WARNING: this is legacy code, which is no longer maintained!! The PersLay model of this repo is now improved and part of the GUDHI library (https://gudhi.inria.fr/python/latest/representations_tflow_itf_ref.html).
Authors: Mathieu Carriere, Theo Lacombe, Martin Royer
Note: This is an alpha version of PersLay. Feel free to contact the authors for any remark, suggestion, bug, etc.
This repository provides the implementation of PersLay, a tensorflow layer specifically desgined to handle persistence diagrams. This implementation follows the work presented in [1].
It contains a jupyter notebook tutorialPersLay.ipynb to try PersLay on graphs and orbits data as in [1].
PersLay can be installed by running the following instructions in a terminal:
$ git clone https://github.com/MathieuCarriere/perslay
$ cd perslay
$ (sudo) pip install .
All the code was implemented in python 3.6. It is likely that more recent versions would also work.
PersLay itself is simply a neural network layer implemented using a tensorflow backend and some numpy.
Tests were run on a Linux environment (Ubuntu 18.04) with the following versions of these libraries
(it is likely that other versions---especially more recent ones---should also work):
numpy: 1.15.4tensorflow: 2.0
Tensorflow can easily be installed using conda, pip, or following the instructions at https://www.tensorflow.org/install.
Remark: We used tensorflow-gpu 1.13 in an older version of the code, please check old releases of PersLay on the GitHub repo if you want to keep using Tensorflow versions less than 2.0.
In order to show how PersLay can be used in a learning pipeline on real-life data, we provide a jupyter notebook. This notebook has a few more dependencies.
sklearn: 0.20.2scipy: 1.1.0pandas: 0.23.4matplotlib: 3.0.3h5py: 2.8.0andhdf5: 1.10.2(used to store and load persistence diagrams)
Furthermore, jupyter notebook (or equivalent) is needed to run tutorialPersLay.ipynb.
In order to generate and/or preprocess persistence diagrams, we rely on the GUDHI library, which is a C++/python library whose python version can be installed using
$ conda install -c conda-forge gudhi
Otherwise, one can follow the steps at http://gudhi.gforge.inria.fr/python/latest/installation.html.
The main repository contains the python file perslay.py that defines the PersLay operation and the different types of layers that can be used with it.
Moreover, the folder tutorial contains the python notebook tutorialPersLay.ipynb, that reproduces the experiments in [1], contains an example of neural network using PersLay that can be used as a template for other experiments, and is hopefully easy to use and self-contained.
This notebook relies on the aforementioned libraries and on the various functions defined in the files utils.py, preprocessing.py and expe.py.
The /data/ repository contains the graph datasets used in the experiments of [1].
Graphs datasets (aside the COLLAB and REDDIT ones, see below) also contains a /mat/ folder where the different graphs
(encoded by their adjacency matrix) are stored (.mat files). Orbit datasets are generated on-the-fly.
Finally, you will also find the python notebook visuPersLay.ipynb in the tutorial folder, which contains examples of PersLay computations on a single persistence diagram. This notebook shows how the usual persistence vectorizations, such as persistence landscapes or images, can be retrieved as special cases of the PersLay architecture.
In [1], we also performed experiments using COLLAB, REDDIT5K and REDDIT12K datasets.
REDDIT5K and REDDIT12K datasets are large datasets (5,000 and 12,000 graphs respectively)
of large graphs (hundreds of nodes and edges).
As such, sharing online the adjacency matrices for these datasets is impossible
(folders are respectively of 18Gb and 30Gb size).
Similarly, COLLAB matrices were not included to not overload the repository (about 400Mb).
Unfortunately, the URL from which we downloaded the initial data, http://www.mit.edu/~pinary/kdd/datasets.tar.gz, appears to be down.
Feel free to contact one of the authors if you want more information.
The perslay package contains a class PerslayModel which implements a Tensorflow / Keras model, and which is initialized with four arguments:
-
name which is the Tensorflow name of the model,
-
diagdim which is the dimension of the persistence diagram points (usually 2),
-
rho which is a Tensorflow / Keras model used for postprocessing the concatenated vectorized persistence diagrams of all channels (see the architecture described in [1]). Use "identity" if you don't want to postprocess.
-
perslay_parameters, which is a python dictionary containing the parameters of PersLay. Examples can be found in
tutorialPersLay.ipynb.
In the following description of PersLay parameters, each parameter, or dictionary key, that contains _init in its name is optimized and learned by PersLay during training. If you do not want to optimize the vectorization, set the keys train_vect and train_weight to False.
-
The following keys are mandatory:
name description layer Either "PermutationEquivariant", "Image", "Landscape", "BettiCurve", "Entropy", "Exponential", "Rational" or "RationalHat". Type of the PersLay layer. "Image" is for persistence images, "Landscape" is for persistence landscapes, "Exponential", "Rational" and "RationalHat" are for structure elements, "PermutationEquivariant" is for the original DeepSet layer, defined in this article, "BettiCurve" is for Betti curves and "Entropy" is for entropy. perm_op Either "sum", "mean", "max", "topk". Permutation invariant operation. keep Number of top values to keep. Used only if perm_op is "topk". pweight Either "power", "grid", "gmix" or None. Weight function to be applied on persistence diagram points. If "power", this function is a (trainable) coefficient times the distances to the diagonal of the points to a certain power. If "grid", this function is piecewise-constant and defined with pixel values of a grid. If "gmix", this function is defined as a mixture of Gaussians. If None, no weighting is applied. final_model A Tensorflow / Keras model used to postprocess the persistence diagrams in each channel. Use "identity" if you don't want to postprocess.
