questions about image shape

[JIBSN]

Hi,
I am now training micro-sam with a custom dataset. The shape of my raw imags is HxWx3 and the label's shape is HxW. I found an error "File "/home/jovyan/.conda/micro-sam/lib/python3.10/site-packages/torch_em/data/segmentation_dataset.py", line 61, in init
assert shape_raw == shape_label, f"{shape_raw}, {shape_label}"
AssertionError: (2048, 3), (2048, 2048)".
I am confused with this error. Please give me some tips why it happens and how to solve it.

[anwai98]

Hi @JIBSN,
From the looks of the error, I am assuming that you are making use of the torch_em.default_segmentation_loader (as for example, here). In this case, the segmentation dataset's expectation is to have channels first for the inputs (i.e. shape of the image should be 3 x H x W). You can use a raw transform to transpose the axes to make use of it, for example:

def custom_raw_transform(raw):
    "Chaning the axes for the input images from H x W X 3 -> 3 x H x W"
    raw = raw.transpose(2, 0, 1)
    return raw

loader = torch_em.default_segmentation_loader(..., raw_transform=custom_raw_transform, ...)