Patched a potential segfault when appending simulations

[VictorSeven]

I had some problems when calling the ``append_simulations` method. In a Jupyter Notebook, it would kill the kernel every time without further notice. In a script, one can see this is due to a segmentation fault.

After some debugging, I spotted the problem when checking the Z-score of the data. In my case, the result of the simulation was two vstacked time series (external input and simulated time series; the tensor shape was (batch_size, 2, n_data)). For testing purposes, I was training the SBI only with a single realization of the external input, thus leading to 0 variance among all batches.

Despite this not being a very common use case, the fact that it kills the Python kernel makes the problem difficult to track for the user: there is no error message and the debugger does not work properly (it will just crash, not going inside the SBI code). Even print statements inside the function append_simulations would not print out in a notebook, being visible only when used on a script.

For this reason, I added a small extra check in the code. If this edge case arises, the computation of the Z-score is skipped, and now the user is adequately warned about the problem so it can be fixed when needed. From what I have seen code would run afterwards without problems.

[janfb]

Hi @VictorSeven, thanks for making this PR and sorry for the delayed response.
I am not sure I understand the use-case of training with only a single training-data point. But I agree that this edge case should not result in a segmentation fault and should be documented with a warning.

I suggest to just check whether num_unique == 1 in line 35.
If yes, raise the warning, if no perform the z-scoring and the original check implemented in the function.
Do you see what I mean?

At the moment the tests are failing at the black formatting check. You can fix this by running black filename (as well as isort and pyright) locally before committing.

Best regards, and thanks again for looking into this
Jan

[janfb]

thanks!
see comment in PR. tests are failing because of black formatting. See https://github.com/mackelab/sbi/blob/main/CONTRIBUTING.md#style-conventions for how to run formatting checks locally.

[VictorSeven]

Hello, sorry for the delay -was a bit sick last weeks.

I have updated the change as suggested to use num_unique instead. I run black and so on as indicated so I hope now everything should be correct. Let me know if there is any additional problem.

About use case: I understand the confusion about the use case of training with a single data point. In my case, my dynamical system was always subject to a single external input, which I was not passing to SBI, because it was always the same. Then, I wanted to train SBI to recognise several different external inputs. As a test, I rewrote the code where now SBI was aware of the external input, and I was to check if giving it only 1 would yield the same results as I had before. I understand it was maybe not intended usage and this led to an error, but the main problem was that the Python kernel would just silently crash, making it difficult to debug the problem.

[janfb]

looks good, thanks!

[codecov-commenter]

Codecov Report

Merging #803 (88f88e3) into main (2a114f4) will increase coverage by 0.08%.
The diff coverage is 100.00%.

📣 This organization is not using Codecov’s GitHub App Integration. We recommend you install it so Codecov can continue to function properly for your repositories. Learn more

@@            Coverage Diff             @@
##             main     #803      +/-   ##
==========================================
+ Coverage   74.76%   74.85%   +0.08%     
==========================================
  Files          80       80              
  Lines        6190     6199       +9     
==========================================
+ Hits         4628     4640      +12     
+ Misses       1562     1559       -3     

Flag	Coverage Δ
unittests	74.85% <100.00%> (+0.08%)	⬆️

Flags with carried forward coverage won't be shown. Click here to find out more.

Impacted Files	Coverage Δ
sbi/utils/sbiutils.py	80.22% <100.00%> (+0.46%)	⬆️

... and 22 files with indirect coverage changes

Help us with your feedback. Take ten seconds to tell us how you rate us. Have a feature suggestion? Share it here.