DiGress - GitLite

KO | EN

gitlite — search

#python #react #typescript #llm #android #rust #deep-learning #docker #java #awesome #cli #kubernetes #awesome-list #mcp

DiGress

(★ 529)

code for the paper "DiGress: Discrete Denoising diffusion for graph generation"

파일 탐색기

최종 버전 다운로드 (.zip)

- - comm20.yaml
  - guacamol.yaml
  - moses.yaml
  - planar.yaml
  - qm9.yaml
  - sbm.yaml
- - comm20.yaml
  - debug.yaml
  - guacamol.yaml
  - moses.yaml
  - nx_graphs.yaml
  - planar.yaml
  - qm9_no_h.yaml
  - qm9_with_h.yaml
  - sbm.yaml
  - test.yaml
- - general_default.yaml
- - continuous.yaml
  - discrete.yaml
- - train_default.yaml
- __init__.py
- config.yaml
- digress_guacamol_smiles.txt
- final_smiles_qm9_noH.txt
- final_smiles_qm9withH.txt
- generated_planar_adj_matrices.npz
- generated_samples_moses.txt
- generated_samples_qm9withH.txt
- generated_samples_sbm.txt
- generated_smiles_moses.txt
- samples_qm9_no_h.txt
- - __init__.py
  - dist_helper.py
  - rdkit_functions.py
  - spectre_utils.py
  - visualization.py
- - __init__.py
  - abstract_dataset.py
  - guacamol_dataset.py
  - moses_dataset.py
  - qm9_dataset.py
  - spectre_dataset.py
- - __init__.py
  - diffusion_utils.py
  - distributions.py
  - extra_features.py
  - extra_features_molecular.py
  - layers.py
  - noise_schedule.py
- - __init__.py
  - abstract_metrics.py
  - molecular_metrics.py
  - molecular_metrics_discrete.py
  - train_metrics.py
- - __init__.py
  - layers.py
  - transformer_model.py
- __init__.py
- diffusion_model.py
- diffusion_model_discrete.py
- main.py
- utils.py
.gitignore
LICENSE
README.md
requirements.txt
setup.py

# 설치 가이드

1. 코드 내려받기

git clone https://github.com/cvignac/DiGress

깃허브에서 프로젝트 코드 전체를 내 컴퓨터로 내려받습니다.

cd DiGress

방금 내려받은 프로젝트 폴더 안으로 이동합니다.

2. Python

쉬움 추천

사전 준비물

Git GitHub에서 프로젝트 코드를 내려받으려면 필요합니다.
Python 3 설치 시 'Add Python to PATH' 옵션을 꼭 체크하세요 (Windows).

```pip3 install torch==2.0.1 --index-url https://download.pytorch.org/whl/cu118```

requirements.txt 등에 명시된 파이썬 라이브러리를 설치합니다.

