Pretrained checkpoints about mdm

[GRAPH-0]

Hi guys，

I have tried running the code, but found it hard to reproduce the paper results. Some test results reported on QM9 are as follows:

Any suggestions for this?

BTW, there are some minor bugs for training Geom Drug. Hope you can fix them in the future.

[Layne-Huang]

Hi,

Sorry for the late response. I have updated the test_eval.py script, which has some little errors.

You could run the command python test_eval.py --ckpt --sampling_type generalized --w_global_pos 1 -- w_global_node 1 --w_local_pos 4 --w_local_node 5. Then you could overserve good performance of MDM.

Regarding the checkpoint, I will upload it later.

[GRAPH-0]

Got it!
Kind reminders about Geom Drug Training:

• soft_edges -> soft_edge

  MDM/configs/geom_full.yml

  Line 18 in abdcd8b

  soft_edges: True

• missing val_loader for Geom

  MDM/train.py

  Line 213 in abdcd8b

  train_set = Geom(pre_transform=transforms)

[Layne-Huang]

Got it! Kindly reminders about Geom Drug Training:

• soft_edges -> soft_edge

  MDM/configs/geom_full.yml

  Line 18 in abdcd8b

  soft_edges: True

• missing val_loader for Geom

  MDM/train.py

  Line 213 in abdcd8b

  train_set = Geom(pre_transform=transforms)

Thanks for your advice, they are very helpful. Actually, we did not utilize the validation dataset in the Geom-drugs dataset. You could split it by yourself.