You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
I wanted to share my observations regarding the correct usage of moco v3 as a backbone, as reported by the moco official repository, as well as a potential inconsistency in the repository.
In the file main_lincls.py of the moco v3 repository, the module self.base_encoder.head is removed assuming that ViT is being used as the backbone. However, in the file get_features_mocov3.py of the same repository, the fc layer self.base_encoder is still present in the backbone, which should be nn.Identity() in this case, as did to Ctranspath.
I just wanted to share my observations and this does not indicate any doubts towards the experiment results. In fact, I really enjoyed this work a lot.
I am open to have a discussion with followers.
The text was updated successfully, but these errors were encountered:
Heyy,
I wanted to share my observations regarding the correct usage of moco v3 as a backbone, as reported by the moco official repository, as well as a potential inconsistency in the repository.
In the file main_lincls.py of the moco v3 repository, the module self.base_encoder.head is removed assuming that ViT is being used as the backbone. However, in the file get_features_mocov3.py of the same repository, the fc layer self.base_encoder is still present in the backbone, which should be nn.Identity() in this case, as did to Ctranspath.
I just wanted to share my observations and this does not indicate any doubts towards the experiment results. In fact, I really enjoyed this work a lot.
I am open to have a discussion with followers.
The text was updated successfully, but these errors were encountered: