Feature request: data augmentation on GPU

[alexandrebone]

Dear nnUNet team / MIC-DKFZ members,

Thank you very much for your amazing work.

I have recently tried nnUNet v2. I was wondering if you may consider, in future releases, to offer the possibility to perform data augmentation on the GPU, for instance leveraging the Kornia or Monai libraries instead of batchgenerators.
From my point of view, this may relieve the hardware constraints (CPU/GPU balancing) that come with nnUNet and facilitate its further dissemination in the community. In particular, I find difficult to accommodate that the optimal hardware architecture varies (greatly) with the number of input modalities.

Thank you again, best regards,
Alex

[dojoh]

Dear Alex,

thanks for the feedback. I will bring this idea to Fabians attention.

Best regards,
Ole