linearGaussian_snre_test fails

[bkmi]

Describe the bug
The following tests fail (stochastically) since the training data produces one Inf for each of them.

test_api_snre_multiple_trials_and_rounds_map[SNRE_B-1]
test_api_snre_multiple_trials_and_rounds_map[SNRE_C-1]

To Reproduce
Please add a minimal code example that reproduces the problem:

1. Name Python version and SBI version: Python 3.8.18 git commit 105008efbacf320758e43a5db5bee4e3a0f9de36
2. Minimal code example pytest tests/linearGaussian_snre_test.py -k test_api_snre_multiple_trials_and_rounds_map
3. [Optional]: error message

Expected behavior
should not create an Inf

Additional context
I am working on #966 , but it shouldn't affect this test.

[famura]

I just ran pytest tests/linearGaussian_snre_test.py -k test_api_snre_multiple_trials_and_rounds_map on the current main (commit b81420b). The first 2 tries passed, on the 3rd one I got. This error seems to be non-deterministic.

FAILED tests/linearGaussian_snre_test.py::test_api_snre_multiple_trials_and_rounds_map[SNRE_B-1] - ValueError: Found 0 NaN simulations and 2 Inf simulations.SNRE does not allow invalid simulations. Replace the invalid values with an unreasonably low or hig...

@michaeldeistler I think that this is related to what I approached you about, even though the tests in our PyMC PR have been successful afterward. However, now I think we just got lucky 😅

[manuelgloeckler]

Just happened on the main branch :o Rerunning the workflow fixed it, so yeah seems to be stochastic.

See: https://github.com/sbi-dev/sbi/actions/runs/8410614947/attempts/1

Seems to be related to Inf/Nan, probably in multi-round SNRE. Likely related to #1060

[famura]

Check out my PR #1107 @manuelgloeckler @bkmi. I think I am about to fix this