Remove self.representation (and other memory intensive attributes) from portilla-simoncelli

[kbonnen]

Oh the other thing I wanted to bring up -- I thought we had decided that models shouldn't keep a self.representation lying around, so that they're more lightweight. Did I remember that correctly? If so, that will require a bit more work here, but it looks possible to me.

Originally posted by @billbrod in #96 (comment)

[kbonnen]

We should do this. It means removing self.representation and the following attributes:

pyr: Steerable_Pyramid_Freq
The complex steerable pyramid object used to calculate the portilla-simoncelli representation
pyr_coeffs: OrderedDict
The coefficients of the complex steerable pyramid.
mag_pyr_coeffs: OrderedDict
The magnitude of the pyramid coefficients.
real_pyr_coeffs: OrderedDict
The real parts of the pyramid coefficients.
representation_scales: list
The scale for each coefficient in its vector form
representation: dictionary
A dictionary containing the Portilla-Simoncelli statistics

[billbrod]

I think we can keep around the pyramid (and I actually added several other pyramids), because not having to re-create them each forward pass will speed up performance (especially on GPU, otherwise we have to create pyramid and then pass to GPU, and copying to GPU takes a while). But just not their coefficients.

[billbrod]

Closed by #225