[ITK] Bspline registration with Mutual Information v4 : memory concern

[Timothee Evain]

Dženan,

Your fix works like a charm. The process take now ~4Gb(in release).
I observed that this memory isn't released after the end of the registration process. Is that intended ?
Also, to write it down for potential users who want to patch their 4.10 version or earlier, there is a small macro to change (ITK_CONSTEXPR_VAR into ITK_CONSTEXPR) since master branch had a modification about macros.

Anyway, many thanks to you for the swift help ! :)

Tim



----- Mail original -----
De: "Dženan Zukić" <dzenanz at gmail.com>
À: "Timothee Evain" <tevain at telecom-paristech.fr>
Cc: community at itk.org
Envoyé: Mardi 6 Septembre 2016 17:43:49
Objet: Re: [ITK] Bspline registration with Mutual Information v4 : memory concern

Hi Tim,

I did some memory profiling using 4.9.0 (which also exhibits the problem).
With current code, the MattesMutualInformationImageToImageMetricv4::
DerivativeBufferManager::DoubleBufferSize() increases the buffer to huge
sizes, because CheckAndReduceIfNecessary() fails to acquire the
synchronization lock. The reason is probably that memory-bound combining
the PDF derivatives is slower than its CPU-bound calculation.

In release mode, the buffers grew to 50GB before I killed the process. In
debug mode, they only grow to around 5GB, again probably because the
calculation speed is on par with PDF combining due to debug overhead.

Can you review the proposed fix?
http://review.source.kitware.com/21501

Regards,
Dženan

On Tue, Sep 6, 2016 at 5:32 AM, Timothee Evain <tevain at telecom-paristech.fr>
wrote:

> Hi Dženan,
>
> Sure!
> You could find here an archive https://www.dropbox.com/s/9scgn9if39i4a1x/
> V4RegistrationMemoryLeak.tar?dl=0
> It contains:
> -Fixed and moving images in mhd format
> -The binary (for windows)
> -The source code with its CMakeLists if you want to compile for another OS
>
> Thanks,
>
> Tim
>
>
> ----- Mail original -----
> De: "Dženan Zukić" <dzenanz at gmail.com>
> À: "Timothee Evain" <tevain at telecom-paristech.fr>
> Cc: community at itk.org
> Envoyé: Lundi 5 Septembre 2016 20:54:29
> Objet: Re: [ITK] Bspline registration with Mutual Information v4 : memory
> concern
>
> Hi Tim,
>
> can you turn this snippet into a runnable example which reproduces the
> memory leak?
>
> Regards,
> Dženan
>
> On Mon, Sep 5, 2016 at 12:27 PM, Timothee Evain <
> tevain at telecom-paristech.fr
> > wrote:
>
> > Hello everyone,
> >
> > I'm trying a Bspline registration based on the Mutual Information metric
> > (v4) and I'm facing a memory problem.
> > When running the registration, the memory usage skyrockets in a second
> > until freezing the PC by lack of memory.
> > I don't know why : this doesn't happen when I'm switching to another
> > metric (e.g. Correlation), the images aren't that big (~400k voxels) and
> a
> > previous rigid registration phase behaved correctly.
> > Here is my code snippet, what am I doing wrong ? (Image overlap had been
> > ensured)
> >
> >    typedef itk::Image<float,3> ImageTypef;
> >
> >    itk::ImageRegistrationMethodv4<ImageTypef,ImageTypef>::Pointer
> > WarpHelper=
> >       itk::ImageRegistrationMethodv4<ImageTypef,ImageTypef>::New();
> >    itk::MattesMutualInformationImageToImageMetricv4<ImageTypef,
> ImageTypef>::Pointer
> > MIMetrics=
> >       itk::MattesMutualInformationImageToImageMetricv4<ImageTypef,
> > ImageTypef>::New();
> >    itk::BSplineTransform<double,3,3>::Pointer BSplineTransform=
> >       itk::BSplineTransform<double,3,3>::New();
> >    itk::LBFGSOptimizerv4::Pointer Optimizer=
> >       itk::LBFGSOptimizerv4::New();
> >
> >    MIMetrics->SetNumberOfHistogramBins(50);
> >    MIMetrics->SetUseMovingImageGradientFilter(false);
> >    MIMetrics->SetUseFixedImageGradientFilter(false);
> >    MIMetrics->SetUseFixedSampledPointSet(false);
> >    MIMetrics->SetVirtualDomainFromImage(TestIm);
> >
> >
> >    WarpHelper->SetFixedImage(TestIm);
> >    WarpHelper->SetMetricSamplingPercentage(0.002);
> >    WarpHelper->SetMovingImage(MovIm);
> >    WarpHelper->SetMetric(MIMetrics);
> >    //WarpHelper->SetMetricSamplingStrategy(WarpHelper->RANDOM);
> >    WarpHelper->InPlaceOn();
> >    WarpHelper->SetInitialTransform(BSplineTransform);
> >    WarpHelper->SetOptimizer(Optimizer);
> >
> >    itk::Array<double> ShrinkFactorPerLevel(1);
> >    ShrinkFactorPerLevel[0]=1;
> >    itk::Array<double> SmoothSigmaPerLevel(1);
> >    SmoothSigmaPerLevel[0]=0;
> >    WarpHelper->SetNumberOfLevels(1);
> >    WarpHelper->SetSmoothingSigmasPerLevel(SmoothSigmaPerLevel);
> >    WarpHelper->SetShrinkFactorsPerLevel(ShrinkFactorPerLevel);
> >
> >    /* ----- Initialize the Bspline Transform ----- */
> >    BSplineTransform->SetTransformDomainOrigin(TestIm->GetOrigin());
> >    BSplineTransform->SetTransformDomainDirection(
> TestIm->GetDirection());
> >
> >    itk::Vector<double,3> PhysicalDim;
> >    for (int dim = 0; dim < 3; dim++)
> >    {
> >       PhysicalDim=(TestIm->GetSpacing()[dim])*float((TestIm->
> > GetLargestPossibleRegion().GetSize()[dim])-1);
> >    }
> >    BSplineTransform->SetTransformDomainPhysicalDimensions(PhysicalDim);
> >
> >    itk::Size<3> meshSize;
> >    meshSize.Fill(10-3);// NbGridNodesEachDim-SplineOrder
> >    BSplineTransform->SetTransformDomainMeshSize(meshSize);
> >
> >    BSplineTransform->SetIdentity();
> >
> >    /* ----- Optimizer setup ----- */
> >    itk::Array<double> ScaleBspline(BSplineTransform-
> > >GetNumberOfParameters());
> >    ScaleBspline.fill(1);
> >
> >    Optimizer->SetLineSearchAccuracy(0.9);
> >    Optimizer->SetDefaultStepLength(1.5);
> >    Optimizer->TraceOn();
> >    Optimizer->SetGradientConvergenceTolerance(1e-3);
> >    Optimizer->SetMaximumNumberOfFunctionEvaluations(20);
> >    Optimizer->SetDoEstimateScales(false);
> >    Optimizer->SetScales(ScaleBspline);
> >
> >
> >    try
> >    {
> >       WarpHelper->Update();
> >    }
> >    catch (itk::ExceptionObject &err)
> >    {
> >       std::cerr<<err;
> >       return 0;
> >    }
> >
> > Thanks,
> >
> > Tim
> > _______________________________________________
> > Community mailing list
> > Community at itk.org
> > http://public.kitware.com/mailman/listinfo/community
> >
>