Softmax Loss
dgomezurita opened this issue · comments
Congratulations for your work.
I would like to test your network using the softmax loss function but I still do not understand how to change your train.prototxt. to train correctly. (I have tried several things), I have relied on the branch v_1 however I do not get the results published in your work.
I would like to ask your help to correctly define the train using softmax.
Thank you very much.
@dgomezurita Not sure what your referred to. Is this layer you are looking for?
Yes, I have seen that layer of your branch v1, I would like to change the lost OIM function in your master branch by the RandomSamplingSoftmaxLoss layer of your v1 branch. Will it be possible to do that?
Should I remove this layer of OIM and this and I replace them with RandomSamplingSoftmaxLoss?
If you want to use RandomSamplingSoftmax rather than OIM loss, you can directly use branch v1.
Yes, that is correct, but I expressed myself badly, I would like to use the convolutional network resnet-50 of its branch master instead of VGG of branch v1, what is your recommendation to try SoftmaxRS or conventional Softmax?
thank you very much.
We have a baseline of using conventional Softmax loss in the paper. I don't think it can work in this person search framework.
We also have an experiment of subsampling unlabeled identities in the paper. I think SoftmaxRS might work but OIM is a better choice due to its non-parametric nature.
thank you!