Unnecessary warning when using next(iter(SubjectsLoader))

[lorinczszabolcs]

Is there an existing issue for this?

• I have searched the existing issues

Related issues, but not completely the same: #1247 #1250

Bug summary

I have some custom code where I iterate through my dataloader with next(iter(loader)), and it shows a warning message, which I believe is unnecessary, as I have changed my code to use a tio.SubjectsLoader.

Code for reproduction

import torchio as tio
from torch.utils.data import Dataset, DataLoader
import numpy as np
import tempfile
import nibabel as nib

# Temporary directory to simulate input files
temp_dir = tempfile.TemporaryDirectory()

# Create example 3D image and label NIfTI files
def create_nifti(data, filename):
    affine = np.eye(4)  # Identity matrix for affine
    nifti_img = nib.Nifti1Image(data, affine)
    filepath = f"{temp_dir.name}/{filename}"
    nib.save(nifti_img, filepath)
    return filepath

# Simulated 3D image and label data
image_data = np.random.rand(10, 10, 10).astype(np.float32)
label_data = np.random.randint(0, 2, (10, 10, 10)).astype(np.uint8)

# Save them as NIfTI files
image_path = create_nifti(image_data, "image1.nii.gz")
label_path = create_nifti(label_data, "label1.nii.gz")

# Create a list of subjects with file paths
subject_list = [
    {
        'image': image_path,
        'label': label_path,
    }
]

# Define a transform pipeline
transform = tio.Compose([
    tio.ToCanonical(),  # Standardize orientation
    tio.RescaleIntensity((0, 1)),  # Normalize intensity to 0-1
])

# Custom Dataset that implements __getitem__
class CustomTorchDataset(tio.SubjectsDataset):
    def __init__(self, subject_list, transform=None):
        self.subject_list = subject_list
        self.transform = transform

    def __len__(self):
        return len(self.subject_list)

    def __getitem__(self, idx):
        subject_dict = self.subject_list[idx]
        subject = tio.Subject(
            image=tio.ScalarImage(subject_dict['image']),
            label=tio.LabelMap(subject_dict['label']),
        )
        if self.transform:
            subject = self.transform(subject)
        return {
            'image': subject['image'].data,
            'label': subject['label'].data,
        }

# Initialize the custom dataset
dataset = CustomTorchDataset(subject_list, transform=transform)

# Create a DataLoader
loader = tio.SubjectsLoader(dataset, batch_size=1, shuffle=False)

# Fetch a batch
batch = next(iter(loader))

# Print the batch keys and their shapes
for key, value in batch.items():
    print(f"{key}: {value.shape}")

# Cleanup temporary files
temp_dir.cleanup()

Actual outcome

image: torch.Size([1, 1, 10, 10, 10])
label: torch.Size([1, 1, 10, 10, 10])

Error messages

/usr/local/lib/python3.11/dist-packages/torchio/data/image.py:248: UserWarning: Using TorchIO images without a torchio.SubjectsLoader in PyTorch >= 2.3 might have unexpected consequences, e.g., the collated batches will be instances of torchio.Subject with 5D images. Replace your PyTorch DataLoader with a torchio.SubjectsLoader so that the collated batch becomes a dictionary, as expected. See https://github.com/fepegar/torchio/issues/1179 for more context about this issue. warnings.warn(message, stacklevel=1)

Expected outcome

I expect to not receive such a warning, given that I am using a SubjectsLoader and SubjectsDataset.

System info

Platform:   Linux-5.15.0-122-generic-x86_64-with-glibc2.35
TorchIO:    0.20.1
PyTorch:    2.4.1
SimpleITK:  2.4.1 (ITK 5.4)
NumPy:      1.26.4
Python:     3.11.0 | packaged by conda-forge | (main, Jan 14 2023, 12:27:40) [GCC 11.3.0]

[fepegar]

Thanks for reporting, @lorinczszabolcs. Unfortunately I won't be able to pick this up any time soon. Feel free to give it a try.

[nicoloesch]

Hi @lorinczszabolcs,

I am pretty sure that warning message arises due to the wrong/unexpected instantiation of your subjects. Generally speaking, a tio.SubjectsDataset should be instantiated with tio.Subject instances instead of the list of dicts you provide with

# Create a list of subjects with file paths
subject_list = [
    {
        'image': image_path,
        'label': label_path,
    }
]

The expected usage is outlined in the official documentation.

The reason why you now get an error message arises from the call batch = next(iter(loader)). Internally, the tio.SubjectsLoader gets converted to either a torch.utils.data.dataloader._SingleProcessDataLoaderIter or torch.utils.data.dataloader._MultiProcessingDataLoaderIter depending on the num_workers. This is just how PyTorch works and differentiates between loading data in a single stream or in parallel (also does not really matter as both result in True in your specific case in torchio.utils.in_torch_loader). As you create a tio.Subject dynamically in __getitem__ of your dataset, the stack already contains the _SingleProcessDataLoaderIter/_MultiProcessingDataLoaderIter as the tio.SubjectsLoader has already been converted under the hood and we iterate over it using next(iter()), which then calls your CustomTorchDataset.__getitem__. The respective check in the instantiation of any tio.Image then flags it and provides you with the warning - this is expected behaviour and informs you of a potential mismatch in what PyTorch > 2.3.x expects.

If you create your tio.Subjects first (you can provide paths to them - they will then be loaded accordingly), put them in a list and provide it to the tio.SubjectsDataset upon instantiation instead of dynamically in __getitem__, the warning vanishes.

See following MWE:

import torch
import torchio as tio

def generate_subject(tensor_size: tuple[int, ...]) -> tio.Subject:
    subject = tio.Subject(
        image=tio.ScalarImage(tensor=torch.randint(0, 2**8,tensor_size)),
        label=tio.LabelMap(tensor=torch.randint(0, 2, tensor_size)),
    )
    return subject

def debug():
    tensor_size = (1,240,240,155)
    num_subjects = 2

    transform = tio.Compose([
        tio.ToCanonical(),  # Standardize orientation
        tio.RescaleIntensity((0, 1)),  # Normalize intensity to 0-1
    ])


    dataset = tio.SubjectsDataset([generate_subject(tensor_size=tensor_size) for _ in range(num_subjects)], transform=transform)
    loader = tio.SubjectsLoader(dataset, batch_size=1)
    for batch in loader:
        print(batch)
        break


if __name__ == '__main__':
    debug()

Hope this helps and we can close the Bug.

Edit: Code formatting for better readability

[lorinczszabolcs]

Thanks for investigating!

Yes, issue can be closed, I somehow missed the part that we need to instantiate with list of Subjects. Thanks for the heads-up!