Net Surgeries

[shelhamer]

The editing model parameters example only hints at the possibilities of net surgery. To make it more useful and better explain pycaffe along the way it should have:

• transplanting custom filters (Gaussian, Sobel, bilinear, ...) and classifier weights (like a separately learned SVM)
• combining layers from different nets / making an ensemble in a single model
• grafting new channels to a model to augment the number of classes
• switch the channel order of conv1, or convert between grayscale and color

These are all possible as-is, but it takes know-how.

[shelhamer]

Partly done in #2058.

[shelhamer]

This is as done as it will be in the foreseeable future.