RandCropByPosNegLabeld self.center mix for multi-process and dataset different shape

[fgbxx]

hi, all
I found a bug when I use multi-process for different shape dataset.
the transform is global , and the self.center will be changed during different processes, so change it to a local parameter in call function is better solution I think.

Best,
Gang in Beijing

[wyli]

If the inputs have different shapes, perhaps you should create separate RandCropByPosNegLabeld instances and set different keys respectively. I'm moving this to the discussion tab.

[tmathai]

@wyli Maybe the spleen/brats 3D segmentation tutorial(s) should be updated to reflect this issue of a randomized transform every time a data point from the data loader is called upon? This would prevent others from spending 3 days thinking that there is a bug somewhere 🤣

[rijobro]

Hi @tmathai, which line(s) of the tutorial do you find confusing? I don't see anywhere in which the transforms are called twice (once for the image and then once for the label).

[tmathai]

@rijobro In the BRATS 3D segmentation tutorial, specifically at box 7 (In 7) called Check data shape and visualize, there are two instances where data is extracted: 1. print statement and in the plt.imshowstatement. That was where I grabbed the code to extract the data point ( image/label) after transformations and plot the masks over the image, and my code above shows that.

[rijobro]

@tmathai thanks for point it out. In this case it works since there are no random validation transforms, but you're right that this should be simplified.

[rijobro]

PR here: Project-MONAI/tutorials#549