Self-attn in decoder layers.

[TsingWei]

I noticed there is a section about DETA does not need self-attention in the decoder. in the paper. The results show that when the self-attn is replaced by ffn in decoder, the performance is better. I wonder whether the final version in the table of compared-with-other-SOTAs using this setting? Because I found in the code that the self-attn is hard-coded in the decoder layer:

DETA/models/deformable_transformer.py

Line 328 in dade176

self.self_attn = nn.MultiheadAttention(d_model, n_heads, dropout=dropout)

[jozhang97]

Thank you for your interest!

I wonder whether the final version in the table of compared-with-other-SOTAs using this setting?
No, our default model still contains self-attention.