You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
** I have some questions about 'Criss-Cross Attention'. Here should be C×W×H or C×H×W in the place marked by the red circle?
At the same time,I have some questions about Affinity.Which variables corresponds to the codes in the place marked by the blue line,proj_query_H and proj_key_H , or proj_query_W and proj_key_W? I don’t understand the definition of Qu and Ωu very well. Is it convenient for you to answer? Looking forward to your answer, thank you very much!
The text was updated successfully, but these errors were encountered:
Q_u is a pixel vector in spatial position u with shape [b, c', 1, 1], so the x_q contains h*w Q_u
x_q=self.q(x) #b, c', h, w
3.\gamma_u is a set which contains the key pixel vectors in the same row and column with position u, so its shape is [b, c', h+w-1].Actually, in code they split the key set to H and W,and calculate the attention scores for h and w respectively.
Q_u is a pixel vector in spatial position u with shape [b, c', 1, 1], so the x_q contains h*w Q_u
x_q=self.q(x) #b, c', h, w
3.\gamma_u is a set which contains the key pixel vectors in the same row and column with position u, so its shape is [b, c', h+w-1].Actually, in code they split the key set to H and W,and calculate the attention scores for h and w respectively.
**
I have some questions about 'Criss-Cross Attention'. Here should be C×W×H or C×H×W in the place marked by the red circle?
At the same time, I have some questions about Affinity. Which variables corresponds to the codes in the place marked by the blue line,proj_query_H and proj_key_H , or proj_query_W and proj_key_W? I don’t understand the definition of Qu and Ωu very well. Is it convenient for you to answer? Looking forward to your answer, thank you very much!
The text was updated successfully, but these errors were encountered: