Convert cropped images/labels to individual samples

[Fredf10]

Hi,

I am using monai and monai transforms together with pytorch_lightning, and have an issue when generating the training data using transforms.RandCropByPosNegLabeld with more than one samples. This results in a dataset where each item (i.e. sample) is a list with length equal to num_samples specified in transforms.RandCropByPosNegLabeld. I wonder if anyone knows how I can reshape the generated dataset such that eached cropped image becomes and individual sample? i.e. in the example below I would like that len(dataset) = len(data_list)*num_samples, rather than len(dataset) = len(data_list) and where each item in the dataset is a list with n=num_samples.

############################################################################################################

from monai import data, transforms
from torch.utils.data import DataLoader

data_list = [{'image': './data/training_sample_1/imaging.nii.gz',
'label': './data/training_sample_1/segmentation.nii.gz'}]

transform_dict = transforms.Compose([transforms.LoadImaged(keys=["image", "label"], ensure_channel_first=True),
transforms.RandCropByPosNegLabeld(keys=["image", "label"],
label_key="label", spatial_size=(160, 160, 160),
pos=3, neg=1, num_samples=2, image_key="image",
image_threshold=0),
transforms.ToTensord(keys=["image", "label"], track_meta=False)])

dataset = data.CacheDataset(data=data_list,
transform=transform_dict,
cache_rate=1.0,
num_workers=1)
training_sample = dataset[0]
print(type(training_sample), len(training_sample))

training_data = DataLoader(dataset)

[KumoLiu]

Hi @Fredf10, your question looks similar to this one #6242.

Hope it can help you, thanks!

[Fredf10]

Hi @KumoLiu, you are correct, this question looks similar, and PatchDataset does indeed seem to work.
Thanks for your help!