import error

[adaaaaaa]

Windows 10 x64
python 3.6.6
install mxnet and import correctly,but install mxnet-mkl import with an error like this:

`import mxnet
Traceback (most recent call last):

File "", line 1, in
import mxnet

File "f:\python36\lib\site-packages\mxnet_init_.py", line 24, in
from .context import Context, current_context, cpu, gpu, cpu_pinned

File "f:\python36\lib\site-packages\mxnet\context.py", line 24, in
from .base import classproperty, with_metaclass, _MXClassPropertyMetaClass

File "f:\python36\lib\site-packages\mxnet\base.py", line 174, in
_LIB = _load_lib()

File "f:\python36\lib\site-packages\mxnet\base.py", line 166, in _load_lib
lib = ctypes.CDLL(lib_path[0], ctypes.RTLD_LOCAL)

File "f:\python36\lib\ctypes_init_.py", line 348, in init
self._handle = _dlopen(self._name, mode)

OSError: [WinError 126] 找不到指定的模块。`
how to deal...

[yajiedesign]

this because lost dll.please run depend(http://www.dependencywalker.com/) check what dll lost and tell me.

[apeforest]

@sandeep-krishnamurthy Please label this Windows, Build

[adaaaaaa]

@yajiedesign what dll should i try to check? never use that software before...

[pengzhao-intel]

#11049

[yajiedesign]

@adaaaaaa libmxnet.dll

[adaaaaaa]

@yajiedesign tons of DLL? how should i get and fix them?

[adaaaaaa]

anyone can help?

[zhreshold]

@adaaaaaa missing mkldnn.dll and mklml.dll , are they included in the package?

You can find mklml.dll here: https://github.com/intel/mkl-dnn/releases/download/v0.14/mklml_win_2018.0.3.20180406.zip
Not sure about mkldnn prebuilt dll, I think it's easier to build though.

[yajiedesign]

@adaaaaaa I do some fix.please wait tomorrow.

[kohillyang]

where can I find mkldnn.dll?

[adaaaaaa]

@yajiedesign gluon's tutorial "cnn-scratch" using mxnet-mkl version 20180731 perform badly...
executed in 3m 40s
while in linux using mxnet-cu92mkl use cpu it executed in 23.8s nearly 10 times less...
what cause this?

[kohillyang]

@adaaaaaa Do you have mkldnn.dll? Or can you tell me where I can find it? Thanks.

[adaaaaaa]

@kohillyang no i don't have...the newest version is fine now , don't need to fix the ***.dll now...

[yajiedesign]

@adaaaaaa i don't know,do you can test mxnet-cu92mkl in windows?

[adaaaaaa]

@yajiedesign no...cuda is too big...i'm not going to try it in windows...
in Windows,i'm sure mxnet-mkl using cpu to executed "cnn-scratch" in 3m 40s
and in Linux using gpu to executed "cnn-scratch" in 5.9s,
in Linux using cpu to executed "cnn-scratch" in 23.8s.

[adaaaaaa]

really a big space to improve...
while other interesting things is i test single float point in aida64...
it shows that 1070ti is 10 times faster than i7-8700K
while in Linux "cnn-scratch benchmark" it's only 4 times faster...
something wrong in gpu version in mxnet?

[kohillyang]

I find mxnet_mkl 3 times slower than mxnet without mkl. The inference time of mxnet is about 1s but it's 3s if mkl.
And whether mxnet_mkl or mxnet cannot make full use of multi cores. I test it on i7 2600, but the utilization is only about 40% if batch size is set to 12.

[pengzhao-intel]

@adaaaaaa @kohillyang where is cnn-scratch?

[adaaaaaa]

ever a part of gluon tutorial in chapter convolutional-neural-networks...
cnn-scratch.zip
@pengzhao-intel

[adaaaaaa]

other "benchmark" is the latest tutorial's lenet tutorial...
in Windows,using mxnet-mkl,cpu,executed in 26 to 27s every epoch
in Linux,using mxnet-cu92mkl,cpu,executed in 3.5 to 3.9s every epoch;
gpu,executed in 1.2 to 1.4s every epoch...
@yajiedesign @pengzhao-intel

[adaaaaaa]

any process now?

[pengzhao-intel]

WIP to verify the performance.

FYI, the performance on Linux is fully validated; however, the Window is w/ lower priority.
And in our local, we don't have much windows system to reproduce the issue.

I suggest you try Linux first. If you see the same perf issue on Linux, it will be very helpful.