this op causing heap corruption? Thanks in advance.

[zhreshold]

@ZhennanQin Can you try the latest nightly build of mkl cu92? Recently I just merged a PR fixing memory issue of BipartiteMatching OP

apache/mxnet#13727

[mli]

Job PR-578-4 is done.
Docs are uploaded to http:https://gluon-vision-staging.s3-website-us-west-2.amazonaws.com/PR-578/4/index.html
Code coverage of this PR: vs. Master:

[ZhennanQin]

@zhreshold After updating mxnet-cu92mkl to 1.5.0b20190115, CI can pass. Hurray! Thanks for pointing me the fix.

Do you have any other trouble for using mxnet mkl build?

[zhreshold]

Looks good as long as CI has no complains. I will keep this one open for one day in case @hetong007 and @zhanghang1989 have extra prompt

[pengzhao-intel]

@zhreshold do you have the plan to cherry pick #13727 into release 1.4.x?

[zhreshold]

@pengzhao-intel It was actually picked into 1.4.x by apache/mxnet#13900

[pengzhao-intel]

@wuxun-zhang please help test how much performance benefit we can gain after changing MKLDNN backend.
@zhreshold any recommended benchmark scripts for benchmarking or only use the CI tests?

[zhreshold]

There's no existing benchmark scripts, but I think we can creating some using exsiting inference code.

[pengzhao-intel]

@zhreshold thanks a lot. Ping us when the code is ready.

[pengzhao-intel]

@zhreshold could you help merge the PR? thanks.