The issue keeps track on developing epistemic scoring for monaibundle app.
[x] Create a scoring inferTask generater, BundleScoringInferer the scoring infer task should instantiate bundle models with dropout, if the model do not support setting dropout, skip the scoring method. Users can set dropout rate.
[x] Add init_scoring_methods to the monaibundle app, corner case check if dropout is not supported. Check if the epistemic scoring can't be created, fall back to no scoring method if so.
[x] Enable --conf epistemic_model <bundle1,bundle2,etc>, so that users can choose which scoring models to use.
[x] Test the feature with spleen_ct_segmentation_v0.1.0 bundel.
[x] Raise the PR as draft version
[x] PR for Review
Need to discuss here:
I'm seeing the dropout parameter in different bundle models are named differently, such as: UNet, FlexibleUNet use "dropout" as the argument, SwinUNETR uses "drop_rate", UNETR uses "dropout_rate", VNet uses "dropout_prob", and some do not support dropout.
How will we handle this argument difference?
you probably don't need this Create a scoring inferTask generater, BundleScoringInferer
Lets discuss tomorrow.. want to understand how you are planning to solve
The issue keeps track on developing epistemic scoring for monaibundle app.
BundleScoringInferer
the scoring infer task should instantiate bundle models with dropout, if the model do not support setting dropout, skip the scoring method. Users can set dropout rate.--conf epistemic_model <bundle1,bundle2,etc>
, so that users can choose which scoring models to use.Need to discuss here: I'm seeing the dropout parameter in different bundle models are named differently, such as: UNet, FlexibleUNet use "dropout" as the argument, SwinUNETR uses "drop_rate", UNETR uses "dropout_rate", VNet uses "dropout_prob", and some do not support dropout. How will we handle this argument difference?