search

camel007 / caffe-moon

This repo is re-implement "MOON: A Mixed Objective Optimization Network for the Recognition of Facial Attributes"
39 stars 16 forks source link

It does not seem to use the mixed objective optimization？ #9

Open zhengge opened 6 years ago

zhengge commented 6 years ago

In train_val.prototxt， I do not find the mixed objective optimization.

layer {
 name: "euclidean-loss"
  type: "EuclideanLoss"
  bottom: "moon-fc"
  bottom: "labels"
  top: "loss"
}

Could you explain it? Thanks.

bikong2 commented 6 years ago

i did not find the moon part, too.

chaipangpang commented 6 years ago

1.this project no relationship with the paper—“MOON”. 2.Why use the loss function —“EuclideanLoss”？It is better use “Sigmoid Cross-Entropy” or “HingeLoss”？

BackT0TheFuture commented 6 years ago

this repo does have noting to do with MOON loss.

@zhengge @bikong2 @chaipangpang

moon

did you guys make MOON loss clear? how to calculate src_dist ？

klintqinami commented 5 years ago

"Since CelebA has identical source and target distributions, we define the loss layer in (6) to weight all elements equally during backpropagation – which is equivalent to Euclidean loss between the network output and the 40 binary attribute values."

[Page 9, MOON : A Mixed Objective Optimization Network for the Recognition of Facial Attributes]