Promote Child Items to Top Level

[bhashemian]

The MONAO dict-based transforms acts on the top level keys, for instance data={"image": ..., "pred": ...}; however, when the output is nested (like in the hovernet model), the actual data will lie one level below like data={"image": ..., "pred": {"a":..., "b":...}}.

Currently, if user want to apply transform on "a" or "b", it can use Lambdad and apply that transfrom on data["pred"] instead and use "a" and "b" as keys. However, if the user want to apply a transform on "image" and "a" for instance, I didn't find any transform that can enable it.

Here a transform that can promote these children items ("a" and "b") to the same level of "image" can solve the problem.

[myron]

Hi, I'm not familiar with this model, but i'm just curious at which place do we get the output in this form "pred": {"a":..., "b":...}} during loading pipeline or post-processing after prediction? I don't think any loading, augmentation transforms produce such format. or if the model produces such output, can we simple reformat the output of the model? thank you

[bhashemian]

Hi @myron, we got this in the output of HoVerNet model.

{"pred": {"horizontal_vertical": ..., "nucleus_prediction": ..., "type_prediction": ...}

MONAI/monai/networks/nets/hovernet.py

Lines 581 to 586 in 77651ad

	output = {
	HoVerNetBranch.NP.value: self.nucleus_prediction(x, short_cuts),
	HoVerNetBranch.HV.value: self.horizontal_vertical(x, short_cuts),
	}
	if self.type_prediction is not None:
	output[HoVerNetBranch.NC.value] = self.type_prediction(x, short_cuts)

Changing that will cause backward incompatibilities internally and externally.

[myron]

I see, thank you. then if you have access to model output in the python code, you could directly add those keys "horizontal_vertical", 'nucleus_prediction".. to you main dict
e.g. batch_data.update(output["pred])
this would save you from writing another monai transform.
but, I guess you don't have access to it in python code, then yes, you'd need a new transfor . I think there is a similar in-spirit transform splitdim https://docs.monai.io/en/stable/transforms.html#splitdim, but that only split a tensor dimension to create new key, value pairs.