Hello,
I am testing Transfernet to modify it to my data. I still face troubles to understand attention module as it is described differently in your equations in paper, your diagram (same paper) and in deploy.prototxt . I tried to visualize the network deploy.prototxt in Netscope - http://ethereon.github.io/netscope/#/editor but it gives me errors. Can you describe me where is the soft_att blob located as there are two contradictory definitions in prototxt, namely:
layer {
name: "soft-att"
type: "Scale"
bottom: "conv5_3"
scale_param { scale_factor: 0.004 } # no scaling
top: "conv5_3_scaled"
}
and
layer {
name: "soft-att"
type: "LinearSum"
bottom: "conv5_3_scaled"
bottom: "att"
top: "soft-att"
}
The Netscope refuse to visualize it (also attention.prototxt) and I am not able to figure out, what is the exact structure of attention module.
Thank you for clarification
Michal
Hello,
I am testing Transfernet to modify it to my data. I still face troubles to understand attention module as it is described differently in your equations in paper, your diagram (same paper) and in deploy.prototxt . I tried to visualize the network deploy.prototxt in Netscope - http://ethereon.github.io/netscope/#/editor but it gives me errors. Can you describe me where is the soft_att blob located as there are two contradictory definitions in prototxt, namely:
layer {
name: "soft-att"
type: "Scale"
bottom: "conv5_3"
scale_param { scale_factor: 0.004 } # no scaling
top: "conv5_3_scaled"
}
and
layer {
name: "soft-att"
type: "LinearSum"
bottom: "conv5_3_scaled"
bottom: "att"
top: "soft-att"
}
The Netscope refuse to visualize it (also attention.prototxt) and I am not able to figure out, what is the exact structure of attention module.
Thank you for clarification
Michal