Add support to use different numbers of samples in the queue

[fepegar]

Related to:

• Queue with different samples_per_volume for each volume #604
• Extend the queue to take different samples per volume for each subject in the dataset #609
• Samplers and Queues supporting different number of sampled patches per each subject of the SubjectsDataset #789

Description
After these changes, one can decide how many patches to extract from each subject, using a num_samples kwarg when instantiating the subjects.

Checklist

• I have read the CONTRIBUTING docs and have a developer setup (especially important are pre-commitand pytest)
• Non-breaking change (would not break existing functionality)
• Breaking change (would cause existing functionality to change)
• Tests added or modified to cover the changes
• Integration tests passed locally by running pytest
• In-line docstrings updated
• Documentation updated, tested running make html inside the docs/ folder
• This pull request is ready to be reviewed
• If the PR is ready and there are multiple commits, I have squashed them and force-pushed

[codecov]

Codecov Report

Merging #795 (4118985) into main (cc9ccb3) will increase coverage by 0.09%.
The diff coverage is 100.00%.

@@            Coverage Diff             @@
##             main     #795      +/-   ##
==========================================
+ Coverage   93.57%   93.67%   +0.09%     
==========================================
  Files          72       72              
  Lines        4562     4566       +4     
==========================================
+ Hits         4269     4277       +8     
+ Misses        293      289       -4     

Impacted Files	Coverage Δ
torchio/constants.py	100.00% <100.00%> (ø)
torchio/data/queue.py	86.60% <100.00%> (+3.12%)	⬆️
torchio/data/dataset.py	100.00% <0.00%> (+1.96%)	⬆️

📣 Codecov can now indicate which changes are the most critical in Pull Requests. Learn more

[jmlipman]

Hi Fernando.

I don't know whether I can download torchio and apply your pull request to try it locally, so, I could only read the code.

I like the idea of having a specific property to store the samples_per_volume (I was doing this, in practice). I have a question/concern. Consider the following scenario:

• Queue's size is 50.
• We are training on 2D slices of four 3D images.
• Each image has 30 slices. (total slices 30*4=120).

I have the feeling that the queue will be filled like this:

First fill: 30 slices of image 1 + 20 slices of image 2.
Second fill: 30 slices of image 3 + 20 slices of image 4.
Consequently, ignoring 10 slices from image 2 and 10 from image 4. This is because every time "fill" is called it loads a new subject without checking if we are done taking the all the slices from the last subject the previous time "fill" was called. At least I don't see how this is done, but, as I've mentioned before, I didn't run the code and I might be wrong.

[fepegar]

I don't know whether I can download torchio and apply your pull request to try it locally, so, I could only read the code.

Maybe you can use

git clone -b 604-queue-different-samples [email protected]:fepegar/torchio.git

About the other concerns, I think that using a much larger queue length would make the stochastic process less sensitive to this issue. We might get fewer patches from a specific subject during an epoch, but that shouldn't happen in the next one if the subjects list is shuffled (which it is, by default). If you still think this is a concern, we can look into it and find alternatives.