Fabian Isensee
Fabian Isensee
Exactly as @Chasel-Chen explains it! The pseudo dice is just to see the training progress and to determine whether overfitting is a problem (green line goes down = bad!). It...
There is no way of 'fixing' this, please just take it as it is! It is not supposed to be the same as the validation dice. Running actual validations would...
EMA stands for exponential moving average. This tracks the pseudo dice following a `updated = old * alpha + (1 - alpha) * new` formula. This is the smoothed green...
they are random patches from the validation set, sampled according to the same rules as the training patches. Dice is computed over all sampled patches (pretending all patches together are...
Seems like 1) your epoch time is really slow, see https://github.com/MIC-DKFZ/nnUNet/blob/master/documentation/benchmarking.md 2) you seem to suffer from some overfitting 3) one class is not learned properly What could concretely be...
It looks like you created a class that has the same name as the nnUNetTrainer but does not inherit from it? When making changes, please remember to create trainer classes...
@constantinulrich can you take a look? You are using this feature extensively
Please install pytorch as specified in the installation instructions. Please use the most recent version with the highest available version of CUDA. I recommend using a conda environment. Triton will...
Hey, really cool stuff! I like the idea of bitwise encoding, but there is a much easier solution: just make the dtype of the regions bool. This is now the...