RandSpatialCropSamples seems to output a strange result

[patriciacs1994]

Hello,

I have a 3D image and I want to select a number of random patches with a shape (96,96,96). So I decided to use RandSpatialCropSamples. Here you can see my piece of code:

patch_size = 96
train_imtrans = Compose(
    [
        LoadImage(image_only=True),
        ScaleIntensity(),
        AddChannel(),
        RandSpatialCropSamples((96, 96, 96),10, random_size=False), 
        ToTensor(),
    ]
)
train_segtrans = Compose(
    [
        LoadImage(image_only=True),
        AddChannel(),
        RandSpatialCropSamples((96, 96, 96),10, random_size=False), 
        ToTensor(),
    ]
)
val_imtrans = Compose([LoadImage(image_only=True), RandSpatialCropSamples((96, 96, 96),10, random_size=False), ScaleIntensity(), ToTensor()])
val_segtrans = Compose([LoadImage(image_only=True),RandSpatialCropSamples((96, 96, 96),10, random_size=False), ToTensor()])


# create a training data loader
train_ds = ArrayDataset(train_img_list[0], train_imtrans, train_img_list[1], train_segtrans)
train_loader = DataLoader(train_ds, batch_size=2, num_workers=8, pin_memory=torch.cuda.is_available())

# create a validation data loader
val_ds = ArrayDataset(val_img_list[0], val_imtrans, val_img_list[1], val_segtrans)
val_loader = DataLoader(val_ds, batch_size=1, num_workers=4, pin_memory=torch.cuda.is_available())

for batch_data in train_loader:
    inputs, labels = batch_data[0], batch_data[1]

My problem is in the inputs and labels parameters. both have a size of (2,1,96,96,96), but when I plot them, both inputs and labels are pieces of the original image and labels should be the correspondent piece of GT image but it is not. I realized that batch_data has an output size of (num_samples*2). Am I doing something wrong or is there an error in the function?

NOTE: train_image_list and val_img_list are lists of lists of strings. with size (2,56)

Thank you in advance for your help

[Nic-Ma]

Hi @patriciacs1994 ,

May I know what DataLoader you are using?
If you use the torch.utils.data.DataLoader, please try with monai.data.DataLoader, we have enhanced collate_fn callback.

Thanks.

[patriciacs1994]

Hi @Nic-Ma
I was using torch.utils.data.DataLoader.
How I can use collate_fn?

Thank you in advance.

[rijobro]

Hi @patriciacs1994, I think that by using ArrayDataset, you might be making life slightly difficult for yourself. I would advise switching to using our normal Dataset (or CacheDataset etc), and using a dictionary of data, with dictionary transforms. We do this in most of our segmentation tutorials, for example this brats one.

The problem you're seeing here is sneaky. If you executed len(batch_data), I think it would return 20. This corresponds to two input types (image and label) and 10 samples produced by RandSpatialCropSamples. Each of the elements would then have batch size of 2, making things even messier. I think if you were to take batch_data[i] and batch_data[i+10], they would correspond with each other. This is why @Nic-Ma suggested playing with the collation function. However, I think it would be simpler in your case to just switch to using a dictionary of data.

[rijobro]

Just opened a related issue: #2111.

[patriciacs1994]

Thank you for your answers.
Changing the format of the dataset to a dictionary solves all the problems.