search

facebookresearch / ToMe

A method to increase the speed and lower the memory footprint of existing vision transformers.
Other
931 stars 67 forks source link

The difference between 85.96 and 85.66 original MAE accuracy #17

Closed LiuDongyang6 closed 1 year ago

LiuDongyang6 commented 1 year ago

Thank you for this work!

I observe that two different numbers (85.96 and 85.66) are reported for MAE ViT-L/16 original accuracy, and I wonder what the underlying difference is.

Thanks!

dbolya commented 1 year ago

Thanks for catching that. The 85.66 is a model we trained ourselves, while the 85.96 is the model available from the official pytorch MAE repo. Since our training setup couldn't reproduce the 85.96 number, we had to train our own model for fair comparison.

LiuDongyang6 commented 1 year ago

Thanks