search

LeapLabTHU / Agent-Attention

Official repository of Agent Attention (ECCV2024)
473 stars 35 forks source link

Agent_num #20

Closed Dasha484 closed 7 months ago

Dasha484 commented 7 months ago

Hello, I encountered several questions while reading your code.

1.What does 'agent_num mean'? I did not find a clear definition in the paper.

2.When defining bias, why is the dimension of bias related to agent_num? If I modify agent_num, it will result in dimension mismatch. However, I noticed that your paper includes comparisons with different agent_num.

I hope to receive your response.

tian-qing001 commented 7 months ago

Hi @Dasha484.

  1. agent_num represents the number of agent tokens, which has been detailed in the paper. We also provide ablation experiments on it in Table 8.
  2. Various agent tokens exhibit distinct biases across spatial positions, so the dimension of agent bias is related to agent_num. When modifying agent_num, consider using interpolation to adjust the size of agent bias. Refer to section C.2.