pytorch_dsb2018

• This is my Pytorch code used for the Data Science Bowl 2018 Challenge. However, the boilerplate code is fairly general and can be used as a template for deep networks on images.
• Implemetation includes:
  • Setting up models, optimizers, and schedulers
  • Loading and saving of checkpoints
  • Logging of metrics
  • Run seamlessly on CPU or GPU
  • Multi-objective learning
  • Instance and class weights
  • A refined plateau learn rate scheduler
  • A custom img_aug augmentor for random 5-crops
  • An implementation of group-normalization
  • Plotting and comparing of learning curves
  • Miscellaneous image pre/postprocessing functions
  • IOU (intersection-over-union) calculations
  • Code for failure mode analysis
  • A statistics class to incrementally compute average, std deviations, min and max