Meige Guan

> When building a Graph with PyG, the molecular dataset already provides the data required by `Torch_geometric.data.data`. For example, the sample in `ZINC`: > > ``` > {'num_atom': 24, >...

> At the end of the paper: > > > **Node Representation**. There is a wide range of node representation tasks on graph structured data, such as > > finance,...

The argument `edge_type` assigns the way to calculate attention bias between every two nodes. It defaults to `multi_hop`, which means using the edge in the shortest path to calculate attention...

Hi, sorry for the late response. We're working on moving the pretrained models to a new site, and they will be available soon, please stay tuned.

To fix the problem, you can change pcqm4mv1 pretrained model's url to https://ml2md.blob.core.windows.net/graphormer-ckpts/checkpoint_best_pcqm4mv1.pt in file `graphormer/pretrain/__init__.py`. We'll update the code soon.

For now you can use our oc20 pretrained model (which will be moved to the new blob container soon), but we can't guarantee best performance because it was pretrained on...

Hi! As for 3D pretrained models, we only released a model pretrained on the oc20 dataset. But the Graphormer 3D model we provided (in `graphormer/models/graphormer_3d.py`) can be trained on any...

Yes, if you want to train a Graphormer 3d model, you'll need 3d position information for sure. Atoms or tags information is optional, but it's suggested to include atom types...

Hi, sorry for the late response. For the node classification problem, you can modify Graphormer's `forward` method to produce predictions for each node. The `GraphormerGraphEncoder`'s `forward` method will return `graph_node_feature`...

Hi! Based on our current implementation, it might take some extra effort to satisfy your need. You can try to wrap up code directly related to Graphormer model (which is...