(Optional) state_dict for each transform (reproducibility) about torchio HOT 2 CLOSED

hello
I am not sure to understand, which exact state you need to save:

Let say you have a RandomAffine transform instantiate with scales=1 degrees=20 translations=0
so each time you have a new data you get a new rotation.

What do you want to save ?
the exacte rotation value of the last transform ? (what for then ?)

If the objective is to start the training from where you stopped, I do not see the point the save the last used transform state (ie rotation value) if you need to continue training you can just used the same RandomTransform you used to start with ... (because it is random ....) no ?

but may be I miss something here

from torchio.

Hi,

I think you might be right! I was thinking that the transform could differ between subjects e.g. w.r.t. to in_min_max for some intensity transforms but as long as the initial list of subjects and the SubjectsDataset can be restored (by just initialising the same transforms and the same subjects), I can restore my state.

This feature request can be therefore closed (I think)!

from torchio.