Yes, for one the weights are totally new, second I dropped a bunch of layers because vgg16 is way too powerful for this task. As I get into more kinds of parts I may have to put those back in though. Finally, I added dropout after the convolutional layers as well as after the fully connected layers, that seems to help a lot with the accuracy. All this was arrived at empirically, I don't know enough about neural net topology to go about that in a more analytical way but it does the job.

As for opensourcing it, yes, but then likely lego would have a trademark case so I would probably use another name. The whole thing would make for a pretty neat Kaggle competition.