Fix an issue in test_criterions: Most likely due to some change to the torch version (?), running the test_infonce_gradients test twice causes non-deterministic behavior. On the first run, non-existent gradients are None as expected, but on the second run, they are set to a zero torch array. I added a hotfix and accept both None and a zero array as possible values now, but this needs to deeper digging to fix properly.
The usage.rst docs had a typo following the intro of discrete multisession training ( #135 ), which was fixed by introducing a separate variable name for the discrete multisession model there.
Docs building failed. Removed pinning of scikit-learn and added a pin to numpy<2 instead. In a future PR, we should look into upgrading some of the docs packages like sphinx, the theme, etc.
Fix the tools/build_docker.sh script. I upgraded the docker container to a more recent python/cuda version and fixed the build process.
Fixes several issues with the test suite:
test_criterions: Most likely due to some change to the torch version (?), running thetest_infonce_gradientstest twice causes non-deterministic behavior. On the first run, non-existent gradients areNoneas expected, but on the second run, they are set to a zero torch array. I added a hotfix and accept bothNoneand a zero array as possible values now, but this needs to deeper digging to fix properly.usage.rstdocs had a typo following the intro of discrete multisession training ( #135 ), which was fixed by introducing a separate variable name for the discrete multisession model there.scikit-learnand added a pin tonumpy<2instead. In a future PR, we should look into upgrading some of the docs packages likesphinx, the theme, etc.tools/build_docker.shscript. I upgraded the docker container to a more recent python/cuda version and fixed the build process.Fix https://github.com/AdaptiveMotorControlLab/CEBRA/pull/150 Fix https://github.com/AdaptiveMotorControlLab/CEBRA/pull/171