Thx for the fancy work.
I wonder is there any plan to release the code of masked patch prediction.
I am very interested in the detail of the training. For example, is there any data augmentation for the training?
we don't plan to release that code. As for your first question, we trained on JFT and used the same data augmentation as for classification (random flip & inception_crop to 224 pixels).
Thx for the fancy work. I wonder is there any plan to release the code of masked patch prediction. I am very interested in the detail of the training. For example, is there any data augmentation for the training?