```pip install -r requirements.txt```

requirements.txt 등에 명시된 파이썬 라이브러리를 설치합니다.

```pip install -e .```

requirements.txt 등에 명시된 파이썬 라이브러리를 설치합니다.

✅ 에러 메시지 없이 실행되고 터미널에 안내 문구가 출력되면 정상입니다.

이 레포의 README에 적힌 실제 명령어를 그대로 가져왔습니다.

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

# DiGress: Discrete Denoising diffusion models for graph generation

Update (Nov 20th, 2023): Working with large graphs (more than 100-200 nodes)? Consider using SparseDiff, a sparse version of DiGress: https://github.com/qym7/SparseDiff

Update (July 11th, 2023): the code now supports multi-gpu. Please update all libraries according to the instructions. 
All datasets should now download automatically

- For the conditional generation experiments, check the `guidance` branch.
  - If you are training new models from scratch, we recommand to use the `fixed_bug` branch in which some neural
network layers have been fixed. The `fixed_bug` branch has not been evaluated, but should normally perform better.
If you train the `fixed_bug` branch on datasets provided in this code, we would be happy to know the results.

## Environment installation
This code was tested with PyTorch 2.0.1, cuda 11.8 and torch_geometrics 2.3.1

- Download anaconda/miniconda if needed
  - Create a rdkit environment that directly contains rdkit:
    
    ```conda create -c conda-forge -n digress rdkit=2023.03.2 python=3.9```
  - `conda activate digress`
  - Check that this line does not return an error:
    
    ``` python3 -c 'from rdkit import Chem' ```
  - Install graph-tool (https://graph-tool.skewed.de/): 
    
    ```conda install -c conda-forge graph-tool=2.45```
  - Check that this line does not return an error:
    
    ```python3 -c 'import graph_tool as gt' ```
  - Install the nvcc drivers for your cuda version. For example:
    
    ```conda install -c "nvidia/label/cuda-11.8.0" cuda```
  - Install a corresponding version of pytorch, for example: 
    
    ```pip3 install torch==2.0.1 --index-url https://download.pytorch.org/whl/cu118```
  - Install other packages using the requirement file: 
    
    ```pip install -r requirements.txt```

- Run:
    
    ```pip install -e .```

- Navigate to the ./src/analysis/orca directory and compile orca.cpp: 
    
     ```g++ -O2 -std=c++11 -o orca orca.cpp```

Note: graph_tool and torch_geometric currently seem to conflict on MacOS, I have not solved this issue yet.

## Run the code
  
  - All code is currently launched through `python3 main.py`. Check hydra documentation (https://hydra.cc/) for overriding default parameters.
  - To run the debugging code: `python3 main.py +experiment=debug.yaml`. We advise to try to run the debug mode first
    before launching full experiments.
  - To run a code on only a few batches: `python3 main.py general.name=test`.
  - To run the continuous model: `python3 main.py model=continuous`
  - To run the discrete model: `python3 main.py`
  - You can specify the dataset with `python3 main.py dataset=guacamol`. Look at `configs/dataset` for the list
of datasets that are currently available
    
## Checkpoints

**My drive account has unfortunately been deleted, and I have lost access to the checkpoints. If you happen to have a downloaded checkpoint stored locally, I would be glad if you could send me an email at vignac.clement@gmail.com or raise a Github issue.**

The following checkpoints should work with the latest commit:

- [QM9 (heavy atoms only)](https://drive.switch.ch/index.php/s/8IhyGE4giIW1wV3) \\
  
  - [Planar](https://drive.switch.ch/index.php/s/8IhyGE4giIW1wV3) \\

- MOSES (the model in the paper was trained a bit longer than this one): https://drive.google.com/file/d/1LUVzdZQRwyZWWHJFKLsovG9jqkehcHYq/view?usp=sharing -- This checkpoint has been sent to me, but I have not tested it. \\

- SBM: ~~https://drive.switch.ch/index.php/s/rxWFVQX4Cu4Vq5j~~ \\
    Performance of this checkpoint:
    - Test NLL: 4757.903
    - `{'spectre': 0.0060240439382095445, 'clustering': 0.05020166160905111, 'orbit': 0.04615866844490847, 'sbm_acc': 0.675, 'sampling/frac_unique': 1.0, 'sampling/frac_unique_non_iso': 1.0, 'sampling/frac_unic_non_iso_valid': 0.625, 'sampling/frac_non_iso': 1.0}`

- Guacamol: https://drive.google.com/file/d/1KHNCnPJmPjIlmhnJh1RAvhmVBssKPqF4/view?usp=sharing -- This checkpoint has been sent to me, but I have not tested it.

## Generated samples

We provide the generated samples for some of the models. If you have retrained a model from scratch for which the samples are
not available yet, we would be very happy if you could send them to us!

## Troubleshooting

`PermissionError: [Errno 13] Permission denied: '/home/vignac/DiGress/src/analysis/orca/orca'`: You probably did not compile orca.

## Use DiGress on a new dataset

To implement a new dataset, you will need to create a new file in the `src/datasets` folder. Depending on whether you are considering
molecules or abstract graphs, you can base this file on `moses_dataset.py` or `spectre_datasets.py`, for example. 
This file should implement a `Dataset` class to process the data (check [PyG documentation](https://pytorch-geometric.readthedocs.io/en/latest/tutorial/create_dataset.html)), 
as well as a `DatasetInfos` class that is used to define the noise model and some metrics.

For molecular datasets, you'll need to specify several things in the DatasetInfos:
  - The atom_encoder, which defines the one-hot encoding of the atom types in your dataset
  - The atom_decoder, which is simply the inverse mapping of the atom encoder
  - The atomic weight for each atom atype
  - The most common valency for each atom type

The node counts and the distribution of node types and edge types can be computed automatically using functions from `AbstractDataModule`.

Once the dataset file is written, the code in main.py can be adapted to handle the new dataset, and a new file can be added in `configs/dataset`.

## Cite the paper

```
@inproceedings{
vignac2023digress,
title={DiGress: Discrete Denoising diffusion for graph generation},
author={Clement Vignac and Igor Krawczuk and Antoine Siraudin and Bohan Wang and Volkan Cevher and Pascal Frossard},
booktitle={The Eleventh International Conference on Learning Representations },
year={2023},
url={https://openreview.net/forum?id=UaAD-Nu86WX}
}
```