Closed joseph-siefert closed 1 year ago
Hi Joe,
Thanks a lot for raising this issue.
It seems that there is an issue with the scipy sparse matrix structure. Could you please tell me what happens if you "densify" the matrix prior to running Sobolev Alignment, e.g.,:
data_an.X = np.array(data_an.X.todense())
If it resolves the problem, I can add it as a pre-processing step within the fit procedure.
Have a great day, Souf
Thanks Soufiane, this did get around that error. Now I have a different issue with Falkon. I have Pytorch 2.0.0 and CUDA 11.7, and I can access GPU and run scVI no problem. I have compiled Falkon with the wheel for these specific versions, which should be tested and supported but is not working. I am following up with them on it.
Thanks a lot. I will add that in the pre-processing pipeline, and in the tutorial.
For Falkon, it took me quite some trial-and-error to install it successfully. I gave some advice here: https://github.com/NKI-CCB/sobolev_alignment#issues-with-the-compiler I am not sure that cuda version plays a role here, but pytorch should be at 1.11 (2.0 is very different). The compiler must also be chosen differently. I'll follow the correspondence you have with them and add it to the README.
Falkon is working now, but I came across another error further down the line:
---------------------------------------------------------------------------
AttributeError Traceback (most recent call last)
Cell In[61], line 2
1 #Estimate number of shared principal vectors
----> 2 n_similar_pvs = sobolev_alignment_clf.permutation_test_number_similar_pvs(n_permutations=10)
AttributeError: 'SobolevAlignment' object has no attribute 'permutation_test_number_similar_pvs'
Hi Joe,
Great that the FalkonML team got it sorted so fast.
For this new issue, it seems that you have version 0.4.1, while the last release is 1.0.0. I fixed this bug in the last release normally. You can install it using pip install sobolev-alignment
command. You do not need to install Falkon again, as long as you re-install in the same environment.
Let me know if the problem remains.
Best, Soufiane
I installed through pip, and running the command gives:
Requirement already satisfied: sobolev-alignment in /mambaforge/envs/sobolev/lib/python3.9/site-packages (1.0.0)
yet:
tail -1 /mambaforge/envs/sobolev/lib/python3.9/site-packages/sobolev_alignment/__init__.py
__version__ = "0.4.1"
I tried to force reinstall:
pip install --upgrade --no-deps --force-reinstall sobolev-alignment
Looking in indexes: https://pypi.org/simple, https://pypi.ngc.nvidia.com
Collecting sobolev-alignment
Downloading sobolev_alignment-1.0.0-py3-none-any.whl (34 kB)
Installing collected packages: sobolev-alignment
Attempting uninstall: sobolev-alignment
Found existing installation: sobolev_alignment 1.0.0
Uninstalling sobolev_alignment-1.0.0:
Successfully uninstalled sobolev_alignment-1.0.0
Successfully installed sobolev-alignment-1.0.0
but still:
tail -1 /mambaforge/envs/sobolev/lib/python3.9/site-packages/sobolev_alignment/__init__.py
__version__ = "0.4.1"
Report
Hi Soufiane, the scVI training runs no problem, but the KRR gives the following error:
There are raw counts in both .X and .layers['counts'] for both source and target
Version information