Incubator-mxnet: memory leak in Executor when inference in cpu !help!

[ZHEQIUSHUI]

i have try only run

mxnet::cpp::NDArray data = GetDataBatch(src, faces)

it haven`t momery leak,,so i think the problem is not in here.
is my code wrong in use Executor?
thank you!

[bgawrych]

@ZHEQIUSHUI Can you provide more details like MXNet version, how you've built this package - also it would be helpful if you will post your net definition and input shapes

[ZHEQIUSHUI]

@bgawrych the mxnet version is v1.7.x lastest.build with mkl(no cuda),input shape is bitchsize3107*108

and you view the other issue that i have report?

issue https://github.com/apache/incubator-mxnet/issues/19430

[bgawrych]

@ZHEQIUSHUI Can you check with debugger stack trace when memory leak happen? I can't reproduce this issue without network definition. Are you sure that this loop with memcpy is right?

[ZHEQIUSHUI]

@bgawrych can I get your email address?I can send your the net Json file and my inference code(if your have free time).thank your so much.I can make sure there is not memory leak except this function.thank you so much!

[ZHEQIUSHUI]

@bgawrych i got the reason...because i use std::cout to log something....i am so sorry....thank you a lot!!!

[ZHEQIUSHUI]

@bgawrych the reason isn`t cout, i use the same code with v1.5.x,there is not memory leak,but v1.7.x.....

[ZHEQIUSHUI]

but the v1.5.x doesn`t compile with mkl....

[bgawrych]

@ZHEQIUSHUI You can create private repository on github with the code and give me an access to it. I will look at this

[ZHEQIUSHUI]

i have try the same code in cpu with the same version mxnet(v1.7.x),but i have compile two libmxnet([just only cpu] and [with mkl]),i make sure the problem is the mxnet with mkl!!my code have not memory leak!

this is only cpu version mxnet

this is with

mkl

[pengzhao-intel]

@ZHEQIUSHUI thanks to reporting this issue and it's really important for us.
Could you help provide a reproducer? Maybe a proxy model is fine which will accelerate our debugging progress?

[ZHEQIUSHUI]

@pengzhao-intel of course.what do you need?

[bgawrych]

@ZHEQIUSHUI we need minimal part of your model that causes this memory leak, currently we don't even know which operators you're using

And as I said before, if you're still ok with that:

@ZHEQIUSHUI You can create private repository on github with the code and give me an access to it. I will look at this

[ZHEQIUSHUI]

hey @bgawrych @pengzhao-intel i just create a new repository and invited you to it,you can come if you need the model json file.

[ZHEQIUSHUI]

i closed this issue

[bgawrych]

@ZHEQIUSHUI Would you mind to share also inference code in github repository?