matrix shapes problem with entropic_fused_gromov_barycenters

[youssef62]

Describe the bug

When I call entropic_fused_gromov_barycenters , I have the following error :
ValueError: shapes (10,1) and (10,1) not aligned: 1 (dim 1) != 10 (dim 0)

Screenshots

Code sample

n = 10 

g1 = nx.erdos_renyi_graph(n,0.5)
g2  = nx.erdos_renyi_graph(n,0.5)

c1 = nx.adjacency_matrix(g1).toarray().astype(np.float64)
c2 = nx.adjacency_matrix(g2).toarray().astype(np.float64)

y = [[[1]]*n]*n
y = np.array(y).astype(np.float64) 

f , c  = fgw_barycenters(n,y,[c1,c2],alpha=0.5 )
f , c  = entropic_fused_gromov_barycenters(n,y,[c1,c2],alpha=0.5)   
print(c)
print(f)

Expected behavior

entropic_fused_gromov_barycenters should act like fgw_barycenters .

Environment (please complete the following information):

• OS (e.g. MacOS, Windows, Linux): MacOS
• Python version:
• How was POT installed (source, pip, conda):pip

Output of the following code snippet:

import platform; print(platform.platform())
import sys; print("Python", sys.version)
import numpy; print("NumPy", numpy.__version__)
import scipy; print("SciPy", scipy.__version__)
import ot; print("POT", ot.__version__)

macOS-12.5.1-arm64-arm-64bit
Python 3.11.5 (main, Aug 24 2023, 15:09:32) [Clang 14.0.0 (clang-1400.0.29.202)]
NumPy 1.26.0
SciPy 1.11.3
POT 0.9.1

Additional context

I think the cause of the problem is in lines 986,987 in _bregman.py

Y = update_feature_matrix(lambdas, Ys_temp, T_temp, p)
Ms = [dist(Ys[s], Y) for s in range(len(Ys))]

dist takes entries in the form (ns,d) where update_feature_matrix returns ouputs of dimension (d,n). Either a transpose of Y or a change or the output type of update_feature_matrix would solve it.

I wanted to do a pr to fix but I could not run the test locally. Do you just make test ?
this is the error I encounter when I do that :

[cedricvincentcuaz]

Hello @youssef62 ,
thank you for your feedback. Indeed the transpose should be considered for the features. Tests were not helpful to spot this as we had n=d...
I also spotted other flaws in these functions, I will open an issue for this matter.

Then if you want to contribute to POT in the future: you'll need to make your install able to read .pyx file.
I think that running this command line in the POT folder should do the trick:
python3 setup.py build_ext --inplace

And then it's explained in the POT doc how to go for a PR. Including local tests, pep8 etc. If you still have doubts you can check at the command lines to run tests in an existing PR ;)

[youssef62]

Thank you for the quick reply. I also noticed that after adding the transpose, there were no errors but the results were always matrices with the same number on all entries.
Otherwise , do you know why I have that message when I run the tests ? should I ask this question in a discussion ?

[cedricvincentcuaz]

I edited my previous message about your message error.
Then for your results, there might be many reasons that you should check to be sure that it is a bug and not just the intended behaviour:

• epsilon plays a crucial role for entropic solvers. Note that using solver through **kwargs will use different entropic FGW solvers.
• sensitivity to init_X and alpha.