gridpatchdataset->apply transformation?

[janetkok]

Is it possible to apply transformation for gridpatchdataset ?
and how to use dataloader to load gridpatchdataset cause gridpatchdataset has no len()

Thanks

[Nic-Ma]

Hi @wyli ,

Could you please help share some comments here?
I think you can show more usage of GridPatchDataset.

Thanks in advance.

[wyli]

perhaps we could now add support of patch-level transforms for the grid dataset? if @ericspod agrees I'll create a feature request

[ericspod]

The linked solution uses a previous data set as input so permits only whole image transforms. Let's add that feature request in case there is a need for patch-level operations.

[Nic-Ma]

Hi @ericspod ,

Maybe you can share more details about the usage of GridPatchDataset in research work?
Thanks in advance.

[ericspod]

There is an example of how to use it in the Nifti Read Example tutorial notebook, does this answer your questions?

[nveeturi2]

I tried the example # 6 from above link to create a data loader which yields patches in regular grid order from loaded images, but ran into issues. Could you please help?
MONAI version: 0.4.0+150.g401ea29
Numpy version: 1.19.5
Pytorch version: 1.7.1
Details below:
---> 29 im, seg = first(loader)
TypeError: Caught TypeError in DataLoader worker process 0.
Original Traceback (most recent call last):
File "/home/ubuntu/anaconda3/lib/python3.7/site-packages/torch/utils/data/_utils/worker.py", line 198, in _worker_loop
data = fetcher.fetch(index)
File "/home/ubuntu/anaconda3/lib/python3.7/site-packages/torch/utils/data/utils/fetch.py", line 28, in fetch
data.append(next(self.dataset_iter))
File "/home/ubuntu/anaconda3/lib/python3.7/site-packages/monai/data/grid_dataset.py", line 161, in iter
for patch, slices, * in self.patch_iter(image): # patch_iter to yield at least 2 items: patch, coords
TypeError: 'tuple' object is not callable
PS:
I was able to successfully use GridPatchDataset with PatchIter, but I need to crop image and mask image both as part of a transformation using Compose(i.e. instead of RandSpatialCropSamples with limited samples, I want to use all the patches created using GridPatchDataset) but ran into an issue with DataLoader.

[ericspod]

I think the patch_iter argument you're passing when constructing GridPatchDataset isn't correct. This has to be a callable which will return an iterator producing patches from the input image, such as PatchIter. It looks like what you've passed in doesn't meet this criteria.

[nveeturi2]

Thanks for the reply. In section 6 example, we dont pass patch_iter. The implementation using patchIter is working fine. Here's the code I tried:

imtrans = Compose([LoadImage(image_only=True), ScaleIntensity(), AddChannel(), ToTensor()]) segtrans = Compose([LoadImage(image_only=True), AddChannel(), ToTensor()]) ds = ArrayDataset(images, imtrans, segs, segtrans) ds = GridPatchDataset(ds, (64, 64, 64)) loader = torch.utils.data.DataLoader( ds, batch_size=10, num_workers=2, pin_memory=torch.cuda.is_available() ) im, seg = first(loader) print(im.shape, seg.shape) 

[ericspod]

It looks like the GridPatchDataset has changed since I mentioned the Nifti reading tutorial. I think GridPatchDataset(ds, PatchIter((64, 64, 64))) would work. We should also fix the example of course. @rijobro

[nveeturi2]

Thanks, I will try this and get back.