DiffSBDD icon indicating copy to clipboard operation
DiffSBDD copied to clipboard

Issue running full atom joint model

Open amitkadan opened this issue 1 year ago • 0 comments

Hi there, I am working on reproducing the results reported in the paper on the crossdocked dataset, and was able to successfully do so when running test.py with the crossdocked_fullatom_cond and the crossdocked_ca_joint model, i.e., by running python test.py crossdocked_ca_joint.ckpt --test_dir <crossdocked_dir>/processed_noH/test/ --outdir <output_dir> --sanitize --timesteps 50 --n_nodes_bias 5 However, when I try to run the crossdocked_fullatom_joint model I get,

0%|                                                                                                        | 0/100
[00:00<?, ?it/s]/shared/akadan/DiffSBDD/test.py:172: UserWarning: Attempt 1/10 failed with error: 'NaN detected in EGNN output'. Trying again...
...
RuntimeError: Maximum number of retries exceeded

I tried running generate_ligands.py with this same model and parameters, and noticed that this error always arises after 5 diffusion steps. Am I doing something wrong when running this model? It seems be the right parameters as specified in the paper (i.e. 50 diffusion steps with 10 resampling steps).

Thanks in advance, Amit

amitkadan avatar Oct 02 '23 23:10 amitkadan