This repository contains the code for our work Graph Filtration Learning which was accepted at ICML'20.

Installation

In the following <root_dir> will be the directory in which you have chosen to do the installation.

1. Install Anaconda from here into <root_dir>/anaconda3, i.e., set the prefix accordingly in the installer.

2. Activate Anaconda installation:

   source <root_dir>/anaconda3/bin/activate
   

3. Install pytorch via conda

   conda install pytorch=1.4.0 torchvision cudatoolkit=<your_cuda_version> -c pytorch
   

4. Install pytorch-geometric and its dependencies following the instructions on its gh-page.

5. Install torchph via

   cd <root_dir>
   git clone -b 'submission_icml2020' --single-branch --depth 1 https://github.com/c-hofer/torchph.git
   conda develop torchph
   

6. Clone this repository into <root_dir>.

Application

1. Generate the experiment configurations you want using the write_exp_cfgs_file.ipynb notebook. It is assumed that the notebook server is started in <root_dir>/graph_filtration_learning.

2. Use the train.py script to run the experiments, e.g.,

   python train.py --cfg_file <my_cfg.json> --output_dir <results/dir/path> --devices 0,1 --max_process_on_device 2 
   

   to use cuda device 0 and 1 with at most 2 experiments on each.

   Each experiment gets a unique id and its output is written to <results/dir/path> as a pickle file. Additionally for each CV run the corresponding trained model is dumped.

3. The notebook results.ipynb contains some code to browse the results.