Official implementation from Stanford University

• Title: Missing Wedge Completion via Unsupervised Learning with Coordinate Networks
  
• Authors: Dave Van Veen, Jesús G. Galaz-Montoya, Liyue Shen, Philip Baldwin, Akshay S. Chaudhari, Dmitry Lyumkis, Michael F. Schmid, Wah Chiu, John Pauly
  
• Contact: {vanveen, pauly} [at] stanford [dot] edu
  

1. Clone the repository:

git clone https://github.com/davevanveen/wucon.git
cd wucon

2. Use these commands to set up a conda environment:

conda env create -f env.yml
conda activate wucon

3. In src/constants.py, create your own project directory DIR_PROJECT which will contain input data, trained models, and generated output.
   • Due to the large amount of data, it's recommended to create DIR_PROJECT outside this code repository to avoid pushing data to GitHub.
4. Download data from this folder and set DIR_DATA in src/constants.py accordingly. More information on data is provided below.
5. (optional) To add your own dataset, follow the format of example datasets.

6. Run cd src; ./main.sh, after you've made the bash script executable.

Below describes a subset of scripts within src/:

• main.sh: generates reconstructions for a particular dataset and set of case_ids.
  • case_id defines a set of experiment parameters. We recommend starting with the default, case_id=0. Then if you wish to modify parameters for your own experiments, see cases in constants.py.
  • Results will be saved in {DIR_PROJECT}/out/{dataset}/c{case_id}.
• constants.py: sets configs for paths, datasets, and experiments.
• recon_volume.py: reconstructs volume chunks in series or parallel, depending on GPU availability, by making individual calls to recon_chunk.py.

• We provide three datasets from our paper here.
  • sph: Spheres of various sizes
  • geo: Geometric shapes
  • p2s: P22 simulated virus particle
• wucon_data_all.zip contains all three datasets. For convenience, we've provided a smaller file wucon_data_p2s.zip which contains only p2s.
• Each dataset contains the following files:
  • {dataset}_tgt.mrc: 3D image volume ("target").
    • Note this is not relevant for experimentally-acquired datasets, as the ground-truth "target" data does not exist.
  • {dataset}_prj.mrc: two-dimensional projections of the 3D volume.
  • tilt_angles.tlt: angles at which the projections were generated.