Simulator for (Topology-Aware) Representation Learning Methods

IMPORTANT: Work in progress...

This repository provides a simulation framework for different representation learning algorithms such as Witness Autoencoder, Topological Autoencoder, Connectivity optimized representation learning, UMAP and t-SNE. It is easily extensible, and allows for reproducible and efficient experimenting with the different algorithms on different datasets.

Installation

pre-commit

Pre-commit is used to protect the master branch from (accidental) commits.

pip install pre-commit
pre-commit --version
pre-commit install

Requirements

pip3 install -r requirements.txt

torchph (only needed for COREL)

https://github.com/c-hofer/torchph

Recommended to install it this way: (not with conda)

git clone https://github.com/c-hofer/torchph.git
cd torchph
python setup.py develop

Usage

The simulator can be used with:

python main.py

One can choose a model, the number of parallel processes and a configuration for the model. Please refer to scripts/config_library/sample.py, where sample configurations can be found. Currently the simulator supports:

• Witness Autoencoder (command: -m 'WCAE')
• Topological Autoencoder (TopoAE)[1]. (command: -m 'topoae')
• t-SNE, UMAP (command: -m 'competitor')
• "COREL"[2] (not integrated, needs to be used manually at this point)

For the competitors, the model is defined directly in the configuration file.

A configuration can be passed to the simulator with '-c #name', e.g.

python main.py -m 'WCAE' -c 'sample.WCAE_sample_config'

The simulator has "access" to new configurations stored in scripts/config_library.

The number of parallel processes can be adjusted by '-n xy', e.g.

python main.py -m 'WCAE' -c 'sample.WCAE_sample_config' -n 2

Note that it defines how many models that are trained in parallel, it does not parallelize the model itself, this can be done in the configuration file if supported.

References

[1] M. Moor, M. Horn, B. Rieck, and K. Borgwardt. Topological Autoencoders. ICML 2020.

[2] C. Hofer, R. Kwitt, M. Dixit, and M. Niethammer. Connectivity-Optimized Representation Learning via Persistent Homology. ICML 2019.