Skip to content

Conversation

chrisway613
Copy link

In the decoder position embedding matrix, the size of first dim is the number of patches + 1, as the 1 for ViT's cls_token. But when embedding the position for masked tokens, their indices have not shifted 1, it may confuse with the position of the ViT's cls_token(although MAE do not use cls_token, but this will lead to weak extensibility if we wanna use the cls_token later)

@lucidrains
Copy link
Owner

@chrisway613 Hi Chris! while this is true, i think leaving untrained parameters in the wrapper class isn't elegant. you can always just concat the CLS tokens onto the decoder_pos_emb after you finished training, something like

decoder_cls_token = nn.Parameter(torch.randn(1, decoder_dim))
pos_embs_with_cls_token = torch.cat((decoder_cls_token, self.decoder_pos_emb), dim = 0)

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

Successfully merging this pull request may close these issues.

2 participants