Top participants are invited to submit their notebooks and it is stated “The judges will run a submitted algorithm on an AWS SageMaker g5.12xlarge instance, and it must run within 24 hours”.
Does the 24hrs limit on AWS include training? or is it only inference?
note that some solution like ensemble, 3d solution, or generative GAN/PDDM takes a long time to train
While it’s meaningless to argue about it now, we have always considered this computational resource limitation to be about training, since in previous competitions we only had to submit training code for retraining and scoring on an external validation set.
So we have designed a lot of network lightweighting and speedups for this computational resource limit, which of course greatly limits our model performance…