Manifold Topology Divergence: a Framework for Deep Generative Models Evaluation

This is an implementation of the algorithms from the paper https://arxiv.org/pdf/2106.04024:

Barannikov, S., Trofimov, I., Sotnikov, G., Trimbach, E., Korotin, A., Filippov, A., & Burnaev, E. (2021).
Manifold Topology Divergence: a Framework for Comparing Data Manifolds. NeurIPS'21.

Installation

1. Requires numpy, scipy, torch.
2. Install ripserplusplus: git clone https://github.com/simonzhang00/ripser-plusplus.git; cd ripser-plusplus; python setup.py install
3. Install MTopDiv: git clone https://github.com/IlyaTrofimov/MTopDiv.git; cd MTopDiv; python setup.py install

Alternatively, you can use dockerfile. It also contains implementations of other algorithms which were tested the paper (Geometry Score, FID, IMD). In the docker, run conda activate py37 after start. The directory MTopDiv/experiments contains jupyter notebooks with experiments from the paper.

Example.ipynb can be executed in Google Colab.

Usage example

import numpy as np
import mtd

np.random.seed(7)
P = np.random.rand(1000, 2)
Q = np.random.rand(1000, 2)

barc = mtd.calc_cross_barcodes(P, Q, batch_size1 = 100, batch_size2 = 1000)

mtd.mtopdiv(P, Q, batch_size1 = 100, batch_size2 = 1000)
0.035247244

For GAN evaluation we recommend to use 0.5 * (mtopdiv(P, Q) + mtopdiv(Q, P)), where P, Q are the point clouds of real and generated objects.