試験
通常
試験時間: 00:00問題時間: 00:00
問題一覧
問題 31 / 65
Which design modification is effective if you want a model that uses positional encodings, like a Vision Transformer, to operate without further training when the input size changes?
①Keep using fixed absolute positional embeddings.
②Always resize the input image to a fixed resolution before tokenizing.
③Increase the number of Transformer layers to enhance expressive power.
④Replace with a convolutional stem that takes optical properties into account.
⑤Introduce relative positional encoding.