Details on training hyperparameters

#5
by egesko19 - opened

Is there any way I can find details on how such high accuracies in Imagenet-1k is reached? I am trying to replicate it by training it from scratch to reach similar metrics, but I usually end up around 60% accuracy before validation loss starts to increase, which is a massive difference compared to the pretrained model that reaches around 90% accuracy on validation dataset.

Sign up or log in to comment