Is the pretrained mae encoder weights available ?

[CheungZeeCn]

in config:

"mae_checkpoint": "mae_models/mae_pretrain_vit_large_full.pth"

in udop_dual:

self.vision_encoder = mae_model(config.mae_version, config.mae_checkpoint, config.image_size, config.vocab_size, 
config.max_2d_position_embeddings)

But I found no pretiraned weights for mae encoder. Is the pretrained mae encoder weights available now?

Thank you!

[zinengtang]

The MAE checkpoint is together with the transformer weights included in the checkpoint.
if you want the original MAE weights you can download it from the original MAE codebase.

[znb899]

In the transformer weights, for mae, there are only weights for patch_embed and special_vis_token (and pos_embed), but not the blocks. And in the forward method, you indeed only use patch_embed to encode the patches.

Do you not use the full mae like in udop_dual? This simple projection carries all the information for font, line spacing, color etc etc?