Loading single channel image with PILReader

[vsk-phi]

Hello,
I am trying to load a PNG image. I want to ensure that at load time, the image should be single channel. I am using something like:

transforms = Compose([
        LoadImaged(keys="image", image_only=True, reader=PILReader(converter=lambda image: image.convert("L"))),
        # LoadImaged(keys="image", image_only=True, ),        
                              EnsureChannelFirstd(keys="image"),
                              Resized(keys="image", spatial_size = (IMG_X, IMG_Y)),
                              ScaleIntensityd(keys="image")])

Unfortunately, I am getting the problem:

> collate dict key "image" out of 2 keys
>> collate/stack a list of tensors
> collate dict key "label" out of 2 keys
Traceback (most recent call last):
  File "C:\Users\3201\AppData\Local\miniforge3\lib\site-packages\torch\utils\data\_utils\collate.py", line 160, in default_collate
    return elem_type({key: default_collate([d[key] for d in batch]) for key in elem})
  File "C:\Users\3201\AppData\Local\miniforge3\lib\site-packages\torch\utils\data\_utils\collate.py", line 160, in <dictcomp>
    return elem_type({key: default_collate([d[key] for d in batch]) for key in elem})
  File "C:\Users\3201\AppData\Local\miniforge3\lib\site-packages\torch\utils\data\_utils\collate.py", line 183, in default_collate
    raise TypeError(default_collate_err_msg_format.format(elem_type))
TypeError: default_collate: batch must contain tensors, numpy arrays, numbers, dicts or lists; found <class 'NoneType'>

I am using the option image_only=True above, so as to avoid the problem with metadata.
However, if I do not explicitly specify the reader, then the transform works without a hitch. In other words,
LoadImaged(keys="image", image_only=True, reader=PILReader(converter=lambda image: image.convert("L"))), does not work while
LoadImaged(keys="image", image_only=True), works in the above transform.

Monai version is 0.9.1.
Any help/suggestion appreciated.

[wyli]

the error message suggests collate dict key "label" out of 2 keys, but your script example doesn't include "label", could you confirm? I can look into this if you provide a complete working example.

[vsk-phi]

hello,
thank you for reply. Here is a more working code snipper of a file say data_something.py

from monai.transforms import (
    Activations,
    EnsureChannelFirst,
    AsDiscrete,
    Compose,
    LoadImage,
    RandFlip,
    RandRotate,
    RandZoom,
    ScaleIntensity,
    Resize,
    LoadImaged,
    EnsureChannelFirstd,
    Resized,
    RandFlipd,
    RandRotated,
    RandZoomd,
    ScaleIntensityd,
)

def get_file_names_labels(data_dir):
    class_names = sorted(x for x in os.listdir(data_dir) 
                     if os.path.isdir(os.path.join(data_dir, x)))
    num_class = len(class_names)
    image_files = [
        [
            os.path.join(data_dir, class_names[i], x)
            for x in os.listdir(os.path.join(data_dir, class_names[i]))
        ]
        for i in range(num_class)
    ]
    num_each = [len(image_files[i]) for i in range(num_class)]
    image_files_list = []
    image_class = []
    for i in range(num_class):
        image_files_list.extend(image_files[i])
        image_class.extend([i] * num_each[i])
    num_total = len(image_class)
    image_width, image_height = PIL.Image.open(image_files_list[0]).size

    print(f"Total image count: {num_total}")
    print(f"Image dimensions: {image_width} x {image_height}")
    print(f"Label names: {class_names}")
    print(f"Label counts: {num_each}")
    val_frac = 0.0
    test_frac = 0.0
    length = len(image_files_list)
    indices = np.arange(length)
    np.random.shuffle(indices)

    test_split = int(test_frac * length) # not used
    val_split = int(val_frac * length) + test_split # not used
    train_indices = indices[val_split:]

    data = [{"image":image_files_list[i], "label": image_class[i]} for i in train_indices]
    return data, num_class

train_data, num_class = get_file_names_labels(train_dir) # please set this as per your location
IMG_X = 128
IMG_Y = 128
    
train_transforms = Compose(
    [
        LoadImaged(keys="image", image_only=True),
        EnsureChannelFirstd(keys="image"),
        Resized(keys="image", spatial_size = (IMG_X, IMG_Y)),
        ScaleIntensityd(keys="image"),
        RandRotated(keys="image", range_x=np.pi / 6, prob=0.3, keep_size=True),
        RandFlipd(keys="image", spatial_axis=0, prob=0.2),
        RandZoomd(keys="image", min_zoom=0.9, max_zoom=1.1, prob=0.2),
    ]
)
train_ds = monai.data.Dataset(train_data, train_transforms)
train_loader = DataLoader(train_ds, batch_size=batch_size, shuffle=False, num_workers=0) # on windows
print("Starting to iterate over training datasets. Please wait...")
for t in train_loader:
    pass
print("Iterating over training dataset passed.")

The above code works. However if i replace the image loader as

LoadImaged(keys="image", image_only=True, reader=PILReader(converter=lambda image: image.convert("L"))),

the code fails with the error mentioned in the first post. As you can see, a lot of code is modified from the examples!
hopefully the above helps to reproduce. I have checked on the MedNist dataset.
please let me know if you need some more information

[vsk-phi]

Sorry for the ugly formatting but i am unable to get it correctly formatted... :-(

[wyli]

thanks, that looks like a bug, I submit a pr to fix it

[vsk-phi]

thank you.