iter_patch memory consumption

[wyli]

Discussed in #5610

^{Originally posted by schellenchris November 30, 2022}
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]

there's additional issue if the padding mode=None and copy_back=True:

MONAI/monai/data/utils.py

Lines 312 to 313 in d0db5fd

	slices = tuple(slice(p, p + s) for p, s in zip(patch_size_, arr.shape))
	arr[...] = arrpad[slices]

the offset only works with padding