search

cvignac / DiGress

code for the paper "DiGress: Discrete Denoising diffusion for graph generation"
MIT License
347 stars 73 forks source link

About node feature in non-molecular datasets #79

Open Jary-lrj opened 9 months ago

Jary-lrj commented 9 months ago

Hi: When I create my own dataset with node feature of 128 dimension, should I set self.node_types = torch.tensor(128*[1])? If I don't do so, how can I aviod RuntimeError: Expected size for first two dimensions of batch2 tensor to be: [200, 128] but got: [200, 1]?

Thanks.

Mutual-Luo commented 4 months ago

I think this paper is targetting attributed graphs, especially molecular graphs. As the author said in "Related Work": "Concurrently, Haefeli et al. (2022) designed a model limited to unattributed graphs, and similarly observed that discrete diffusion is beneficial for graph generation."

If you want to create your own dataset, you can refer the paper "Diffusion Models for Graphs Benefit From Discrete State Spaces-Kilian Konstantin Haefeli et al.".This is a concurrent work, so I think the author doesn't compare Digress with it.