AC-STEM atom localization

Atom localization on AC-STEM (Aberration-Corrected Scanning Transmission Electron Microscopy) images.

Requirements

This code uses python 3.7.5 (defined by the .python-version file in case you using pyenv). Highly recommended to work using a virtual environment (virtualenv venv && source venv/bin/activate). Install requirements via:

pip install -r requirements.txt

Run

To replicate SAC_CNN results reported in our publication, use the following script:

/bin/bash scripts/dl_replicate_results.sh

This will run the original SAC_CNN model models/model_existing.ckpt.

Alternatively, to re-run the entire pipeline use the following command:

/bin/bash scripts/dl_train_evaluate.sh

This includes all stages:

  1. Generate a crops dataset.
  2. Training a SAC-CNN architecture.
  3. Inference using the trained model.
  4. Evaluate performance results.

Custom executions

It is possible to run SAC-CNN in your own data. To do so, use python commands with input arguments as follows:

PYTHONPATH=$PROJECTPATH python atoms_detection/dl_detection.py dataset/my_custom_dataset.csv

where dataset/my_custom_dataset.csv is a CSV file specifying to all images that will be used to run detection. All images must be in TIF format and must be included inside the data/tif_data folder. The CSV file should be formatted as follows:

Filename,Coords,Split
my_custom_image_1.tif,,test
my_custom_image_2.tif,,,test
my_custom_image_3.tif,,,test
my_custom_image_4.tif,,,test
...

Credits

  • High Performance Artificial Intelligence (HPAI) group, Barcelona Supercomputing Center (BSC).
  • Department of Chemistry and Applied Biosciences, ETH Zurich.
  • School of Chemical and Process Engineering and School of Chemistry, University of Leeds.
  • SuperSTEM Laboratory, SciTech Daresbury Campus.
  • Department of Physics, University of York, Heslington.
  • School of Chemical and Process Engineering and School of Physics, University of Leeds.
  • Institute of Chemical Research of Catalonia and The Barcelona Institute of Science and Technology.