フリー問題

問題文

A team has one dataset directory and needs separate training and validation pipelines: augmentation only for training, but the same normalization for both. Which arrangement fits?

選択肢

1. Apply the augmentation inside the collate function for the training loader.
2. Apply augmentation to both and average the validation results.
3. Build two transform pipelines that share the normalization step and give each one to its own dataset instance.
4. Use a single dataset instance and switch its transform attribute at the start of each phase, which is the documented approach because it guarantees that the two phases never see inconsistent preprocessing.