search

dvlab-research / Mask-Attention-Free-Transformer

Official Implementation for "Mask-Attention-Free Transformer for 3D Instance Segmentation"
59 stars 6 forks source link

Question about mask in cross attention component #5

Closed triton99 closed 3 months ago

triton99 commented 1 year ago

Hi @X-Lai , Thanks for sharing this great work!

What is the purpose of attn_masks in your transformer decoder? In your paper, you mentioned that mask-attention-free-transformer.

https://github.com/dvlab-research/Mask-Attention-Free-Transformer/blob/4b5048c0e08c2fc42f660dfea3209043179ace1b/maft/model/transformer.py#L108-L115

Thank you.

NiccoloCavagnero commented 3 months ago

If I understand correctly, only the first cross-attention layer is not masked, all subsequent layers are.