leiden and umap not reproducible on different CPUs
Nobody has claimed this yet.
- Dominant language
- Python
- Stars
- 2.6k
- Forks
- 779
- Avg merge
- 1d 4h
- Merged PRs (30d)
- 27
Description
- I have checked that this issue has not already been reported.
- I have confirmed this bug exists on the latest version of scanpy.
- (optional) I have confirmed this bug exists on the master branch of scanpy.
I noticed that running the same single-cell analyses on different nodes of our HPC produces different results.
Starting from the same anndata object with a precomputed X_scVI latent representation, the UMAP and leiden-clustering looks different.
On
- Intel(R) Xeon(R) CPU E5-2699A v4 @ 2.40GHz
- AMD EPYC 7352 24-Core Processor
- Intel(R) Xeon(R) CPU E7-4850 v4 @ 2.10GHz
adata.obs["leiden"].value_counts()
0 4268
1 2132
2 1691
3 1662
4 1659
5 1563
...
On
- Intel(R) Xeon(R) CPU E7- 4870 @ 2.40GHz
0 3856
1 2168
2 2029
3 1659
4 1636
5 1536
...
Minimal code sample (that we can copy&paste without having any data)
A git repository with example data, notebook and a nextflow pipeline is available here:
https://github.com/grst/scanpy_reproducibility
A report of the analysis executed on four different CPU architectures is available here:
https://grst.github.io/scanpy_reproducibility/
Versions
WARNING: If you miss a compact list, please try `print_header`!
-----
anndata 0.7.5
scanpy 1.6.0
sinfo 0.3.1
-----
PIL 8.0.1
anndata 0.7.5
backcall 0.2.0
cairo 1.20.0
cffi 1.14.4
colorama 0.4.4
cycler 0.10.0
cython_runtime NA
dateutil 2.8.1
decorator 4.4.2
get_version 2.1
h5py 3.1.0
igraph 0.8.3
ipykernel 5.3.4
ipython_genutils 0.2.0
jedi 0.17.2
joblib 0.17.0
kiwisolver 1.3.1
legacy_api_wrap 0.0.0
leidenalg 0.8.3
llvmlite 0.35.0
matplotlib 3.3.3
mpl_toolkits NA
natsort 7.1.0
numba 0.52.0
numexpr 2.7.1
numpy 1.19.4
packaging 20.7
pandas 1.1.4
parso 0.7.1
pexpect 4.8.0
pickleshare 0.7.5
pkg_resources NA
prompt_toolkit 3.0.8
ptyprocess 0.6.0
pycparser 2.20
pygments 2.7.2
pyparsing 2.4.7
pytz 2020.4
scanpy 1.6.0
scipy 1.5.3
setuptools_scm NA
sinfo 0.3.1
six 1.15.0
sklearn 0.23.2
sphinxcontrib NA
storemagic NA
tables 3.6.1
texttable 1.6.3
tornado 6.1
traitlets 5.0.5
umap 0.4.6
wcwidth 0.2.5
yaml 5.3.1
zmq 20.0.0
-----
IPython 7.19.0
jupyter_client 6.1.7
jupyter_core 4.7.0
-----
Python 3.8.6 | packaged by conda-forge | (default, Nov 27 2020, 19:31:52) [GCC 9.3.0]
Linux-3.10.0-1160.11.1.el7.x86_64-x86_64-with-glibc2.10
64 logical CPU cores, x86_64
-----
Session information updated at 2021-10-15 09:58
Contributor guide
First steps
- Read the whole issue, then the project's contributing guide.
- Comment on the issue to say you are picking it up — it saves two people doing the same work.
- Fork the repository and make your change on a branch.
- Open a pull request that references the issue number.
Research direction
Start with the linked scanpy_reproducibility repository and its report, then reproduce the UMAP and Leiden differences across the listed CPU architectures using the provided example data and pipeline. Compare the reported environments and execution results; done would require identifying a reproducible cause and documenting or validating a concrete resolution.
Written by the indexing model from the issue text.
Assessment
- Tech stack
- python
- Domain
- bioinformatics, machine-learning
- Issue type
- Bug
- Difficulty
- 5/5
- Estimated time
- Over a week
- Activity status
- Stale
- Clarity
- Needs clarification
- Newbie friendliness
- 25/100