Installed MKL - no improvement in face detection

[sergei-sh]

I've installed MKL, recompiled DLIB and see almost no time difference in face detection.
Ensured importing the correct module.
MKL was found during compilation, no errors there.
I use get_frontal_face_detector from Python and run 64 consecutive detections on the same images.

Using the latest dlib under CentOs 7.5. I also compile with CUDA at the same time.

[davisking]

Hog doesn’t use these so this is expected. Turn on avx or neon instructions instead.

[sergei-sh]

Hi, thanks for the answer. But I'm still confused. You use BLAS in matrix.h and matrix is used in scan_fhog_pyramid. To me this means the HOG detector should benefit from MKL. Am I wrong?

[davisking]

Yes, you are wrong.

[sergei-sh]

But I clearly see my Dlib is linked against MKL libraries which contain BLAS functions which are being used by matrix class.

Could you help with any hint to understand which parts of Dlib do utilize this functionality?

[davisking]

Just look at the fhog extraction code. You can see what it does. That goes for anything else, look and see what it does.

[xsacha]

Is there much benefit for MKL if you are using CUDNN?

[davisking]

The hog based detectors don’t use mkl or cudnn. So it doesn’t matter either way.

[xsacha]

What about if we use DNN face detector? With CUDNN, would MKL have any benefit?

[davisking]

If you are using the DNN tooling the network is running either all on the GPU or all on the CPU. So installing a CPU acceleration library like the MKL will do nothing if you are already using cudnn.

[xsacha]

The feature detector (shape) for face cannot be run on the GPU at this stage right? So, even with everything on GPU, this part (which is quite expensive) is on the CPU. Would it make use of MKL?

[davisking]

The shape_predictor should be super fast. But yes, it's just on the CPU. I don't think it benefits from the MKL either.

[xsacha]

It uses more than everything else on CPU (which is close to 0%) ;) So it ends up in profiling.

So if we are using CUDNN, then ideally we don't need BLAS?

[davisking]

I don't think we are talking about the same thing. The shape_predictor runs in about a millisecond on a normal computer.

[xsacha]

It does add up on my machine. Consider the scenario where you have 10 HD cameras and up to a dozen faces per camera. It seems to matter more about the size of the face too so I attempt to shrink them first.
The CUDNN keeps up fine but the CPU begins to be used up.

[davisking]

Yeah that’s a lot. Can’t you run multiple shape models on multiple cpu cores?

[xsacha]

It's over multiple cores already but with database and everything else, it adds up.
Was just wondering if it uses MKL. If not, I guess I can remove it.

[dlib-issue-bot]

Warning: this issue has been inactive for 30 days and will be automatically closed on 2018-12-24 if there is no further activity.

If you are waiting for a response but haven't received one it's likely your question is somehow inappropriate. E.g. you didn't follow the issue submission instructions, or your question is easily answerable by reading the FAQ, dlib's documentation, or a Google search.

[dlib-issue-bot]

Notice: this issue has been closed because it has been inactive for 35 days. You may reopen this issue if it has been closed in error.