Applying dictionary transforms to a list of tensors

[guillaume-troadec]

Hi,

I am wondering if there is a good way for me to apply dictionary transforms to a list of tensors.

I am using a pattern inspired by in this tutorial notebook (see below).

Here is an example of my current code. The motivation behind this is that I want the non-deterministic transforms (e.g. RandFlipd) to be applied with the same random seed to the corresponding input and target tensors.

transforms = Compose(
    [
        RandFlipd(keys=["input", "targets"], prob=0.1),
        Spacingd(keys=["input", "targets"], pixdim=[5, 5, 5]),
    ]
)

# tensors are currently monai metatensors 
data = {"input": input_tensor, "targets": target_tensor}
result = transforms(data)

The main crux of my problem is that I wish to be able to run this code with this line instead: data = {"input": input_tensor, "targets": [target_tensor_a, target_tensor_b]}, that is, with an arbitrary number of targets.

However I get an AttributeError because dictionary transforms cannot currently handle non-tensor values

Does anyone have any suggestions on the proper pattern to apply dictionary transforms to a list of tensors? A hacky way would be to flatten out the values but ideally I'd like to avoid this - is there a better way to go about it? Thanks!

[KumoLiu]

Hi @guillaume-troadec, I think you could do something like:

transforms = Compose(
    [
        RandFlipd(keys=["input", "targets_a", "targets_b"], prob=0.1),
        Spacingd(keys=["input", "targets_a", "targets_b"], pixdim=[5, 5, 5]),
    ]
)

data = {"input": input_tensor, "targets_a": target_tensor_a, "targets_b": target_tensor_b}
result = transforms(data)

For more details, you can refer to https://github.com/Project-MONAI/tutorials/blob/main/modules/transforms_metatensor.ipynb.

Hope it helps, thanks!

[guillaume-troadec]

Hi @KumoLiu , thanks for your help!

I did consider this, but I was hoping to perhaps avoid having an arbitrary number of keys (it would be easier for downstream code if everything was under the targets key). Do you have any suggestions on where to go from here?

[KumoLiu]

I don't think we can specify a list for each key, it depends on our implementation. I think you can also reorganize it into one key after the transforms.

MONAI/monai/transforms/spatial/dictionary.py

Line 505 in 50e66fa

_init_shape, _pixdim = d[key].peek_pending_shape(), d[key].pixdim

result["target"] = [result["targets_a"], result["targets_b"]]

[wyli]

you can also consider monai.transforms.apply_transform, something like:

result = apply_transform(transforms, data, map_items=True)

MONAI/monai/transforms/transform.py

Lines 101 to 115 in 50e66fa

	def apply_transform(
	transform: Callable[..., ReturnType],
	data: Any,
	map_items: bool = True,
	unpack_items: bool = False,
	log_stats: bool | str = False,
	lazy: bool | None = False,
	overrides: dict | None = None,
	) -> list[ReturnType] | ReturnType:
	"""
	Transform `data` with `transform`.
	If `data` is a list or tuple and `map_data` is True, each item of `data` will be transformed
	and this method returns a list of outcomes.
	otherwise transform will be applied once with `data` as the argument.

[guillaume-troadec]

Thank you both - I'll give that a go :)