summaryrefslogtreecommitdiff
path: root/ot
AgeCommit message (Collapse)Author
2017-07-24add test sinkhornRémi Flamary
2017-07-21Added and passed tests for dual variablesarolet
2017-07-21Compute cost with primalarolet
2017-07-21Added dual variables computationsarolet
2017-07-21Cleaned optimal plan and optimal cost computationarolet
2017-07-20pep8Alexandre Gramfort
2017-07-20pt not pltAlexandre Gramfort
2017-07-20pep8Alexandre Gramfort
2017-07-20pep8Alexandre Gramfort
2017-07-20pep8 + pimp plot1D_mat renderingAlexandre Gramfort
2017-07-14Removed some references to node_id_typearolet
node_id_type is really always int, it makes code hard to read though. In lemon they needed the typedef because they have more complicated graphs.
2017-07-14Removed unused variable maxarolet
Probably a legacy normalization variable
2017-07-13Added optional maximal number of iterationAntoine Rolet
2017-07-13Removed references to matlabAntoine Rolet
Also: - added error message when maxiter is reached - added debug logs
2017-07-12add init WDARémi Flamary
2017-07-11update doc gpu fixes #9Rémi Flamary
2017-07-11update doc gpu fixesRémi Flamary
2017-07-11add documentationRémi Flamary
2017-07-11Merge branch 'master' of github.com:rflamary/POTRémi Flamary
2017-07-11version 0.3.1Rémi Flamary
2017-07-10correction on Win64 bugNicolas Courty
2017-07-07add sinkhorbn2 +v3Rémi Flamary
2017-07-05add FDA for comparisonRémi Flamary
2017-07-03doc updateRémi Flamary
2017-07-03debug bregman stabilizedRémi Flamary
2017-06-20add clean zeros function for sparse distributionsRémi Flamary
2017-06-13implement paralell sinkhorn stabilizedRémi Flamary
2017-06-13implement paralell sinkhornRémi Flamary
2017-06-13make sinkhorn more general with method selectionRémi Flamary
2017-06-09add doc and correct encodingRémi Flamary
2017-06-09add doc ot.gpu.bregmanRémi Flamary
2017-04-24Remove unnecessary parameterLeo gautheron
2017-04-24add GPU implementation sinkhorn lpl1Leo gautheron
2017-04-21performance improvement sinkhorn lpl1Leo gautheron
- instead of updating individually for each target examples, update for all target examples at once using numpy functions. This allows for a faster computation (for me, divided by 4 on 3000*100 random matricies and random labels in [0,1]). - if I understoud correctly, a value of -1 in the array labels_a meant that we didn't have a label for this example. But in machine learning, we often encounter the binary case where we say we have the positive class (+1) and negative class (-1); thus with a dataset like this, the algorithm wouldn't work as expected. I replaced the default value for 'no label' to '-99' instead of '-1', and I added a parameter to modify it.
2017-04-21small normalization bugRémi Flamary
2017-04-20more changes from feebackLeo gautheron
in addition add the posibility to normalize the cost matrix through the function fit
2017-04-20missing argument function sinkhornLeo gautheron
2017-04-20changes from feedbackLeo gautheron
2017-04-20sinkhorn GPU implementationLeo gautheron
2017-04-18little cleanup sinkhornLeo gautheron
2017-04-18Performance improvement sinkhornLeo gautheron
Doing the computation this way is equivalent and allows to reduce the space complexity required from O(max(a, b)^2) to O(a*b) (especially usefull to transport a small number of sources example to a lot of target) This also allows to decrease the computation time.
2017-04-13update bregman fileLeo gautheron
- change commented prints to python3 compatible syntax - Correct issue that could cause the sinkhorn algo to stop with u and v containing nan/infinite numbers: - Assign uprev and vprev before changing u and v. - Then update u and v. - Then check if u and v contain nan, but ALSO infinite values. - if there are issues, then display error (with 2 r, not 3 :p) along with the iteration number (there may have errors at iteration 0)
2017-04-07doc updateRémi Flamary
2017-04-07V0.2Rémi Flamary
2017-04-07doc wdaRémi Flamary
2017-04-07better dicumentationRémi Flamary
2017-04-07add WDARémi Flamary
2017-04-07avoid filename conclict in windowsRémi Flamary
2017-03-21cleanupt cpp wrapper nameRémi Flamary
2017-03-14add emd2+ multiprocRémi Flamary