Setting NUM_SAMPLES when using sampler with Queue about torchio HOT 3 OPEN

Hello

I am curious about the cases where it happen.
I work with full brain MRI segmentation and there the probability map are often very large. (ie number of possible patche center << NUM_SAMPLE

but let's imagine I want to focus on a small region with only five (connected) voxels. If I choose large enough patch size it seems to me that event the five possible distinct patches (each center on the five voxel of my cdf) will already be very similar. So taking only five patches will not solve the issue to have almost identical patches ... no ?

may be your proposition makes sense if the five voxel are spatially distinct ... but it looks weird to me to have single voxel regions ...

from torchio.

Hi,

The reason this came up is because I am using torchio for 2D samples, with each slice representing a patch in the classical sense. As a result, I sometimes have less available slices than the patches_per_subj set upfront. I am aware that this case is not usually encountered but it raises then the question: Why have that mechanism in place if it is not checked anyways? Why not remove the entire NUM_SAMPLES attribute alltogether if it is not set at all or at the wrong time?
For my application it makes sense to set the attribute but I understand this is not always the case. However, if the check and system is already there, why not use it in the way it was intended. I am happy to utilise my own classes that overwrite this functionality but the question still remains why the mechanism is there but not being used.

from torchio.

I was just questioning the use case, but I now better understand your's so it makes senses

from torchio.