#positional-embedding

Question 1: Single choice

A ViT was pretrained on $224 \times 224$ images with $16 \times 16$ patches and learned position embeddings for its $14 \times 14$ patch grid. You fine-tune it on $384 \times 384$ images with the same patch size, which gives a $24 \times 24$ grid. What do you do with the position embeddings?

Interpolate them in 2D to the new $24 \times 24$ grid1
Keep them unchanged, since attention works for any number of tokens2
Use $27 \times 27$ patches so that the grid stays $14 \times 14$ 3
Center-crop the images to $224 \times 224$ 4

#positional-embedding

Timed test