c-hofer/torchph

The essence of my research, distilled for reusability. Enjoy 🥃!

torchph

This repository contains PyTorch extensions to compute persistent homology and to differentiate through the persistent homology computation. The packaging structure is similar to PyTorch's structure to facilitate usage for people familiar with PyTorch.

Documentation

Read the docs!

The folder tutorials (within docs) contains some (more or less) minimalistic examples in form of Jupyter notebooks to demonstrate how to use the PyTorch extensions.

Associated publications

If you use any of these extensions, please cite the following works (depending on which functionality you use, obviously :)

@inproceedings{Hofer17a,
  author    = {C.~Hofer, R.~Kwitt, M.~Niethammer and A.~Uhl},
  title     = {Deep Learning with Topological Signatures},
  booktitle = {NIPS},
  year      = {2017}}

@inproceedings{Hofer19a,
  author    = {C.~Hofer, R.~Kwitt, M.~Dixit and M.~Niethammer},
  title     = {Connectivity-Optimized Representation Learning via Persistent Homology},
  booktitle = {ICML},
  year      = {2019}}

@article{Hofer19b,
  author    = {C.~Hofer, R.~Kwitt, and M.~Niethammer},
  title     = {Learning Representations of Persistence Barcodes},
  booktitle = {JMLR},
  year      = {2019}}
  
@inproceedings{Hofer20a},
  author    = {C.~Hofer, F.~Graf, R.~Kwitt, B.~Rieck and M.~Niethammer},
  title     = {Graph Filtration Learning},
  booktitle = {arXiv},
  year      = {2020}}
  
@inproceedings{Hofer20a,     
  author    = {C.~Hofer, F.~Graf, M.~Niethammer and R.~Kwitt},     
  title     = {Topologically Densified Distributions},     
  booktitle = {arXiv},    
  year      = {2020}} 

torchph's Issues

pershom.cpp seems to be missing from repo

When trying to make the pershom_cpp_src file, there is a reference to a 'pershom.cpp' that does not exist.

Related error:

make: *** No rule to make target 'pershom.cpp', needed by 'pershom.o'. Stop.

Internal variables/tensors in modules should be created with __new__ instead of __init__

See http://pytorch.org/docs/master/notes/cuda.html

How to compute loss for 1-dim and 2-dim VR Pershom?

@c-hofer @rkwitt

Renaming chofer_torchex to torchph

Planning to rename chofer_torchex as torchph for easier use;

remove utils

Utils needs to be refactored to pytorch_utils repository and removed from chofer_torchex

Address compiler warnings in cuda compilation

At the moment there are several warnings when compiling the cuda stuff. This should be investigated and removed.

solved!

Thanks for your work!

About how to use vr_persistence_l1 to parallel count

Hi, thanks for your work! But i want to know if i can count persistence homological parallel in gpu through torchph?

eg:
n pointclouds size:[batch,pointcloud_number,dim]
to
n persistence homologicals size:[batch,persistence_barcode]

Respecting the help and the answer that you give

pip deployement

Batch processing to get the topology features

Hi, thank you so much for sharing this useful library! I am wondering how to use it for handling batch input. For example, when I want to calculate the vr persistence l1 features of a batch with "BxNx3", one simple way is to use your "vr_persistence_l1" function with a for loop, which will cost a lot of time (B times compared with 1 input). Do you have any suggestion for this scenario?

torch.mul

Torch.mul behaviour is changed. Adapt to new behaviour for element wise multiplication.

vr complex on distance matrix

We need to refactor the vr computation such that distance matrices can be directly passed into the pipeline.

conda deployment

investigate if VR-complex should return distance as death times or distance/2

Currently the vr computation returns the distance between to vertices as a death time, it is not obvious if this is expected or not.

html documentation

make kernel calls compatible with pytorch streams

Currently all kernels operate on default stream. This makes the pytorch stream context manager useless and leads to crashes.

Allow documentary to be built without nvcc

Currently for sphinx we need to import all modules which fails if no nvcc is available as cuda extansions won't compile.
This dependency should be removed.

Migration from PyTorch 0.3 to 0.4

SLayer's derivative wrt input

Hello, I have a question about using SLayer. Could you please clarify why isn't it possible to differentiate w.r.t. its input?
Thank you

Rolling docs build at external server

So far we have to build docs locally and push them to gh-pages.

Wanted new behavior:

1. docs are updated automatically with each push of master

2. there remains a possobility to build docs locally in some gitignored folder for testing befor pushing

@rkwitt Research if u want to. i'll do the coding.