DELG implementation does not have the auto-encoder as cited in paper.

The DELG paper

According to the latest version DELG paper, the local feature part of DELG is slightly different from DELF. It apply an auto-encoder on features and then use the attention weights on reconstructed features. losses of auto-encoder reconstruction and attention classifier are used together to train the local features. Just like the two screenshot below:

The implementation of DELG in DELF repo

However, the code implementation of DELG here does not include the auto-encoder.

1. In delg_model.py, The class DELG is a subclass of DELF and use the exact local feature part as DELF.

2. in train.py, we can look at the train_step function (line 242 -322). Only SparseCategoricalCrossentropy losses of two classifier (descriptor and attention) are taken into account.

Question about the auto-encoder feature

Does the implementation of DELG here miss the auto-encoder? or the paper has a wrong description about its model structure?

Is there any plan to add the auto-encoder inside?