Question about MDA

[bilipa]

In the 2nd training stage, we use conv_1x1 generate alpha_i. Then we use mulitiplication for alpha_i and inception, and get the output.(The output shape: batch x L*C x H x W). In your paper, you said: Each attentive feature map is then passed through the following blocks thereafter..... Does it mean use the output feed into the inception module ? for example, in F(alpha_2), the output incept_1_alpha_2_output feed into [incept_1, incept_2, incept_3], incept_2_alpha_2_output feed into [incept_2, incept_3] and incept_3_alpha_2_output feed into [incept_3] or all of output feed into [incept_1, incept_2, incept_3]?

[bilipa]

I sent the email to author. Here is the answer:

The output attentive feature maps are fed into the following inception modules in this way:

incept_1_alpha_2_output are fed into [incept_2, incept_3], incept_2_alpha_2_output are fed into [incept_3] and incept_3_alpha_2_output are not fed into inception blocks, and directly fed into global average pooling layer.