This repository contains the code for our work Graph Filtration Learning which was accepted at ICML'20.
In the following
<root_dir> will be the directory in which you have chosen to do the installation.
Install Anaconda from here into
<root_dir>/anaconda3, i.e., set the prefix accordingly in the installer.
Activate Anaconda installation:
Install pytorch via conda
conda install pytorch=1.4.0 torchvision cudatoolkit=<your_cuda_version> -c pytorch
pytorch-geometricand its dependencies following the instructions on its gh-page.
cd <root_dir> git clone -b 'submission_icml2020' --single-branch --depth 1 https://github.com/c-hofer/torchph.git conda develop torchph
Clone this repository into
Generate the experiment configurations you want using the
write_exp_cfgs_file.ipynbnotebook. It is assumed that the notebook server is started in
train.pyscript 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.
results.ipynbcontains some code to browse the results.