using a GridSampler with a Queue

Right now this doesn’t exactly make sense, e.g. you need to pass samples_per_volume to Queue which is not quite coherent, and it’s not clear how to pass arguments (specifically patch_overlap) to the GridSampler.

(Use case: I want my validation to be done deterministically and from the entire image which is however too big to be segmented in a single pass, even without gradients. I guess there are also some minor issues with computing validation metrics on the patches rather than after stitching the patches but this is a start.)

[fepegar]

If you don’t need to assemble the volume, you can still use the code from the notebook minus the aggregator bits. Here’s some pseudocode:

metrics = {}
for each sample in validation dataset:
    get tensor from sample
    sampler = grid_sampler(tensor)
    metrics_subject = []
    for patch in sampler:
        patch_x, patch_y = prepare_patch(patch)
        metric = get_metric(patch_x, patch_y)
        metrics_subject.append(metric)
    metrics[subject] = metrics_subject

---

After thinking a bit more and looking at the code, I think the problem with the snippet above is that the GridSampler should take a sample as input, not just a tensor, like the other samplers. I’ll see what I can do and come up with a solution.

[allcontributors[bot]]

@fepegar

I’ve put up a pull request to add @bcdarwin! 🎉