Hi
I found the LeNet example very helpful.
However I've had problems adapting to GoogLeNet.
I am training a face recogniser using GoogLeNet adapted for the last layer. I am trying to make the last layer work with triplet size 128 or 256.
However when I train either using Naive or standard triplet training, I find that my training and test loss, and test accuracy, jump all over the place. This happens even when I reduce the LR to 0. Does this fork not use the LR?
If you have had any success applying this fork to GoogLeNet, please could you share the train_test.prototxt file and solver file? The LeNet example was helpful but GoogLeNet is closer to a real life use case and would be extremely helpful!
Thank you very much!
Hi I found the LeNet example very helpful. However I've had problems adapting to GoogLeNet. I am training a face recogniser using GoogLeNet adapted for the last layer. I am trying to make the last layer work with triplet size 128 or 256. However when I train either using Naive or standard triplet training, I find that my training and test loss, and test accuracy, jump all over the place. This happens even when I reduce the LR to 0. Does this fork not use the LR? If you have had any success applying this fork to GoogLeNet, please could you share the train_test.prototxt file and solver file? The LeNet example was helpful but GoogLeNet is closer to a real life use case and would be extremely helpful! Thank you very much!