search

cvignac / DiGress

code for the paper "DiGress: Discrete Denoising diffusion for graph generation"
MIT License
353 stars 71 forks source link

How long does the code run in your GPU server? #32

Open FairyFali opened 1 year ago

cvignac commented 1 year ago

For what dataset do you want to know the runtime?

FairyFali commented 1 year ago

For what dataset do you want to know the runtime?

such as QM9 and the longest time in any datasets you test.

cvignac commented 1 year ago

It's very approximate, but it's in the order of 6h for QM9, 2 days for planar, and one week for sbm, guacamol and moses.

I'm not sure that all models had converged, though. For example, it was very easy to beat all previous methods on planar, so we did not run the model for as long as the SBM one.

cvignac commented 1 year ago

I seem to have similar statistics after 3k epochs: Sampling statistics {'spectre': 0.0158276848560015, 'clustering': 0.213101743441406, 'orbit': 0.05987776607931439, 'planar_acc': 0.0, 'sampling/frac_unique': 1.0, 'sampling/frac_unique_non_iso': 1.0, 'sampling/frac_unic_non_iso_valid': 0.0, 'sampling/frac_non_iso': 1.0}

Here is the training curve:

image

If your planar acc is still 0 after a long time of training, check that the package versions match the one in the latest commit.

xinyangATK commented 5 months ago

Hi, I would like to know which types of GPU did you use in training these dataset, including small and large dataset. Thanks!