https://github.com/pvnieo/diffusion-net-plus
https://github.com/pvnieo/diffusion-net-plus
Last synced: about 2 months ago
JSON representation
- Host: GitHub
- URL: https://github.com/pvnieo/diffusion-net-plus
- Owner: pvnieo
- License: gpl-3.0
- Created: 2023-08-09T14:25:07.000Z (almost 2 years ago)
- Default Branch: master
- Last Pushed: 2025-02-06T14:02:14.000Z (4 months ago)
- Last Synced: 2025-03-26T21:47:38.976Z (2 months ago)
- Language: Python
- Size: 41 KB
- Stars: 4
- Watchers: 2
- Forks: 0
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
- License: LICENSE
Awesome Lists containing this project
README
:rotating_light::construction: **WORK IN PROGRESS** :construction::rotating_light:
# DiffusionNet++ :speaking_head: :spider_web: :heavy_plus_sign::heavy_plus_sign:
[](https://arxiv.org/abs/XXXX.XXXXX)
> Add batching support to the original DiffusionNet implementation, better performance, more features, and simpler API.
---
## :negative_squared_cross_mark: ToDos
- [ ] Add documentation to all functions
- [ ] Add installation files for pip
- [ ] push the code to PyPI
- [ ] clean the code (style, comments, variable names, unused parts, legacy code, etc.)
- [ ] reduce the number of dependencies
## :construction_worker: Installation
This implementation requires Python >= 3.8, you can install it locally using the following commands:
```bash
python setup.py sdist bdist_wheel
pip install .
```
An installation method using PyPI will be available soon.
```bash
pip install diffusion-net
```
#### Conda Environment
You can create a conda environment that runs this code as follows:
```bash
conda create -n diffusion -y
conda activate diffusion
conda install -y python=3.8
# use the correct cuda version, e.g., `11.7` for cuda 11.7
conda install pytorch=2.2 torchvision torchaudio pytorch-cuda=${CUDA-VERSION} -c pytorch -c nvidia -y
conda install pyg -c pyg
conda install pytorch-sparse -c pyg
pip install git+https://github.com/pyg-team/pyg-lib.git
pip install trimesh potpourri3d
```
## :book: Usage
The DiffusionNet++ has a simple API that can be used to instantiate the model, prepare the data, and batch it using the PyG library. The following code snippet demonstrates how to use the DiffusionNet++ model to extract features from a batch that has shapes of different sizes.
```python
import numpy as np
import trimesh
import torch
from torch_geometric.data import Batch
from diffusion_net import DiffusionData, DiffusionOperatorsTransform, DiffusionNet
# load some meshes
mesh1 = trimesh.load("path/to/mesh1.obj") # a mesh with N1 vertices
mesh2 = trimesh.load("path/to/mesh2.obj") # a mesh with N2 vertices, N2 != N1
v1, f1 = np.array(mesh1.vertices), np.array(mesh1.faces)
v2, f2 = np.array(mesh2.vertices), np.array(mesh2.faces)
# create the data objects
data1 = DiffusionData(pos=torch.from_numpy(v1), face=torch.from_numpy(f1).T)
data2 = DiffusionData(pos=torch.from_numpy(v2), face=torch.from_numpy(f2).T)
# compute the diffusion operators
diffusion_transform = DiffusionOperatorsTransform(n_eig=97) # compute the diffusion net operators with 97 eigenvalues
data1 = diffusion_transform(data1)
data2 = diffusion_transform(data2)
# create a batch
my_batch = Batch.from_data_list([data1, data2])
my_batch.x = my_batch.pos.clone() # set the input features to the positions
# create the model and do a forward pass
diffusion_net = DiffusionNet(3, 69) # input features are 3D positions, output features dimension is 69
output = diffusion_net(my_batch)
print(output.x.shape)
>>> torch.Size([(N1 + N2), 69])
```
## :chart_with_upwards_trend: Results
We reproduced the results of the original paper on the XX dataset. The results are detailed in the technical report [here](https://arxiv.org/abs/XXXX.XXXXX).
## :mortar_board: Citation
If you find this work useful in your research, please consider citing:
```bibtex
@inproceedings{mallet2024atomsurfsurfacerepresentation,
title={AtomSurf : Surface Representation for Learning on Protein Structures},
author={Vincent Mallet and Souhaib Attaiki and Yangyang Miao and Bruno Correia and Maks Ovsjanikov},
year={2025},
booktitle={The Thirteenth International Conference on Learning Representations (ICLR)},
}
```