Add PortillaSimoncelliMinimal to tutorial get original paper's statistics

[dherrera1911]

Add a new section to Portilla Simoncelli metamers tutorial, with a demonstration of how to recover the original set of statistics reported in the paper from the output of the model. closes #151

[review-notebook-app]

Check out this pull request on 

See visual diffs & provide feedback on Jupyter Notebooks.

---

Powered by ReviewNB

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Comparison is base (8573cb2) 86.82% compared to head (664da25) 86.79%.

❗ Current head 664da25 differs from pull request most recent head 0f8f094. Consider uploading reports for the commit 0f8f094 to get more accurate results

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #216      +/-   ##
==========================================
- Coverage   86.82%   86.79%   -0.03%     
==========================================
  Files          35       35              
  Lines        3346     3347       +1     
==========================================
  Hits         2905     2905              
- Misses        441      442       +1     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[review-notebook-app]

View / edit / reply to this conversation on ReviewNB

dherrera1911 commented on 2023-08-03T21:28:40Z
----------------------------------------------------------------

Cut line

[review-notebook-app]

View / edit / reply to this conversation on ReviewNB

dherrera1911 commented on 2023-08-03T21:28:41Z
----------------------------------------------------------------

Maybe should be 'dropped from full set'

[review-notebook-app]

View / edit / reply to this conversation on ReviewNB

dherrera1911 commented on 2023-08-03T21:28:42Z
----------------------------------------------------------------

Typo, text -> test

[review-notebook-app]

View / edit / reply to this conversation on ReviewNB

dherrera1911 commented on 2023-08-03T21:28:43Z
----------------------------------------------------------------

This 2-level list rendered better in jupyter lab.

[billbrod]

This looks great to me, thanks for putting this together! I've made some small changes to the following:

• change to index tensor type to make it run on python 3.7
• use coarse-to-fine in final synthesis example, update model_min's forward pass to support that
• small wording changes, line wraps, spacing, capitalization convention (camel case -> snake case)

Outside of the changes you made, I also made some we had discussed earlier:

• in PS docstring, clarifies that, when use_true_correlations=True, the diagonals are still un rescaled, so that they're covariances)
• Adds a brief discussion of how to get proper coarse-to-fine integration when extending the PS model (to the notebook)
• makes the steerable pyramid's default / "all" argument for scales match that of the PS model (None, rather than [])

There's only one thing:

• Can we find something to link to about the auto-covariance matrix symmetry? And maybe why it's difference from the regularly symmetry of covariance matrices.

[eerosim]

Daniel,
Thanks so much for putting the time into tracking down the redundancies in the parameter file! A few notes:

• In the list at the top, I think it's worth providing the math expressions for the number of non-redundant correlational parameters (auto-correlation, cross-correlation, etc). You currently provide those later, when tallying up the numbers, but I think it's preferable to state them up front when you describe the issue.
• The magMeans are different (for reasons I cannot remember, they are in the code but do not seem to do much). Unlike the correlations, they they are only approximately redundant, and I'm wondering if we can make the case for their elimination more precise. Specifically, in your comparison of textures generated by the full and minimal model:
  1. I believe you're running your examples from the same starting image. Can you also use identical seeds for the random number generator (to guarantee that the stochastic Adam optimizer draws the same random values)? I think that should yield images that are much more similar, perhaps even visually indistinguishable.
  2. Report the RMSE between the two synthesized images.
  3. Compare the values of the magMeans in the two results. I think they should be nearly the same, even though they are not explicitly "enforced" in the minimal model (and if so, this demonstrates that they are statistically redundant).

Thanks much,
Eero

[billbrod]

There are some more changes we'd like to see to the text at the bottom of the notebook, but given the significant changes to the PS code in #225 , I'm going to merge this PR and we'll finish the changes over in the new PR. I'll copy over Eero's comments.