[AI Boostcamp Day 21] Training, Training and Training

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!