--- How to train your Dragon ---

• Given the model architecture (implemented in PyTorch), candidates have to train the given model, and submit the weights.
• Host competition on kaggle for auto grading of submissions, and manually verify the scores/outputs of winners by their weights.
• Participants can explore
  • Different data augmentation techniques, including the promising ones like cutmix, mixin etc
  • Various optimizers & activation functions
  • Various ensemble strategies that kagglers use
  • Hacky tricks like Pseudo Labeling
  • Knowledge Distillation to further improve accuracy
• The main focus of the contest would be to give participants a hands on experience with PyTorch, and give some feeling about training neural nets (which we had by our 6th sem).
• The dataset of the contest could be an ensemble of existing image classification datasets out there
• We can purposefully make the dataset imbalanced, to give participants the feel of how this changes the accuracy.