Small Angle Scattering recommendation system based on CNNs and trained from Virtual Experiments

This repository is related to the paper "Learning from virtual experiments to assist users of Small Angle Neutron Scattering in model selection"

The authors of the paper are Jose Robledo, Peter Willendrup, Henrich Frielinghaus, and Klaus Lieutenant, and the code was written by Jose Robledo.

Note current software works well with PyTorch 2.2.1 .

If you found this repo useful, please cite us.

Prerequisites

• python 3
• CPU or NVIDIA GPU (recommend GPU for acceleration of training, but testing may be done in the CPU)

Installation

• Clone this repo:

git clone https://github.com/jorobledo/sas_recommendation_model
cd sas_recommendation_model

• Install PyTorch and other dependencies:

pip install -r requirements.txt

Download dataset

The full dataset of virtual experiments at the KWS-1 beamline of the FRM-II reactor in Garching used in our paper is published in Zenodo and can be downloaded here.

• Download the KWS-1 virtual experiments dataset from zenodo by running:

bash ./utils/get_data.sh

This will download the full dataset in the data folder. Make sure that the script is run from the current folder for a correct download path. If not, modify the download path in ./utils/get_data.sh to the desired location.

note: The dataset is large, therefore this might take a while.

Instructions

To run the models as in the paper you will need the data and the pretrained parameters. Assuming you already downloaded the data and that it is in the data folder of repository directory and that you are standing in the main repo directory, you can run the following command to download the pre-trained parameters:

bash ./utils/get_pretrained.sh all

The keyword all may be changed to resnet, densenet, or inception to download only the specified model parameters. They will be downloaded in the parameters folder (in which test.py will search for them).

Once you are on a working environment with pytorch installed you can run

python test.py --batch_size=200 --avg_batches=5 --num_workers=4

Setting a smaller batch size will run faster, as well as averaging on less amount of batches. You can also set the number of workers for the dataloader (check the test.py script for further understanding). This should print out the average Top-1, Top-3, and Top5 accuracies for all of the models.

Citation

If you use this code for your research, please cite our paper.

[1] J.I. Robledo, H. Frielinghaus, P. Willendrup and K. Lieutenant, Learning from virtual experiments to assist users of Small Angle Neutron Scattering in model selection, Scientific Reports 14 (2024), 14996. doi:https://doi.org/10.1038/s41598-024-65712-y.

BibTex:

@article{robledo2024learning,
  title={Learning from virtual experiments to assist users of Small Angle Neutron Scattering in model selection},
  author={Robledo, Jos{\'e} Ignacio and Frielinghaus, Henrich and Willendrup, Peter and Lieutenant, Klaus},
  journal={Scientific Reports},
  volume={14},
  number={1},
  pages={14996},
  year={2024},
  publisher={Nature Publishing Group UK London}
}