how to use image streaming to reduce memory usage and cpu usage in image stitching

Beginner Questions
,
1

The cpu usage and memory usage is pretty high when stitching tiles into a single one. 140
tiles(2592*1944) stitching occupy nearly 100% cpu usage for more than 60s. i wonder if the usage won’t be so high in the image streaming mode as they are computed one by one. if so, is there a demo code or guidance how to use image streaming in itk, thanks.
this article mention that itk support image streaming mode.

2

Using streaming will reduce memory use, but not CPU use. To reduce CPU use, you should limit number of threads ITK uses. There are a few ways. With 5.0 refactoring, MultiThreader is now called MultiThreaderBase.

Here is how to use streaming:

1 Like
3

to limit the cpu usage, this works for python

itk.MultiThreaderBase.SetGlobalMaximumNumberOfThreads(5)
itk.MultiThreaderBase.SetGlobalDefaultNumberOfThreads(5)

but even if the thread number setting is effective, the cpu usage still get to 100% for a few seconds while resampling

8d03a1f1eaf477a553038a0d745479b
8d03a1f1eaf477a553038a0d745479b1829×781 85 KB

is there any way to limit the cpu usage in TileMergeImageFilter, or bind cpu core also a acceptable solution that spare other cpu core to be idle. now i add time.sleep() to limit the resample cpu usage like this

resampleF = itk.TileMergeImageFilter[type(color_images[0])].New()
resampleF.SetMontageSize(stage_tiles.GetAxisSizes())
# resampleF.SetNumberOfThreads(10)

for t in range(stage_tiles.LinearSize()):
  resampleF.SetInputTile(t, color_images[t])
  index = stage_tiles.LinearIndexToNDIndex(t)
  transform = montage.GetOutputTransform(index)
  resampleF.SetTileTransform(index, transform)
  time.sleep(0.1)
resampleF.Update()

yet i think that’s not the best way to do it and will increase processing time significantly
by the way the TileMergeImageFilter does not support SetNumberOfThreads

Traceback (most recent call last):
  File "E:\projects\ITKMontage\examples\SimpleMontage.py", line 122, in <module>
    resampleF.SetNumberOfThreads(10)
AttributeError: 'itkTileMergeImageFilterIRGBUS2' object has no attribute 'SetNumberOfThreads'
4

Try SetNumberOfWorkUnits: resampleF.SetNumberOfWorkUnits(10).

I wouldn’t expect that adding sleep in the for loop would reduce CPU use, as the computations happen when you invoke resampleF.Update().

1 Like
5

Maybe this line should have been MultiThreaderBase::Pointer mt = this->GetMultiThreader(); for resampleF.SetNumberOfWorkUnits(10) to have the desired effect. But

itk.MultiThreaderBase.SetGlobalMaximumNumberOfThreads(5)
itk.MultiThreaderBase.SetGlobalDefaultNumberOfThreads(5)

should work.

1 Like
6

If you don’t want your computer to “lock up”, you could give the process “itk_test” priority lower than “normal”. Also, how many physical cores (not logical, hyper-threaded cores) does your CPU have? You need to set number of threads below that number to ensure CPU use does not reach 100%.

7

my computer has 6 physical cores so i set the number 5, bu t it seems still reach 100% usage, yet resampleF.SetNumberOfWorkUnits(10) worked i think. i will update my results after test more times

8

the max cpu usage is under 70% now after hundreds times test these days

9

here I found a python version link

1 Like
10

replace

resampleF.Update()

by

writer = itk.ImageFileWriter.New(resampleF.GetOutput().astype(itk.RGBPixel[itk.UC]))
writer.SetFileName(out_file)
writer.SetNumberOfStreamDivisions(5)
writer.Update()

it works great for me, the memory usage line no waves like nothing happened
1645096260(1)
and the total process time is really quick, 160 tiles for 30.11 seconds
image

1 Like
11

I guess that lower memory use leads to no use of swap file on hard disk, hence the overall processing is faster.

1 Like