Stellarscope bug reproduce

It seems that different runs of stellarscope on the same data are giving different results.

TO REPRODUCE

setup environment

mamba env create -n stbug environment.yml
conda activate stbug

pull the latest stellarscope, devmode branch

git clone git@github.com:nixonlab/stellarscope.git stellarscope_devmode.git
git fetch
git switch -c devmode origin/devmode
git rev-parse --short HEAD

The commit should be 520b379

install stellarscope into stbug environment

pip install -e .
cd ..

run the test script twice

the test script includes the commit hash in the output directory name and handles creating a new output directory when the current one exists.

. runtest1.sh
. runtest1.sh

final outputs are different

diff -q output.520b379.rep?/pseudobulk-TE_counts.mtx

Output:

Files output.520b379.rep1/pseudobulk-TE_counts.mtx and output.520b379.rep2/pseudobulk-TE_counts.mtx differ

Check intermediate results

The --devmode setting outputs several intermediate matrices that can be compared between runs.

diff output.520b379.rep?/pseudobulk-rawscores_before_fit.txt
diff output.520b379.rep?/pseudobulk-rawscores_after_fit.txt
diff output.520b379.rep?/pseudobulk-probs_after_fit.txt
diff output.520b379.rep?/pseudobulk-assignments_exclude.txt