Official pytorch implementation for "Efficient and Degree-Guided Graph Generation via Discrete Diffusion Modeling". Our code is devloped based on https://github.com/ehoogeboom/multinomial_diffusion.
We use the evaluation modules provided by https://github.com/uoguelph-mlrg/GGM-metrics and https://github.com/hheidrich/CELL.
dgl
prettytable
scikit-learn
tensorboard
tensorflow
tensorflow-gan
torch
torch-geometric
tqdm
wandb
See node.ipynb, once you train the model, it's saved to the "./graphs" directory.
By default we use empirical degree sampler, which randomly takes a degree sequence from the training data as empirical with neural in the option --empty_graph_sampler if you have trained your neural degree sampler.
#!/bin/bash
python train.py \
--epochs 50000 \
--num_generation 64 \
--diffusion_dim 64 \
--diffusion_steps 128 \
--device cuda:1 \
--dataset Ego \
--batch_size 8 \
--clip_value 1 \
--lr 1e-4 \
--optimizer adam \
--final_prob_edge 1 0 \
--sample_time_method importance \
--check_every 500 \
--eval_every 500 \
--noise_schedule linear \
--dp_rate 0.1 \
--loss_type vb_ce_xt_prescribred_st \
--arch TGNN_degree_guided \
--parametrization xt_prescribed_st \
--empty_graph_sampler empirical \
--degree \
--num_heads 8 8 8 8 1
#!/bin/bash
python train.py \
--epochs 50000 \
--num_generation 64 \
--num_iter 256 \
--diffusion_dim 64 \
--diffusion_steps 512 \
--device cuda:0 \
--dataset polblogs \
--batch_size 4 \
--clip_value 1 \
--lr 1e-4 \
--optimizer adam \
--final_prob_edge 1 0 \
--sample_time_method importance \
--check_every 50 \
--eval_every 50 \
--noise_schedule linear \
--dp_rate 0.1 \
--loss_type vb_ce_xt_prescribred_st \
--arch TGNN_degree_guided \
--parametrization xt_prescribed_st \
--degree \
--num_heads 8 8 8 8 1
Evaluation is done every eval_every epochs. You can also re-evaluate a specific checkpoint using the script below.
python evaluate.py \
--run_name 2023-05-29_18-29-35 \
--dataset polblogs \
--num_samples 8 \
--checkpoints 5500
Training results can be found in wandb/{dataset_name}/multinomial_diffusion/multistep/{run_name}
We are still working on integrating the following two features into our code:
-
Even faster sampling by incrementally modifying graph.
-
Attributed graph generation.
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent