What did I do today

  • Found wrong photos (labelled wrong from the first place) -> revised and distributed so that the teammates could use
  • Found out that the model was transforming not only the train data, but also the validation data in ways that it was intended to make the inference harder. Changed this.
  • Ran several models
    • AdamW vs SGD -> similar result, AdamW faster
    • mutliplicative LR scheduler vs CosineAnnealingLR scheduler -> cosine much better
    • simpler transformation vs compound transformation -> couldn’t tell because of the problem mentioned above

Peer Session

Reflections

- Make a plan always!