iter_patch memory consumption

[schellenchris]

Hello,

I was playing around with the GridPatchDataset / training with smaller Slices than the original volume, when I ran into some weird RAM OOM problems. After some investigation i noticed that line 290 in data/utils.py cause the problem:

MONAI/monai/data/utils.py

Lines 288 to 290 in d0db5fd

	# pad image by maximum values needed to ensure patches are taken from inside an image
	if padded:
	arrpad = np.pad(arr, tuple((p, p) for p in patch_size_), look_up_option(mode, NumpyPadMode).value, **pad_opts)

is there a reason to pad the image that much (especially padding the channels doesn't make sense for me)?
I managed to reduce the memory by 15! GB by modifying the code a bit and just padding the axis slices are generated from. I was just wondering if I am missing out on something.

[wyli]

thanks for reporting, I create an issue to track it #5611

[wyli]

this is addressed by #5612 by no padding when the corresponding dimension of the patch size is set to None or 0.

[schellenchris]

works like a charm, no memory problems anymore. thank you!