A question about the the equation 4

[goldentimecoolk]

Hi, in the equation 4 of the paper, there is a multiplication of MLP(C_q) and PE(x_q, y_q). But in the specific implementation, the query_pos (PE(x_q, y_q)) is directly input to the cross-attention and not multiplied by MLP(C_q). Instead, the pos_transformation is multiplied by MLP(C_q) two times, which seems to be inconsistent to the paper. Is there something wrong in my understanding?