Importance Sampling

[pavan-tummala-tcgls]

Hello,

I had a question on how I could improve the performance of nnUNet on certain kind of training data or weigh them more in general during training so I was wondering if I could assign higher weights so instead of sampling uniformly from the training dataset, it can use the assigned weights to sample more frequently from important examples.

Thank You.

[ZxnSnowy]

I have the same problem, like there are 200 image.nii.gz, 50 of which I want one or more to appear in the batch every time. How should I deal with it?

[GregorKoehler]

Hi @pavan-tummala-tcgls and @ZxnSnowy,
as far as I can tell, nnUNet does not offer configuration options to oversample certain samples out of the box. So I'd propose you implement your own DataLoader class which implements this. Take a look at the classes in: https://github.com/MIC-DKFZ/nnUNet/tree/master/nnunetv2/training/dataloading
As a pointer, the best choice is to adapt the get_indices method. Here's the default get_indices method in batchgenerators' DataLoader class, from which the nnUNet DataLoaders inherit:
https://github.com/MIC-DKFZ/batchgenerators/blob/f53025bd05fe9642133ac539a0df3559d82415b4/batchgenerators/dataloading/data_loader.py#L185-L213

[pavan-tummala-tcgls]

Thanks for the info. Will try this out.