5.1 Case study

The common protocol for reproducing the papers includes:

• paper with code available

  • clone the git repo and create a container with the same environment (ubuntu/python/tf/pytorch etc) as the README suggested

  • download the dataset they used for training and create a mini dataset for debugging

  • run on the whole training pipeline to make sure it is reproducible (compare the loss/metrics)

  • modified the data loader/loss func/#channel etc to make it compatible with the customized data we are working at

  • get a baseline

  • run training and record the experiments

• paper without code available

  • chose the paper in the fields you're familiar

  • build an MVP first (don't investigate too much time before you're convinced by the performance of the MVP)

  • change a module once a time