Simple implementation of MAE, mask out some patches in input image, send patches not masked into a transformer encoder, cat encoder output with mask tokens into transformer decoder to reconstruct original image.
model trained with mask_ratio=0.75, infer in different mask ratio.
-
mask_ratio = 0.5
-
mask_ratio = 0.75
-
mask_ratio = 0.90
-
mask_ratio = 0.95
- train a vit on this dataset, and pretrain MAE on it, varify whether MAE pretrained can improve acc