Error in Compose of Array implementation of transforms (Spacing and Orientation)

Describe the bug Compose for Array implemented spacing transform and orientation transform are throwing error. This is primarily due to the reason that implementation returns more than one output which doesn’t work well as input for the next transform.

To Reproduce Code to reproduce the error (the error exists in both MONAI 3.x and 4.x versions):

import os
import glob
import numpy as np
import torch
import monai
from monai.transforms import *
from monai.utils import first
root_data_path = '<data path>'
resample_spacing = [1.83723155, 1.83723155, 2.42438879]
patch_size=[128, 128, 128]
train_images = sorted(glob.glob(os.path.join(root_data_path, 'imagesTr', '*.nii.gz')))
train_labels = sorted(glob.glob(os.path.join(root_data_path, 'labelsTr', '*.nii.gz')))
train_image_transform = Compose(
                                [LoadNifti(),  
                                Spacing(resample_spacing, mode='bilinear'), 
                                Orientation(axcodes = "RAS"), 
                                ScaleIntensity(), 
                                AddChannel(), 
                                ToTensor()])
train_label_transform = Compose(
                                [LoadNifti(),  
                                Spacing(resample_spacing, mode='nearest'), 
                                Orientation(axcodes = "RAS"), 
                                AddChannel(), 
                                ToTensor()])
train_files = [{"image": img, "label": seg} for img, seg in zip(train_images, train_labels)]
ds = monai.data.ArrayDataset(img = train_images, img_transform = train_image_transform, seg = train_labels, seg_transform = train_label_transform)
train_ds = monai.data.GridPatchDataset(ds, patch_size = patch_size)
train_loader = torch.utils.data.DataLoader(train_ds, batch_size=2, num_workers=1, pin_memory=torch.cuda.is_available())
im, seg = first(train_loader)

Expected behavior This should work seamlessly without throwing any errors. The same set of transform compose is working for dictionary implementation, but not for array implementation. This is because the dictionary implementation of Spacingd and Orientationd returns only one variable While Array implementation of Spacing and Orientation returns three arguments leading to the following error as could be seen in the screenshot below.

This required to primarily to generate patches for the datasets. I need to pass sliding window patches even during training as RandSpatialCrop or RandCropByPosNegLabel are not performing that well for my problem. I used the following code based on the tutorial for patch generation.

Screenshots

Environment

================================
Printing MONAI config...
================================
MONAI version: 0.4.0
Numpy version: 1.19.4
Pytorch version: 1.7.1
MONAI flags: HAS_EXT = False, USE_COMPILED = False
MONAI rev id: 0563a4467fa602feca92d91c7f47261868d171a1

Optional dependencies:
Pytorch Ignite version: NOT INSTALLED or UNKNOWN VERSION.
Nibabel version: 3.2.1
scikit-image version: NOT INSTALLED or UNKNOWN VERSION.
Pillow version: NOT INSTALLED or UNKNOWN VERSION.
Tensorboard version: NOT INSTALLED or UNKNOWN VERSION.
gdown version: NOT INSTALLED or UNKNOWN VERSION.
TorchVision version: NOT INSTALLED or UNKNOWN VERSION.
ITK version: NOT INSTALLED or UNKNOWN VERSION.
tqdm version: NOT INSTALLED or UNKNOWN VERSION.
lmdb version: NOT INSTALLED or UNKNOWN VERSION.
psutil version: NOT INSTALLED or UNKNOWN VERSION.

For details about installing the optional dependencies, please visit:
    https://docs.monai.io/en/latest/installation.html#installing-the-recommended-dependencies


================================
Printing system config...
================================
`psutil` required for `print_system_info`

================================
Printing GPU config...
================================
Num GPUs: 1
Has CUDA: True
CUDA version: 10.2
cuDNN enabled: True
cuDNN version: 7605
Current device: 0
Library compiled for CUDA architectures: ['sm_37', 'sm_50', 'sm_60', 'sm_70', 'sm_75']
Info for GPU: 0
        Name: Quadro RTX 5000
        Is integrated: False
        Is multi GPU board: False
        Multi processor count: 48
        Total memory (GB): 15.7
        Cached memory (GB): 0.0
        Allocated memory (GB): 0.0
        CUDA capability (maj.min): 7.5

[Nic-Ma]

Hi @architraj29 ,

Almost all the transforms in MONAI expect data shape CHWD, so you need to apply AddChannel transform before Spacing.

Thanks.

[Nic-Ma]

Hi @architraj29 ,

You are right, the GridPatchDataset can’t support dict data so far, we may need to enhance it later. For general Dataset, I think you can try below pseudo code:

images = sorted(glob.glob(os.path.join(root_dir, "im*.nii.gz")))
segs = sorted(glob.glob(os.path.join(root_dir, "seg*.nii.gz")))
train_files = [{"img": img, "seg": seg} for img, seg in zip(images[:20], segs[:20])]

train_trans = Compose([
    LoadImaged(keys=["img", "seg"]),
    AddChanneld(keys=["img", "seg"]),
    ScaleIntensityd(keys=["img"]),
    ToTensord(keys=["img", "seg"]),
])
ds = Dataset(data=train_files, transform=train_trans)
dataloader = DataLoader(ds, ...)

If you must use GridPatchDataset and Spacing & Orientation transforms, you can define a simple function to execute transforms instead of Compose them:

def exec_trans(data):
    img = LoadImage(image_only=True)(data)
    img = AddChannel()(img)
    img, affine = Spacing()(img)
    img = ScaleIntensity()(img)
    return img
...

Thanks.

Thanks.