Question about the temporal span I1 and I2

[chantsss]

Thanks for sharing your excellent work as open source. I have a question about your code. I couldn't find the specific parameters for the temporal span I1 and I2 that you mentioned in your paper. Could you please explain how to pass them and how they work in the attention mechanism?

[XiaolongTang23]

I apologize for any confusion caused by the terminology used in our paper. The parameters referred to as "temporal span I1 and I2" in the paper are named pos_duration and pred_duration in the code. These parameters control the window size for past positions and past predictions, respectively.

[chantsss]

Thank you for your prompt response. Just one more question, I see the pos_duration I1 defaults 20, which is already the maximum value for history length. In other words, the window size for past positions is [t − I1, t]==>[t − 20, t]. In such case, t − I1− I2<t − 20, and it is out of the interval [t − 20, t]. How do you handle such case?

[XiaolongTang23]

You are right. During normal training, although the visible window is larger, we are limited to only the 20 frames of data. However, in the experimental section of our work, I set a longer history length to test the advantages of a larger visible window. For more details, you can refer to Figure 3 in the paper.