about attribute and content net ?

[ranjiewwen]

• in this picture , the (b)and(c)net are one of branch for(a); it is (b)and(c) net also input ranks pair of image?
• another question is how the feature fusion ? and only concated?
• and the attribute net how we get the attribute score? from the picture only get attr_feat ,and last get total scores , calcuate the total scores loss?

@aimerykong thanks!

[aimerykong]

hi, you can think of (b) ansd (c) as a branch of (a)

…

On Mon, Jun 11, 2018, 01:09 ranjiewen ***@***.***> wrote: [image: 20180611160545] <https://user-images.githubusercontent.com/9885492/41219610-66518138-6d91-11e8-9910-480840877524.png> - in this picture , the (b)and(c)net are one of branch for(a); it is (b)and(c) net also input ranks pair of image? — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#19>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AGKZJJ4JIiRuJ5na4wilAV-Al4r8j7lXks5t7iW9gaJpZM4UiPUi> .

[ranjiewwen]

thanks, can you explain the rest question ?

[aimerykong]

I see...your questions do not show up completely in email but in github. attribute features are concatenated with aesthetic features; but content features are multiplicatively gated with the content classification scores. We train attribute classification layer so that we can get attribute scores.

…

On Mon, Jun 11, 2018, 08:17 ranjiewen ***@***.***> wrote: thanks, can you explain the rest question ? — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#19 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AGKZJA2tZoQ4rQcC9lpcz6JMwtcUBKdCks5t7oojgaJpZM4UiPUi> .

[ranjiewwen]

thanks , that is fisrt train attribute classification layer ,get Att_fea, optimize the loss_att ; and then use the (b)net optimize the loss_reg and loss_rank?