Loading Atomrefs in Pretrained DimeNet mode

[Frank-LIU-520]

In the qm9_pretrained_dimenet.py of the example codes, there read # DimeNet uses the atomization energy for targets U0, U, H, and G. in line 12. But there is no way to find out how did DimeNet load and use these atomrefs in QM9 class. Can anybody tell me where I can trace how the atomrefs were loaded and used in the DimeNet training routine?

`import os.path as osp

import torch

from torch_geometric.datasets import QM9
from torch_geometric.data import DataLoader
from torch_geometric.nn import DimeNet

path = osp.join(osp.dirname(osp.realpath(file)), '..', 'data', 'QM9')
dataset = QM9(path)

# DimeNet uses the atomization energy for targets U0, U, H, and G.
idx = torch.tensor([0, 1, 2, 3, 4, 5, 6, 12, 13, 14, 15, 11])
dataset.data.y = dataset.data.y[:, idx]`

[rusty1s]

In the QM9 dataset, there exists both properties U_0 and U_0^ATOM. While some papers train against the U_0 property (target 7), DimeNet trains against U_0^ATOM (target 12). This is also mentioned in the DimeNet paper: "We only use the atomization energy for U0, U, H, and G, i.e. subtract the atomic reference energies, which are constant per atom type."

[Frank-LIU-520]

Thanks for the answer. Got it.