Contrastive loss layer bug

[happynear]

Contrastive loss layer bug

[happynear]

Because I have came across difficulty when reproducing DeepID2. It optimize the softmax loss and contrastive loss simultaneously. I must add an additional label layer to tell contrastive loss whether the two samples are the same/different class. It is stupid.

Then I modified the contrastive layer to determine the same/different class by itself. With my implementation, I just need two usual data layers created by convert_imageset.exe, where 'usual' means a data blob and its corresponding label blob. No need to write program to create the weird data db as in siamese network example.

Here (https://github.com/happynear/FaceVerification/blob/master/caffe_proto/mnist_siamese_train_test.prototxt) is an example of how to use my contrastive layer. I think it is the best way for using the contrastive layer. All I need is to ensure the two image list files are mutually correspond to each other.

[d4nst]

Thanks for the clarification. I am about to start using Caffe for a face verification project, so I might be using it in the same way as you do. I will keep an eye to your other repo as well since I am very interested on reproducing DeepID/CASIA/DeepFace results.

In any case, don't you think that it would be better to have two different contrastive loss layers, one with the original implementation and another one with your custom implementation? In this way, the original siamese network example would still work.

[happynear]

I would rather to write a new example to introduce my contrastive layer. The usage of the original siamese network is really weird. Not only for the contrastive loss layer, but also for its two-way network definition. You can find that I used the concat_layer to combine the two-way network together. I think this is the right way to realize siamese network, too.