Non-homogenous Vectorization

[pgkirsch]

Related to #1465:

Is there an elegant way to enable the following concept of vectorization with GPkit:

with Vectorize(4): 
   self.HLM = HighLevelModels()
    with Vectorize([2, 3, 1, 1]):
        self.MLM = MidLevelModels()
        with Vectorize(2):
            self.LLM = LowLevelModels()

where the length of the middle level of vectorization must necessarily be equal to the degree of the top level of vectorization.

I'm trying to figure out an elegant way of allowing the 7 parallel mid-level models to share common parents. So in this example, the first two models share the same high-level model, models 3-5 share the same high-level model, and so forth.

Ultimately the lowest level variables in this example should have shape (4,7,2).

Any suggestions?

[1ozturkbe]

Alright, I toyed with some ideas just now, and none of them were elegant.

[1ozturkbe]

I think this is most elegant when there is a mapping... will post in a little bit, need breakfast.

[1ozturkbe]

I don't have time to do this, but I think you could do a sequential mapping, where you generate LLMs, MLMs and HLMs with indices [x,y,z], [x,y] and [x], and recursively nest them as arguments to each other. This way you can generate arbitrary shapes of vectored models.