I know you have abandoned that repo - I need a running version of AlexNet for baseline issues for my work and I have an implementation for keras2/tf available. However, when I train my models I achieve top-5-error rate 30.3% on ILSVRC2012 validation data, which is far off from the 20% you have published. Since I basically use your architecture with some minor modifications for keras 2 support I am wondering why that is. Could you share some information how you trained your models (optimizer, data augmentation etc.) so that I could train AlexNet by myself?
Thanks a lot!
Hi,
I know you have abandoned that repo - I need a running version of AlexNet for baseline issues for my work and I have an implementation for keras2/tf available. However, when I train my models I achieve top-5-error rate 30.3% on ILSVRC2012 validation data, which is far off from the 20% you have published. Since I basically use your architecture with some minor modifications for keras 2 support I am wondering why that is. Could you share some information how you trained your models (optimizer, data augmentation etc.) so that I could train AlexNet by myself? Thanks a lot!