Depending on what pweight is, the following additional keys are requested:
-
if pweight is "power":
name description pweight_init Initializer of the coefficient of the power weight function. It can be either a single value, or a random initializer from tensorflow, such as tensorflow.random_uniform_initializer(0., 1.). pweight_power Integer used for exponentiating the distances to the diagonal of the persistence diagram points. -
if pweight is "grid":
name description pweight_size Grid size of the grid weight function. It is a tuple of integer values, such as (10,10). pweight_bnds Grid boundaries of the grid weight function. It is a tuple containing two tuples, each containing the minimum and maximum values of each axis of the plane. Example: ((-0.01, 1.01), (-0.01, 1.01)). pweight_init Initializer for the pixel values of the grid weight function. It can be either a numpy array of values, or a random initializer from tensorflow, such as tensorflow.random_uniform_initializer(0., 1.). -
if pweight is "gmix":
name description pweight_num Number of Gaussian functions of the mixture of Gaussians weight function. pweight_init Initializer of the means and variances of the mixture of Gaussians weight function. It can be either a numpy array of values, or a random initializer from tensorflow, such as tensorflow.random_uniform_initializer(0., 1.).
Depending on what layer is, the following additional keys are requested:
-
if layer is "PermutationEquivariant":
name description lpeq Sequence of permutation equivariant operations, as defined in the DeepSet article. It is a list of tuples of the form (dim, operation). Each tuple defines a permutation equivariant function of dimension dim and second permutation operation operation (string, either "max", "min", "sum" or None). Second permutation operation is optional and is not applied if operation is set to None. Example: [(150, "max"), (75, None)]. lweight_init Initializer for the weight matrices of the permutation equivariant operations. It can be either a numpy array of values, or a random initializer from tensorflow, such as tensorflow.random_uniform_initializer(0., 1.). lbias_init Initializer for the biases of the permutation equivariant operations. It can be either a numpy array of values, or a random initializer from tensorflow, such as tensorflow.random_uniform_initializer(0., 1.). lgamma_init Initializer for the Gamma matrices of the permutation equivariant operations. It can be either a numpy array of values, or a random initializer from tensorflow, such as tensorflow.random_uniform_initializer(0., 1.). -
if layer is "Image":
name description image_size Persistence image size. It is a tuple of integer values, such as (10,10). image_bnds Persistence image boundaries. It is a tuple containing two tuples, each containing the minimum and maximum values of each axis of the plane. Example: ((-0.01, 1.01), (-0.01, 1.01)). lvariance_init Initializer for the bandwidths of the Gaussian functions centered on the persistence image pixels. It can be either a single value, or a random initializer from tensorflow, such as tensorflow.random_uniform_initializer(0., 3.). -
if layer is "Landscape":
name description lsample_num Number of samples of the diagonal that will be evaluated on the persistence landscapes. lsample_init Initializer of the samples of the diagonal. It can be either a numpy array of values, or a random initializer from tensorflow, such as tensorflow.random_uniform_initializer(0., 1.). -
if layer is "BettiCurve":
name description lsample_num Number of samples of the diagonal that will be evaluated on the Betti curves. lsample_init Initializer of the samples of the diagonal. It can be either a numpy array of values, or a random initializer from tensorflow, such as tensorflow.random_uniform_initializer(0., 1.). theta Sigmoid parameter used for approximating the piecewise constant functions associated to the persistence diagram points. -
if layer is "Entropy":
name description lsample_num Number of samples on the diagonal that will be evaluated on the persistence entropies. lsample_init Initializer of the samples of the diagonal. It can be either a numpy array of values, or a random initializer from tensorflow, such as tensorflow.random_uniform_initializer(0., 1.). theta Sigmoid parameter used for approximating the piecewise constant functions associated to the persistence diagram points. -
if layer is "Exponential":
name description lnum Number of exponential structure elements that will be evaluated on the persistence diagram points. lmean_init Initializer of the means of the exponential structure elements. It can be either a numpy array of values, or a random initializer from tensorflow, such as tensorflow.random_uniform_initializer(0., 1.). lvariance_init Initializer of the bandwidths of the exponential structure elements. It can be either a numpy array of values, or a random initializer from tensorflow, such as tensorflow.random_uniform_initializer(3., 3.). -
if layer is "Rational":
name description lnum Number of rational structure elements that will be evaluated on the persistence diagram points. lmean_init Initializer of the means of the rational structure elements. It can be either a numpy array of values, or a random initializer from tensorflow, such as tensorflow.random_uniform_initializer(0., 1.). lvariance_init Initializer of the bandwidths of the rational structure elements. It can be either a numpy array of values, or a random initializer from tensorflow, such as tensorflow.random_uniform_initializer(3., 3.). lalpha_init Initializer of the exponents of the rational structure elements. It can be either a numpy array of values, or a random initializer from tensorflow, such as tensorflow.random_uniform_initializer(3., 3.). -
if layer is "RationalHat":
name description lnum Number of rational hat structure elements that will be evaluated on the persistence diagram points. lmean_init Initializer of the means of the rational hat structure elements. It can be either a numpy array of values, or a random initializer from tensorflow, such as tensorflow.random_uniform_initializer(0., 1.). lr_init Initializer of the threshold of the rational hat structure elements. It can be either a numpy array of values, or a random initializer from tensorflow, such as tensorflow.random_uniform_initializer(3., 3.). q Norm parameter.
If you use this code or refer to it, please cite
[1] PersLay: A Neural Network Layer for Persistence Diagrams and New Graph Topological Signatures. Mathieu Carriere, Frederic Chazal, Yuichi Ike, Theo Lacombe, Martin Royer, Yuhei Umeda Proceedings of the Twenty Third International Conference on Artificial Intelligence and Statistics, PMLR 108:2786-2796, 2020.