Is your feature request related to a problem? Please describe.
I was looking to train a 2D segmentation network on all the slices of a 3D volume. Currently, the GridPatchDataset is a bit odd to use? I get a few errors from just using the utils.misc.first function

grid_ds = monai.data.GridPatchDataset(check_ds_array, patch_size = (160, 160, 1))
check_loader = monai.data.DataLoader(grid_ds, batch_size=1)
check_data = monai.utils.misc.first(check_loader)

AttributeError: 'str' object has no attribute 'shape'

There also is no current getitem function for GridPatchDataset.

Describe the solution you'd like
Ideally it would be a transform that would take in an image of size (160, 160, N) and generate N arrays of size (160, 160).

Describe alternatives you've considered
Currently, there are issues #299 and #606 discussing random patch style datasets, both of which have been helpful. Issue 299 however seems to only take 1 single 2D sample from each input volume, instead of taking N samples along the Z direction.

In issue 606 you mention RandSpatialCropSamples which I have tried, but my concern is that it may re-sample the same random_center multiple times. A version or mode of this function that can generate samples in order, or a mode that can ensure each center is only randomly selected once would be great!

Thanks!
Jillian