Segmentation fault (core dumped)

[ncyue2]

Description: Hi, I ran into a segmentation fault (core dumped) when trying to train the GTM-decon using the provided scRNA-seq data (/data/reference_data.csv)

Changes that is made to the github code:

• change from g++-8 to g++ in ./gtm-decon-code/Makefile;
  • the modified code now looks like gpp := g++ -std=c++17 -I/usr/local/boost-1.68.0/include -I/usr/local/include -L/usr/local/lib and complied with make

Steps to reproduce: the app directory is the forked gtm-decon repo.

1. Enter the directory ./data/ to unzip the "tutorial data.tar.gz" to get the "reference_data.csv", with the code tar -xf tutorial_data.tar.gz
2. Set current directory: ../../standard, and create a new directory named outputs.
3. Enter the outputs directory and create another new directory named splits.
4. Return to the directory: standard
5. Run the preprocess steps:
   • python3 ../app/pre_processing/prepare_single_cell_input_PI_Segerstolpe.py --path_input ../app/data/reference_data.csv --path_save ./outputs/
   • python3 ../app/pre_processing/train_test_validation_split.py --path_input ./outputs/ --path_save ./outputs/splits/
   • ../app/gtm-decon-code/singleCellInput ./outputs/splits/counts_matrix_pp_train.tab ./outputs/splits/cell_labels_oh_train.csv ./outputs/splits/ 1
   • ../app/gtm-decon-code/singleCellInput_TestData ./outputs/splits/counts_matrix_pp_test.tab ./outputs/splits/testData.txt
   • ../app/gtm-decon-code/singleCellInput_TestData ./outputs/splits/counts_matrix_pp_validation.tab ./outputs/splits/validationData.txt

6. Run the training step: (Segmentation fault occurs)

• ../app/gtm-decon-code/gtm-decon -f ./outputs/splits/trainData.txt -m ./outputs/splits/meta.txt -trp ./outputs/splits/priorData.txt -k 14 -i 5 --inferenceMethod JCVB0 --maxcores 6 --outputIntermediates
• also tried ../app/gtm-decon-code/gtm-decon -f ./outputs/splits/trainData.txt -m ./outputs/splits/meta.txt -trp ./outputs/splits/priorData.txt -k 12 -i 5 --inferenceMethod JCVB0 --maxcores 6 --outputIntermediates

Screenshot:

Other information: Operating system: Fedora Linux version: 6.4.12-200.fc38.x86_64

Packages installed:

Name : armadillo
Version : 10.8.2
Release : 3.fc38
Architecture : x86_64
Size : 103 k
Source : armadillo-10.8.2-3.fc38.src.rpm
Repository : @System
From repo : fedora

Name : openblas Version : 0.3.21 Release : 4.fc38 Architecture : x86_64 Size : 82 k Source : openblas-0.3.21-4.fc38.src.rpm Repository : @System From repo : anaconda

Name : openblas-devel Version : 0.3.21 Release : 4.fc38 Architecture : x86_64 Size : 1.6 M Source : openblas-0.3.21-4.fc38.src.rpm Repository : @System From repo : fedora

Name : lapack
Version : 3.11.0
Release : 2.fc38
Architecture : x86_64
Size : 14 M
Source : lapack-3.11.0-2.fc38.src.rpm
Repository : @System
From repo : fedora

Name : lapack-devel Version : 3.11.0 Release : 2.fc38 Architecture : x86_64 Size : 1.6 M Source : lapack-3.11.0-2.fc38.src.rpm Repository : @System From repo : fedora

Name : arpack
Version : 3.8.0
Release : 6.fc38
Architecture : x86_64
Size : 633 k
Source : arpack-3.8.0-6.fc38.src.rpm Repository : @System From repo : fedora

Name : arpack-devel Version : 3.8.0 Release : 6.fc38 Architecture : x86_64 Size : 23 k Source : arpack-3.8.0-6.fc38.src.rpm Repository : @System From repo : fedora

Name : SuperLU Version : 5.3.0 Release : 4.fc38 Architecture : x86_64 Size : 471 k Source : SuperLU-5.3.0-4.fc38.src.rpm Repository : @System From repo : fedora

Name : SuperLU-devel Version : 5.3.0 Release : 4.fc38 Architecture : x86_64 Size : 101 k Source : SuperLU-5.3.0-4.fc38.src.rpm Repository : @System From repo : fedora

[aosakwe]

Hi @ncyue2,

Thank you for using gtm-decon.

I looked into your issue running the code through the CLI and jupyter notebook and was able to get the model to run on both. Could you try the following?

• run make clean in the /gtm-decon-code/ folder to remove the *.o files
• run make again to recompile. It is possible that some of the *.o files are not being recompiled as required.
• check that the splits folder does contain the required files for training (trainData.txt, meta.txt, PriorData.txt).

I tested the jupyter notebook and the code you ran on armadillo version 9.900.2 and 12.6.4. If the above bullet points do not help I would suggest trying one of these versions and seeing if that helps. Feel free to reach out if you have more questions.

Edit: Also just tested it out with Armadillo 10.8.2 (what you used). If all of the above doesn't work then you may want to revisit how you installed armadillo and share the steps you followed here.