Use Alexnet for two-class classification scenrio?

[Saeed Izadi]

As you know, the Alexnet has been trained on the 1000 category scenario, containing many classes. Now I want to use the architecture  for a two class problem, say cars vs. boats, What should I do? I should fine-tune the network again on all the images from cars and boats or re-training the last classifier layer suffices? 

Bests,

[Cyndi]

Depending on how much data you have, I would first try to fine-tune fc8.  If this is doing okay, and not over-fitting, then you can try fine-tuning the fully connected layers and conv layers.  Additionally, since you only have two classes, you might consider using SigmoidCrossEntropyLoss as opposed to SoftmaxWithLoss.