Hi.
I have some problems with retrianing an Alexnet trained on ImageNet.
I hope to retraining an Alexnet on some natural image datasets and use it as a feature extractor for image retrieving. This work is originally written on caffe and I hope to reproduce the same results using Pytorch.
But I notice that in the original caffe code, Alexnet has LRN (local response normalization) in Layer 1 and Layer2. But in torchvision model, we don’t have such normalization and even without BN as well. I am rather concerned if it will influence the results? And is it feasible that I add BN after the first two layers while loading the weights trained without BN for retraining?
Thanks