search

scikit-adaptation / skada

Domain adaptation toolbox compatible with scikit-learn and pytorch
https://scikit-adaptation.github.io/
BSD 3-Clause "New" or "Revised" License
54 stars 16 forks source link

[MRG] Implementation of the TJM method #140

Closed BuenoRuben closed 3 months ago

BuenoRuben commented 4 months ago

see: https://citeseerx.ist.psu.edu/document?repid=rep1&type=pdf&doi=a279f53f386ac78345b67e13c1808880c718efdf#page=3&zoom=100,144,841

codecov[bot] commented 4 months ago

Codecov Report

Merging #140 (63672d4) into main (ff06a8b) will decrease coverage by 14.16%. The diff coverage is 98.87%.

Additional details and impacted files ```diff @@ Coverage Diff @@ ## main #140 +/- ## =========================================== - Coverage 97.39% 83.24% -14.16% =========================================== Files 48 41 -7 Lines 4422 4225 -197 =========================================== - Hits 4307 3517 -790 - Misses 115 708 +593 ```
antoinecollas commented 4 months ago

Now the constraints are always verified and the loss decreases when tradeoff=0. However, when tradeoff is different from 0, the loss oscillates... So there is still a problem.

BuenoRuben commented 4 months ago

I managed to make the loss decrease when we have a tradeoff, but the conmition of one of the matrix being identity is still not solved, however the decision boundaries are closer to those of TCA, so we should be fine, however I don't understand why before, when the loss was oscillating, we had a better adapter (we had 100% acc)