How to manually release memory of Image after data was loaded?

[alabamagan]

Suppose I created a list of LabelImage, I need to loop them in a covariant fashion:

for a in im_list:
  for b in im_list:
    if a == b: continue
    # Do somthing with a & b
    # need to release memory used by b
...

So I only need simultaneously a & b, but after the data was loaded in the inner loop, it persist and there is a waste of memory. Is there a way I can release the memory and revert the LabelImage to the stage before the data tensor was loaded?

What I've tried:

del b # in inner loop, doesn't release mem
del b.data # can't delete attribute

Cheers

[fepegar]

Hi, @alabamagan.

This is probably because instances in a and b are repeated, so even if you delete them with del, you will still be keeping a pointer to them in one of the lists and memory will not be released. A possible solution is iterating over SubjectsDataset, which makes copies of the subjects. Then, maybe you can use the path of the image or the tensor data to compare two images.

import torchio as tio
subjects = [tio.Subject(im=im) for im in im_list]
dataset = tio.SubjectsDataset(subjects)
for sub1 in dataset:
    for sub2 in dataset:
        if compare(sub1.im, sub2.im):
            do_sth()

[alabamagan]

Thanks!
What you proposed seems to work.