Label Reordering

[mschrimpf]

When we test how well Alexnet pre-trained on ILSVRC performs on PASCAL-VOC, the label ordering is different, e.g. neuron 5 does no longer represent a dog but a tree.
I can think of two ways to adjust the models to this labeling difference:

1. throw out the last layer and train an SVM on the second-to-last layer
2. freeze all layers but the last and re-train on the specific dataset

I guess it doesn't matter which approach we choose, right? @ncheney

[ncheney]

I'm indifferent to the two approaches. I'm also not sure whether it's best to retrain just the last layer or all 3 of the fully connected layers (thought I'd be interested in seeing how much better of result we get with more retraining -- or if the combination of convolutional features is robust to the different class sets).

Thought I think that this issue will come up only after we try the simpler case with training on only half of the imagenet classes and then adding in the other half.

[ncheney]

p.s. that's for the tag, as it send a notification to my email for me to respond to it