Steer pyr demo

[ThomasYerxa]

Demonstrates the use of steer pyr by (1) using it as a front end for a conv net classifier on FashionMNIST and (2) comparing the eigendistortions w/ a baseline classifier that doesn't use a steer pyr frontend.

[review-notebook-app]

Check out this pull request on 

See visual diffs & provide feedback on Jupyter Notebooks.

---

Powered by ReviewNB

[codecov-io]

Codecov Report

Merging #119 (23ae52d) into master (37758e1) will decrease coverage by 47.42%.
The diff coverage is 12.50%.

@@             Coverage Diff             @@
##           master     #119       +/-   ##
===========================================
- Coverage   75.42%   27.99%   -47.43%     
===========================================
  Files          33       33               
  Lines        2629     2568       -61     
  Branches      526      516       -10     
===========================================
- Hits         1983      719     -1264     
- Misses        505     1777     +1272     
+ Partials      141       72       -69     

Impacted Files	Coverage Δ
plenoptic/tools/signal.py	16.66% <ø> (-34.07%)	⬇️
...e/canonical_computations/steerable_pyramid_freq.py	6.64% <6.77%> (-82.96%)	⬇️
plenoptic/tools/display.py	10.23% <50.00%> (-56.31%)	⬇️
plenoptic/tools/data.py	51.58% <100.00%> (-21.49%)	⬇️
plenoptic/synthesize/mad_competition.py	9.97% <0.00%> (-75.88%)	⬇️
plenoptic/metric/model_metric.py	28.57% <0.00%> (-71.43%)	⬇️
plenoptic/metric/perceptual_distance.py	15.09% <0.00%> (-70.76%)	⬇️
...simulate/canonical_computations/non_linearities.py	20.51% <0.00%> (-64.11%)	⬇️
plenoptic/synthesize/synthesis.py	22.10% <0.00%> (-51.88%)	⬇️
... and 8 more

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 37758e1...23ae52d. Read the comment docs.

[ThomasYerxa]

converted to PR prematurely. closing for now

[ThomasYerxa]

When the dust settles, the only significant change coming from this branch should be the Demo_Steerable_Pyramid.ipynb notebook. That should be relatively stable to the one appearing in this PR, so feel free to glance at the notebook if curious but ignore everything else.

[ThomasYerxa]

Questions:

1. Was there a particular reason we opted for a frequency space implementation (boundary conditions are nicer etc.)? In "1.1: Visualizing wavelets" it would be nice to mention if so.
2. For the steerability demo: is this sufficient/interesting? would it be more compelling i.e. to steer to an intermediate orientation and then show that if you initialize a pyramid that includes the steered-to-orientation you can recover the same coefficients.

[billbrod]

I've gone through and edited the text here, take another look at it and tell me what you think. @nikparth can you also take a look at this?

• I think the final section, on probing the effect of using the steerable pyramid as a frontend, is a great idea, but I think we need some more text at the end to explain. I tried adding some, but it still feels like we ran out of steam while writing it and are just saying "whelp, you make sense of it" . I also think need a bit more detail about what the networks are being trained to do, some framing, etc.
• I'm tempted to actually split that whole section off into a separate notebook, so we have two notebooks: steerable pyramid basics and steerable pyramid-as-model component. I could add something about classic visual models there too. Thoughts?
• Re: your question 1. I think it's because that was the only one we really want people to use: it handles any order, reconstruction is exact, and allows for the complex-valued output (which is really useful). @nikparth do you remember if there was another reason?
• Re: question 2. I think this is fine for now, but agree that steering an intermediate representation would be more compelling. Can you add info to Add demo showing steering of intermediate representation #204 ? Add all the notes you have, any code, etc, so that you or someone else can pick it up later.

I also pushed some small updates to (unrelated) docstrings, because I noticed them and they were annoying me.

And here it is on the readthedocs

[nikparth]

I think overall this seems fine as a first release. I added some comments on wording in the notebook, but my main comments would be to remove the standard CNN demonstration and comparison to the pyramid. I think it's fine to just show that the pyramid can be added as a module to build on with standard pytorch modules. And then I'd say remove the eigendistortion section as well and we could have another notebook that is about using plenoptic methods with canonical visual models.