Great work @yucornetto! One question I couldn't find the answer to in the paper: how is linear probing done? I get that the details for the training loop and such are from the MAE paper. However, I'm not sure about where you extract the representations for the linear probe. Is it right before the quantizer (16-dimensional embedding), or 1024-dimensional representations of the latent tokens?
Great work @yucornetto! One question I couldn't find the answer to in the paper: how is linear probing done? I get that the details for the training loop and such are from the MAE paper. However, I'm not sure about where you extract the representations for the linear probe. Is it right before the quantizer (16-dimensional embedding), or 1024-dimensional representations of the latent tokens?