An Error (RuntimeWarning: invalid value encountered in log) in ot.da.SinkhornL1lL2 after some iterations

[EnayatAria]

Using ot.da.SinkhornL1l2Transport for a domain adaptation problem, I faced an error as follows:

Datasets used:

Xs.txt
Xt.txt
ys.txt

** To Reproduce**

If you download the input files in C:\ , then the code is:

import numpy as np
import to

Xs = np.loadtxt("C: / Xs.txt").reshape(604, 5)
Xt = np.loadtxt("C: / Xt.txt").reshape(601, 5)
ys = np.loadtxt("C: / ys.txt")

ot_base = ot.da.SinkhornL1l2Transport(reg_e=10000, reg_cl=100, max_iter=100, verbose=True)
ot_base.fit(Xs=Xs, ys=ys, Xt=Xt)

Result and Error

 It.  |Loss        |Relative loss|Absolute loss
------------------------------------------------
0|5.193677e+06|0.000000e+00|0.000000e+00
1|3.150847e+05|1.548343e+01|4.878593e+06
2|2.668420e+05|1.807914e-01|4.824274e+04
3|2.663638e+05|1.795333e-03|4.782117e+02
4|2.663590e+05|1.786689e-05|4.759007e+00
5|2.663590e+05|1.312580e-07|3.496174e-02
6|2.663588e+05|7.339658e-07|1.954982e-01
7|2.663106e+05|1.808094e-04|4.815146e+01

C:\Users\enayat.aria\PycharmProjects\pythonProject\venv\lib\site-packages\ot\optim.py:357: RuntimeWarning: 
invalid value encountered in log
return np.sum(M * G) + reg1 * np.sum(G * np.log(G)) + reg2 * f(G)
Traceback (most recent call last):
File "C:\Program Files\JetBrains\PyCharm Community Edition 2021.1.2\plugins\python- 
ce\helpers\pydev\pydevd.py", line 1483, in _exec
pydev_imports.execfile(file, globals, locals)  # execute the script
File "C:\Program Files\JetBrains\PyCharm Community Edition 2021.1.2\plugins\python- 
ce\helpers\pydev\_pydev_imps\_pydev_execfile.py", line 18, in execfile
exec(compile(contents+"\n", file, 'exec'), glob, loc)
File "C:/Users/enayat.aria/PycharmProjects/pythonProject/OT_for_DA.py", line 258, in <module>
ot_base.fit(Xs=Xs, ys=ys, Xt=Xt)
File "C:\Users\enayat.aria\PycharmProjects\pythonProject\venv\lib\site-packages\ot\da.py", line 1950, in fit
returned_ = sinkhorn_l1l2_gl(
File "C:\Users\enayat.aria\PycharmProjects\pythonProject\venv\lib\site-packages\ot\da.py", line 239, in 
sinkhorn_l1l2_gl
return gcg(a, b, M, reg, eta, f, df, G0=None, numItermax=numItermax,
File "C:\Users\enayat.aria\PycharmProjects\pythonProject\venv\lib\site-packages\ot\optim.py", line 388, in gcg
G = G + alpha * deltaG
TypeError: unsupported operand type(s) for *: 'NoneType' and 'float'

Checking the parameters, I found that the G matrix obtained in the optim.py code has negative values in the last iteration; due to the last update.

Please let me know how to solve the problem, or if I should provide more information.

Best,

Environment (please complete the following information):

• OS (e.g. MacOS, Windows, Linux): Windows 10
• Python version: 3.9
• How was POT installed (source, pip, conda): pip

[hichamjanati]

@EnayatAria can you elaborate ? your 'to reproduce' section is missing

[EnayatAria]

Hi. I have just updated the problem with a clearer title. Looking forward to hearing from you Thanks

…

On Tue, Nov 16, 2021 at 3:19 PM Hicham Janati ***@***.***> wrote: @EnayatAria <https://github.com/EnayatAria> can you elaborate ? your 'to reproduce' section is missing — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#311 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ALAEZF74JPVORVIITYA4USDUMJR77ANCNFSM5IEK756A> . Triage notifications on the go with GitHub Mobile for iOS <https://apps.apple.com/app/apple-store/id1477376905?ct=notification-email&mt=8&pt=524675> or Android <https://play.google.com/store/apps/details?id=com.github.android&referrer=utm_campaign%3Dnotification-email%26utm_medium%3Demail%26utm_source%3Dgithub>.

[hichamjanati]

can you provide a small code snippet to reproduce the error so i can debug it ?

[EnayatAria]

The code is a simple call of the function

ot_base = ot.da.SinkhornL1l2Transport(reg_e=10000, reg_cl=100, max_iter=100, verbose=True)
ot_base.fit(Xs=Xs, ys=ys, Xt=Xt)

The error is not from the code, but from the input values; since it is working with another dataset. However, when I changed the data set it sends the error after a couple of iteration.

[EnayatAria]

Datasets used:

Xs.txt
Xt.txt
ys.txt

The Xs, Ys, and Xt are attached. Xs = {ndarray: (604, 5)} ; Xt = {ndarray: (601, 5)} ; ys = {ndarray: (604, )}

Considering the dimensions the files should be reshaped such as:
original_Xs = np.loadtxt("Xs.txt").reshape(604, 5)

[EnayatAria]

To reproduce, the code and the datasets are updated in the first comment.

[hichamjanati]

These numerical errors are probably caused by the large values in your data, try normalizing them first or normalize the cost with the norm argument:

ot_base = ot.da.SinkhornL1l2Transport(reg_e=1, reg_cl=100, max_iter=100, verbose=True, norm="median")

hope this fixes your problem

[rflamary]

There should never be negative values in G along the optim iterations. It might come from the solver we should look into it.

[EnayatAria]

Thank you for the solution, but it just breaks the iterations and returns the final G with negative values from optim.py.

I am wondering if the result would be the optimal solution for the general regularized OT problem.

'alpha' is 'None' when none of the conditions in the scalar_search_armijo function in lineserach.py is met. There is a comment above the last 'return' saying that # Failed to find a suitable step length.

What does that mean? Does it mean that it can not converge?

alpha values for the provided datasets returned from line_search_armijo in optim.py are

iter alpha
0 0.99
1 0.99
2 0.99
3 0.99
4 0.99
5 -340.96758804319694
6 -271.4559025260829
7 None

The code again sent the same error as already mentioned

....\site-packages\ot\optim.py:357: RuntimeWarning: invalid value encountered in log
return np.sum(M * G) + reg1 * np.sum(G * np.log(G)) + reg2 * f(G)

but this time does not stop.