[MRG] Gaussian Mixture OT #649
Conversation
Codecov ReportAll modified and coverable lines are covered by tests ✅
Additional details and impacted files@@ Coverage Diff @@
## master #649 +/- ##
==========================================
+ Coverage 96.71% 96.78% +0.07%
==========================================
Files 86 88 +2
Lines 17148 17502 +354
==========================================
+ Hits 16584 16940 +356
+ Misses 564 562 -2 |
… to reflect PR changes
ot/gmm.py
Outdated
| Cs12 = nx.sqrtm(C_s[i]) | ||
| Cs12inv = nx.inv(Cs12) | ||
|
|
||
| for j in range(k_t): |
There was a problem hiding this comment.
challenge 2 (optional) :here the second loop should be doable with einsum and batch sqrtm too
There was a problem hiding this comment.
for both maps, I ended up doing a loop over all nonzero indices in the plan, I don't know if this is better than a smart einsum.
| # i and j, b[i, j] is the translation part | ||
| rng = np.random.RandomState(seed) | ||
|
|
||
| A = nx.zeros((k_s, k_t, d, d)) |
There was a problem hiding this comment.
this is a BIIG tensor!, again mayb can be done with einsum
There was a problem hiding this comment.
looping over nonzero entries in the plan is probably faster for this function, since there are at most k_s + k_t - 1 nonzero entries. I'll look into that
ot/plot.py
Outdated
|
|
||
|
|
||
| def plot1D_mat(a, b, M, title=''): | ||
| r""" Plot matrix :math:`\mathbf{M}` with the source and target 1D distribution | ||
| def plot1D_mat(a, b, M, title='', a_label='Source distribution', |
There was a problem hiding this comment.
thanks for the big and necessary update but keep in mind that yteh function must keep a very similar behavior by default than teh previous version.
That is the color and source/target should be the same as previous version . It is OK to have parameter that change this default behavior and you can use the values you want in your examples but this update should not change too much the other examples that use the function (except for the fill of the densities it is much better than the previous plot).
There was a problem hiding this comment.
I made the default behaviour basically identical to the old ones (just quality upgrades, everything is in the same place), and the 'xy' version can be used with plot_style which defaults to yx. Colours also default to blue and red :)
…nsity backend + gmm_density does the meshgrid stuff
TODO
nx.detnx.detot.plot.plot1D_matupdateot.plotnx.sqrtmot.gmmadditions (hopefully)Types of changes
ot.gmmfor Gaussian Mixture Model OT, with a GMM plan example and a GMM flow exampleot.gaussianot.utils.proj_SDPfor a broadcastable and backend-compatible projection of a symmetric matrix onto the SDP conenx.detnx.sqrtmot.plot.proj1D_matwith the same API with visual improvementsMotivation and context / Related issue
new features implementing (GMM OT)[https://arxiv.org/pdf/1907.05254] +
ot.plot1D_mapimprovementHow has this been tested (if it applies)
pystestfull coverage + two example in the galleryPR checklist