問題文
The training loss rises and falls violently from update to update and never settles. The same run behaves normally when one particular number is divided by 50. Which setting was wrong?
選択肢
- The setting that was wrong is step size of the optimizer.
- The setting that was wrong is initialization range of the optimizer.
- The setting that was wrong is penalty strength of the optimizer.
- The setting that was wrong is batch capacity of the optimizer.