The data loaders most likely are pretty inefficient at the moment. Let's try to profile and optimise GPU utilisation.
Potential problems atm:
- data is concatenated from potentially many data loaders into a big batch
- distmap (for boundary loss) is calculated in the data loader
- one hot encoded representations are created multiple times
- tensors are potentially moved from gpu to cpu and back multiple times (plus some bumpy operations)
The data loaders most likely are pretty inefficient at the moment. Let's try to profile and optimise GPU utilisation.
Potential problems atm: