Triplet loss test

[happynear]

Some users reported that the program would crash when using triplet loss. Similar issue https://github.com/BVLC/caffe/issues/2995 . It seems that the caffe_gpu_dot() is unproperly used.

I have merged another implementation of triplet loss. It uses gemv instead. I am out for business now. Can someone have a try and report the result to me?

[cheer37]

Did you test your triplet implementation? Does it work correctly? I am using triplet instead of siamese for face verification. I think you have superb experience about it. Please guide me, if you post your triplet train_val.prototxt, it will be so helpful. Thanks.

[cheer37]

Happy, i will try to test it. but i dont figure out how to write prototxt for train_val about data layer and loss layer. It seems there is not need label information in data layer because triplet loss layer does not have label information as bottom. If you share its sample, i will test this triplet loss module soon. Thanks.

[yzk0281]

@cheer37 I am training models for face verification too， maybe we can help each other， can you add my qq 594617334？

[cheer37]

I dont have qq. let me know your gmail.

[cheer37]

@happynear. Your triplet seems work correctly.

[yzk0281]

@cheer37 how to use tripletloss layer?

[yzk0281]

@cheer37 how is your retraining for Wu's model? Did you test tripletloss on Wu's model?

[cheer37]

Yes, but not gaining performance boost yet.

[yzk0281]

How did you pre-processing you data? Like Wu says 5 feature points to align?

[cheer37]

Yes

On Wed, Apr 27, 2016 at 5:08 PM, yzk0281 notifications@github.com wrote:

How did you pre-processing you data? Like Wu says 5 feature points to align?

— You are receiving this because you were mentioned. Reply to this email directly or view it on GitHub https://github.com/happynear/caffe-windows/issues/42#issuecomment-215020023

[yzk0281]

Could you share your tripletloss train_val.prototxt please？