Conversation
Codecov Report
@@ Coverage Diff @@
## master #249 +/- ##
==========================================
+ Coverage 92.10% 93.06% +0.95%
==========================================
Files 16 17 +1
Lines 3054 3445 +391
==========================================
+ Hits 2813 3206 +393
+ Misses 241 239 -2 |
|
Just updated the doc one last time, it should build ;). The PR feels OK, wdyt @ncourty ? |
|
|
||
| lst_tot = [] | ||
|
|
||
| for nx in [ot.backend.NumpyBackend(), backend]: |
There was a problem hiding this comment.
NumpyBackend is not already in backend_list? if not so why?
There was a problem hiding this comment.
because we run it for numpy and the current backend and check that they provide the same output (see below the loop)
| v = rnd.randn(10) | ||
| c = rnd.randn(1) | ||
|
|
||
| if torch: |
There was a problem hiding this comment.
the fact that you have to branch for torch here suggests that the backend mechanism
is not consistent
There was a problem hiding this comment.
for gradients no we cannot be consistent because each backend handes gradients differently. The fact that we have a branch mans that we can handle different ways to compute the gradients (when they can be computed).
if you want to test if the backend works, you have to do a test utiings functiosn specfis to ecah backend.
Co-authored-by: Alexandre Gramfort <[email protected]>
Co-authored-by: Alexandre Gramfort <[email protected]>
Co-authored-by: Alexandre Gramfort <[email protected]>
Co-authored-by: Alexandre Gramfort <[email protected]>
Co-authored-by: Alexandre Gramfort <[email protected]>
Co-authored-by: Alexandre Gramfort <[email protected]>
Co-authored-by: Alexandre Gramfort <[email protected]>
Co-authored-by: Alexandre Gramfort <[email protected]>
Co-authored-by: Alexandre Gramfort <[email protected]>
|
here is the new example. I love how simply the Wasserstein loss can be used and optimized. Just corrected the legend in the last figure. Will merge when tests pass. |
|
really cool !
… |
Types of changes
We are working on a new multi-backend API for POT with the objective to be able to handle numpy/torch/jax/cupy arrays seamlessly.
This will be a big one.
Expected features :