The function 'cost_normalization' do not use the backend and only use numpy

[framunoz]

Describe the bug

I was using the ot.da.SinkhornTransport class when I got the following warning:
UserWarning: Warning: numerical errors at iteration 0.

Thinking about why I got this error, I remembered that cost matrices should be normalised, so I changed the norm parameter to "max" to normalise it to the maximum value. However, this gave me an error.

The reason was that the utils.cost_normalization function is implemented only using numpy, when I'm currently using pytorch tensors. Internally, I got the backend of the C array that uses this function, and changed the line C /= float(np.max(C)) to C /= float(nx.max(C)), and then the program worked fine.

I mention this mainly because it may be a bug that happens to someone else in another context.

Environment (please complete the following information):

• OS (e.g. MacOS, Windows, Linux): Window
• Python version: 3.10
• How was POT installed (source, pip, conda): pip
• Build command you used (if compiling from source):
• Only for GPU related bugs:
  • CUDA version:
  • GPU models and configuration:
  • Any other relevant information:

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__)

Windows-10-10.0.19045-SP0
Python 3.10.4 (tags/v3.10.4:9d38120, Mar 23 2022, 23:13:41) [MSC v.1929 64 bit (AMD64)]
NumPy 1.24.1
SciPy 1.9.3
POT 0.8.2

[rflamary]

Good point and your fix seems perfect, those are the edge cases that were not detected when converting to backend.

Could you do this short PR for us please (and add a test that check that normalization does not breaks with non-numpy arrays)?

[framunoz]

Hi,

I was writing the code to try to fix this bug in the branch "cost_normalization_debug". In order to fix this bug, I had to create the median method in the backend class, and implement this method in the rest of the backends. However, I realised that tensorflow doesn't have a median function, although it may be replicable with the percentile function (as seen in this Stack Overflow question Tensorflow median value), but this requires the tensorflow_probabilities library to be installed. Also, I couldn't run the test scripts, as I didn't get the Jax library installed correctly.

Should I proceed with the PR even though there is still a method missing in the tensorflow backend to implement?

[rflamary]

hello, thanks @framunoz ,

you ca,n open a pull request and we can run all texts on gitrhub action on all backends (jax included). Note that It is also OK to test for simple men/max normalization and to rais a not implemented error for median normalization if tensorflow backend (which would be a simpler PR)

[framunoz]

Hello,

I tried to do the PR, but I got the following message:

Can't create a new pull request: Push failed: remote: Permission to PythonOT/POT.git denied to framunoz. unable to access 'https://github.com/PythonOT/POT.git/': The requested URL returned error: 403

EDIT: Nevermind, I do the PR here: #472