Fix for patch commander

[JoeySpronck]

Fix for patch commander skipping patches that should be sampled on the borders. Also adds new optional functionality to not sample patches that do not have mask in the write-shape. This write shape is the shape to which the user will center crop the patches provided by the iterator for sliding window configs with overlap, where you crop off the overlapping borders

[JoeySpronck]

@martvanrijthoven Important fix here, happy to see it merged so I can rebuild my dockers and rerun inference.

[martvanrijthoven]

Hey Joey, thanks!

One test failed:
FAILED tests/patch_iterator_test.py::test_patch_iterator - assert 1026 == 900

Could you check that out? Please let me know if you have any questions

[JoeySpronck]

The issue

BrokenPipeError: [Errno 32] Broken pipe

    def test_patch_iterator():
        patch_configuration = PatchConfiguration(patch_shape=(1024,1024,3),
                                                spacings=(0.5,),
                                                overlap=(0,0),
                                                offset=(0,0),
                                                center=False)
    
        with create_patch_iterator(image_path='/tmp/TCGA-21-5784-01Z-00-DX1.tif',
                                patch_configuration=patch_configuration,
                                cpus=4,
                                backend='openslide') as patch_iterator:
    
            print(f"Number of patches {len(patch_iterator)}\n")
>           assert len(patch_iterator) == 900
E           assert 1026 == 900
E            +  where 1026 = len(<wholeslidedata.iterators.patchiterator.PatchBufferIterator object at 0x7f884fff3850>)

tests/patch_iterator_test.py:22: AssertionError

Comment

I think this is actually the issue Im trying to resolve, but I could make asome improvements.

The problem was that the iterator skipped tiles that should be sampled. For more complex configs many things could go wrong with the current setup.

What I solved

1. What happened now for example: the double for loop on rows and columns stopped once the border of the image was reached, but with center=True the next row or column (for example coordinates that exeed the dimensions only 1 pixel) would also still sample parts within the dimensions of the image. Therefore the max row and col values should be increased with half the patch_size if center=True. I also added the offset to the max value, but that im not 100% sure off.

2. Another issue though is that with a positive offset, you may skip patches that should be sampled left and above the first coordinate (with a negative offset likely the opposite). To be sure that multiple types of complex configs work as intended, I added a full additional patch on all sides (left, right, top, bottom). This will result in more patches in the messages, of which many will be ignored (not sampled) because the mask will be empty.

About this specific test

• The increased x and y max with 1) the offset and 2) 1/2 patch shape if center is true, didnt happen in this test because they are conditional on those variables.
• However, the additional step on all sides did happen
• Becasue there is no mask in the test the unnecessary patches are not filtered out.

Proposal

I will make the addition of the border conditional on whether there is offset or overlap, otherwise the additional border will be skipped.

Happy to hear your feedback

[JoeySpronck]

@martvanrijthoven tests now successful

[JoeySpronck]

quick heads up that im completely reworking it, the full loop should be reworked instead of adjusted

[JoeySpronck]

Ill close this PR and open a new one for the rework