FAQ

This page is not meant to be an extensive documentation of how to address all questions that may arise when using pySIPFENN as a result of things outside our control, such as interpreter configuration, permissions, operating system issues. These can often be resolved by searching the internet for the error message you are receiving.

First things to check

In case you are having issues, it is good to first check the following few things that helped other users.

How to see if it works?

The easiest way to see if pySIPFENN is installed correctly is to run the following command:

python -c "import pysipfenn; print(pysipfenn.__version__)"

Then, to see if its core functionalities are working, you can run a small profiling script that will featurize (calculate KS2022 feature vector of) a typical atomic structure we use to profile its speed:

python -c "from pysipfenn.descriptorDefinitions import KS2022; KS2022.profile()"

You can also profile in parallel to see if your system is configured correctly for parallelization:

python -c "from pysipfenn.descriptorDefinitions import KS2022; KS2022.profileParallel()"

If you got to this point, it means that pySIPFENN is almost certainly working correctly, because most complex operations that could go wrong didn’t. If something did go wrong, look through this FAQ and the installation instructions to see if you can find a solution. If you still can’t, please open an issue or email ak@psu.edu and we will be happy to help you.

Now, you can try to run the tutorial notebook to see if you are able to run SIPFENN models and make predictions by opening the example.workshop2023Feb/sipfenn_examples_clean.ipynb notebook in Jupyter Notebook or Jupyter Lab and following the instructions. To see how this notebook should look like after successful execution, you can look at the example.workshop2023Feb/sipfenn_examples.ipynb or this page in the pySIPFENN documentation.

System and Architecture

pySIPFENN should generally run well on any operating system and computer architecture. In addition to automated CI testing through GitHub Actions across platforms and Python version every time code is modified (see pySIPFENN Repository Actions), we and our collaborators test it on 6 common combinations, which are listed below:

• Windows 10/11 64-bit on x86_64 (i.e. Intel/AMD)

• Windows 11 on ARM64 (e.g. Surface Pro X, Parallels VM on M1/M2 Macs)

• Linux on x86_64 (i.e. Intel/AMD)

• Linux on ARM64 (e.g. Raspberry Pi 4, Jetson Nano, etc.)

• MacOS on x86_64 (most Macs from before 2021)

• MacOS on ARM64 (Macs with M1/M2 chip)

Thus, it is unlikely that your operating system (PC vs Mac) or architecture (Intel vs M1/M2 Mac) is the issue. However, if you are using a different platform than above, please let us know, and we will try to help you.

Python Version

Make sure you are using the correct version of Python. pySIPFENN requires Python 3.9 or higher. We recommend 3.10 for longer support. The 3.11 is officially supported starting from pySIPFENN v0.12.0 (April 2023).

If you are using a different version, you can either install a new version of Python or use a virtual environment to install pySIPFENN in. We recommend using Conda to create a virtual environment. See the installation instructions making sure to specify the version of Python.

Up-to-date Conda / pip

Some users had problems with (1) loading the models or (2) making predictions using them and we traced the issue to an outdated version of Conda installed on their work station by their IT department. Updating Conda, pip, setuptools, and reinstalling pySIPFENN fixed the issue. You can try:

conda update -n base conda -c anaconda
conda update pip
pip install --upgrade setuptools
pip install --upgrade --force-reinstall pysipfenn

If you downloaded the models before, they should be retained as long as you do not change the Conda environment. If you do, you will have to download them again.

Model Download

All default pySIPFENN models are hosted in our Zenodo repository (linked in the main page of this documentation and GitHub README). They are ready to be downloaded automatically by pySIPFENN by simply running the downloadModels() function over any Calculator instance and will be permanently stored in the pySIPFENN installation directory (inside the models folder). Printing the Calculator instance will show you the exact path where the models are stored if you want to check it.

If you have issues downloading the model files in reasonable time, this is usually correlated with a workshop, after which many users try to download them from Zenodo concurrently and exhaust the bandwidth. If that is the case, please try to download the models in a few hours or the next day; otherwise, please let us know, and we will try to help you by providing the files directly.

We also received a report that some MacOS installations of python itself can lack SSL certificate support out-of-the-box and it needs to be enabled manually to enable downloading pySIPFENN models from Zenodo, with the following one-line command (after replacing Python 3.9 with your Python version):

open /Applications/Python\ 3.9/Install\ Certificates.command

Known problems with easy solutions

Torch and ONNX Issues

• Some users recently reported getting ImportError: cannot import name 'COMMON_SAFE_ASCII_CHARACTERS' from 'charset_normalizer.constant' error when using both recent (mid-2023) torch and Mac OS on M1/M2 Macs. It seems to be some dependency issue between onnx2torch and torchvision, but it can be quickly solved with installing missing chardet package with pip as

    pip install chardet
  

More Complex Issues

Out-Of-Memory Error / Models Cannot Load

RAM requirements to run:

While pySIPFENN shouldn’t have much trouble running on most modern desktop and portable computers, it may have issues on older machines or on machines with limited RAM such as VMs on cloud instances. In general, to run the models, you will need:

• 512MB of RAM for Light Model (NN24)

• 2.5GB of RAM for KS2022 Novel Materials Model (NN30)

• 6GB of RAM to run all default models (NN9, NN20, NN24, NN30) and follow the workshop tutorial

RAM requirements to load:

Now, one needs to consider how much memory is needed to get the models loaded into the memory. This is a bit complicated, but in general, your system will need to use around twice the memory given above.

At the same time, if you are using any modern version of Windows or MacOS, you shouldn’t have any issues with this limits if your system has at least 8GB of RAM (assuming no memory-intensive applications running in the background) because they will utilize pagefile (Windows) or swap (MacOS) to automatically store the temporary data (as much as needed) that does not fit into RAM.

Memory on Linux

However, if you are using Linux, the swap is not dynamic and if it’s smaller than the amount of memory you need, you will get an error or the process will just be ‘killed’ by the system. You can check its size with

free -h

And if the sum of RAM and swap is less than 12GB, you likely won’t be able to use all of pySIPFENN models. Fortunately, it’s an easy fix for your administrator, or if you have sudo access, yourself. To temporarily increase swap you just need to allocate some space on your drive (e.g. 8GB)

sudo fallocate -l 8G /swapfile

limit access permissions for security reasons

sudo chmod 600 /swapfile

tell Linux that it is a swap file

sudo mkswap /swapfile

and enable it

sudo swapon /swapfile

Now, you should have no memory issues with running all pySIPFENN networks as of the time of writing (NN9, NN20, NN24, NN30). Please note that, as mentioned, this swap change is temporary and will disappear when you reboot. To make it permanent, you can follow one of many guides available online.

Errors related to initializing OpenMP library

This error was reported to us by pySIPFENN user running Windows 11 64bit, but it wasn’t seen by our team at any point on any of our 12 test systems and doesn’t seem to be related to pySIPFENN directly, but to some conflict between scikit-learn and numpy, possibly because of problems with base conda installation. If you see it, we would recommend you re-install conda and recreate environments to avoid unexpected behavior of pySIPFENN and other Python tools you are using. However, if you are in a hurry and want to try this not recommended fix, you can try to run the following commands before importing pySIPFENN:

import os
os.environ['KMP_DUPLICATE_LIB_OK']='True'

Which fixed the issue for the user who reported it. If you also run into this issue, we would appreciate if you could let us know and we will investigate it further.